Policy Research Working Paper 11285 South Africa’s Fragmented Cities The Unequal Burden of Labor Market Frictions Javier E. Baez Varun Kshirsagar Poverty and Equity Global Department A verified reproducibility package for this paper is January 2026 available at http://reproducibility.worldbank.org, click here for direct access. Policy Research Working Paper 11285 Abstract Using high-resolution administrative, census, and satellite wellbeing) and 4.9-percentile drop in employment. In Cape data, this paper shows that South African cities are char- Town, the declines are 4.0 and 3.7 percentiles, respectively. acterized by spatial mismatches between where people live Employment is 87 percent lower in the poorest decile than and where jobs are located, relative to 20 global peers. Areas the richest in Johannesburg and 61 percent lower in Cape within 5 kilometers of commercial centers have 9,300 fewer Town. These findings suggest that South Africa’s spatial residents per square kilometer than expected, which is 60 organization of people and economic activity constrains percent below the global median. Poor, dense neighborhoods agglomeration and reinforces inequality. This methodology are most affected. In Johannesburg, a 10-percentile increase provides a scalable and standardized data-driven framework in distance from the nearest business hub corresponds to a to analyze spatial accessibility and agglomeration frictions 3.7-percentile drop in asset wealth (a proxy of household in complex, data-constrained urban systems. This paper is a product of the Poverty and Equity Global Department. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at jbaez@worldbank.org and varun.kshirsagar@gmail.com. A verified reproducibility package for this paper is available at http://reproducibility.worldbank.org, click here for direct access. RESEA CY LI R CH PO TRANSPARENT ANALYSIS S W R R E O KI P NG PA The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team South Africa’s Fragmented Cities: The Unequal Burden of Labor Market Frictions Javier E. Baez* Varun Kshirsagar World Bank World Bank Authorized for distribution by Rinku Murgai, Practice Manager, Poverty and Equity Global Department, World Bank Group KEY WORDS: Poverty, Spatial Inequality, Local Labor Markets, The Economics of Cities JEL Classification: R12, R14, R23, R41, R52, O18, J61 paper reflects the views of the authors and does not reflect the official views of the World Bank, its * This Executive Directors, or the countries they represent. Corresponding Author: Javier E. Baez, email: jbaez@worldbank.org. We sincerely thank Olive Nsababera and Nozipho Shabalala for their invaluable insights at the project’s outset. We also thank Edward Beukes and Mark Roberts for their valuable peer review comments, as well as Rinku Murgai and participants at the World Bank’s Poverty Department seminar in Washington DC and at the 9th Urbanization and Development Conference in Cape Town for their feedback. We also thank Mahin Tariq and Kunal Singh for the diligence and skill with which they conducted the World Bank’s reproducibility package analysis. 1 Introduction The success of South Africa’s democracy depends on our ability to change the material condi- tions of our people’s lives so that they not only have the right to vote but also bread and work. — Nelson Mandela (1994) Thirty years after apartheid, steady work remains out of reach for many of South Africa’s urban poor. The country has made notable gains in education and health, but unemployment remains persistently high.1 Thriving cities serve as labor markets by effi- ciently matching people to jobs (Moretti (2010), Bertaud (2018)). Why do South Africa’s largest cities struggle to create jobs? Are these constraints more severe than in other ma- jor African cities? This paper focuses on spatial frictions within cities that act as a tax on labor exchange. These distortions raise search costs, inhibit job matching, and blunt the agglomeration economies that underpin urban productivity.2 We show that South Africa’s cities exhibit unusually fragmented urban structures. A greater share of residents live far from business districts compared to other major African cities. This pattern, often termed spatial mismatch (Kain (1968)), refers to the physical separation of low-income populations from employment hubs.3 In most cities, population density declines with distance from employment centers. South African cities diverge sharply from this pattern. Despite widespread recognition of this pattern, spatial mismatch is rarely quantified in ways that enable consistent monitoring or cross-city comparison. Most existing stud- ies focus on commuting time, rely on household travel surveys, or are limited to a single city. These approaches are costly, infrequent, and difficult to scale. Policymakers lack a consistent, data-driven way to monitor spatial labor market frictions over time and across cities. This paper proposes such a framework, with a global analysis using building vol- umes and population density to establish global benchmarks and then combining census and administrative data on assets, employment and job density with open-source data 1 Beforethe transition, 13% of the population was unemployed; in the past decade, this figure has ranged from 25% to 30%, and has exceeded 40% when discouraged workers are included. Nearly two in three youths are jobless. Meanwhile, economic growth has been sluggish, often lagging behind population growth. By 2024, GDP per capita had fallen back to 2007 levels. Duval et al. (2021) estimate that if South Africa’s employment rate aligned with the OECD average, GDP per capita would be 50% higher than it is today. 2 See, for example, Duranton and Puga (2004), Glaeser and Gottlieb (2008), Marx et al. (2013), Bryan et al. (2020), and Barza et al. (2024). More generally, because of data limitations, the empirical literature has focused on spatial scales at which this consequential divergence between economic and population density is absent (e.g. Fay and Opal (2000), Jedwab and Vollrath (2015), Combes and Gobillon (2015), and Henderson and Turner (2020)). Our analysis would have been more challenging before the necessary data were easier to access and analyze at the more granular scales needed to address these issues (e.g., Donaldson and Storeygard (2016), Melchiorri et al. (2018)). 3 See Gobillon et al. (2007) and Gobillon and Selod (2021) for useful discussions on the spatial mismatch hypothesis. 2 on building footprints, volumes, and road networks to construct spatial accessibility and employment exposure metrics at a high resolution for three of South Africa’s largest cities. We show that spatial mismatch in South African cities significantly exceeds global norms, identifying them as outliers in international comparisons. Further, our approach advances prior work by quantifying spatial mismatch using harmonized geospatial data and network-based routing algorithms. This framework can be applied at scale across diverse urban geographies, including contexts where detailed household surveys are un- available. Our overarching contribution is to combine remote-sensed and administrative data into a unified spatial diagnostic that links international urban structure comparisons to within-city labor market access. Specifically, we find that population densities within 5 km of business districts are 60% lower in South African cities than global norms, and that poor, dense neighborhoods are disproportionately remote. In Johannesburg, a 10-percentile increase in distance from a business district corresponds to a 3.7-percentile drop in the asset index and 4.9-percentile drop in employment, and we see similar relationships in other major cities. Finally, we show that, in 2020 (25 years after apartheid ended), clusters of neighborhoods that were formerly segregated townships drive these aggregate spatial mismatch relations.4 This spatial configuration is inverted relative to international patterns, with the poorest resi- dents disproportionately concentrated in peripheral, low-access neighborhoods. This paper makes four contributions. First, we develop a transparent and scalable, data-driven procedure that adapts morphological center-detection methods (e.g. Tauben- böck et al. (2017)) to open GHSL built-up volume and population grids to delineate busi- ness districts (BDs). Our approach does not rely on monocentric assumptions or admin- istrative boundaries. We identify candidate commercial cells based on building volumes and low residential population, and then cluster them using a distance-based algorithm – single-linkage clustering (Friedman et al. (2001)) – that allows for corridor-shaped ag- glomerations.5 This produces a globally consistent definition of business districts using only open and observable data. We use these delineated BDs to measure how population density declines with distance to the nearest BD across 25 cities.6 4 These include Soweto in Johannesburg, Cape Flats in Cape Town, and Umlazi in Durban. 5 Arribas-Belet al. (2021) delineate employment centers in Spain using building-height information and a density-based clustering algorithm (DBSCAN), which emphasizes compact, high-density clusters. Our approach instead uses single-linkage clustering, which imposes fewer assumptions about cluster compact- ness and relies on scale-dependent parameters that are transparent and interpretable, making it easier to apply consistently across cities with very different urban forms. 6 These gradients are not mechanically driven by spatial autocorrelation in population density. By construc- tion, business districts are identified as locations with high built volume and low residential population, and distances are measured to the nearest such district rather than to population centers. A gravity-based accessibility metric combining employment mass and distance would be ideal, but the absence of globally comparable, fine-grained employment data precludes this approach. GHSL population grids disaggregate 3 Second, we show that South African cities are statistical outliers. They exhibit among the lowest population densities near business districts globally, with poor, densely popu- lated areas systematically located farther from economic centers. We quantify this diver- gence, complementing the anecdotal and historical narratives that have dominated both policy conversations and public discourse. Third, our results temper the prevailing optimism around urban polycentricity (Anas et al. (1998)). Using granular data, we show that in South Africa’s major cities vibrant business districts have formed away from the dense and poor former townships.7 Fourth, we link this spatial mismatch to local labor market outcomes.8 Asset-deprived neighborhoods are more remote and show lower levels of formal employment and higher volatility, particularly in lower-wage segments. These findings demonstrate the value of a scalable methodology for identifying mismatch hotspots. This may, in turn, inform geographically targeted infrastructure and housing investments as well as other urban development policies that enhance agglomeration and access to opportunities.9 These spatial patterns reflect a legacy of urban planning shaped by apartheid-era laws and reinforced by post-transition housing policy. Many authors argue that these policies are a key driver of South Africa’s persistently high urban unemployment (e.g., Banerjee et al. (2008), Hausmann et al. (2023)). The Urban Areas Act (1923) and the Group Areas Act (1950) played a central role in institutionalizing residential segregation and shaping the urban spatial structure that persists today (Simpson (2021), Thompson (2001)). These laws relegated Black South Africans to peripheral settlements, many of which continue to census counts to grid cells using ancillary spatial information and therefore reflect census timing and the lowest administrative level available in each country. Despite differences in temporal alignment and spa- tial granularity, GHSL remains one of the only globally consistent, openly available population surfaces suitable for cross-city analysis. For South Africa, GHSL population is accurate at the subplace (ADM5) level using inputs from Statistics South Africa. 7 Johannesburg has business districts in the traditional downtown, as well as in Sandton and Randburg, while Cape Town has business districts that run along a highway northeast of the traditional downtown through Parow and Bellville. Two major growth areas in Johannesburg (Sandton and Randburg) are north of the traditional downtown which is even further from the asset-deprived areas that are south of the downtown. Similarly, growth within Cape Town has been along a highway that runs northeast from the traditional downtown, far from the formerly segregated and still poor Cape Flats area. In Durban, Umh- langa has developed into a vibrant employment hub in a wealthy area north of the traditional downtown. 8 While analyzing household survey data offers many advantages, it cannot capture granular spatial rela- tionships, as household surveys are not representative at the neighborhood level (e.g., Banerjee et al. (2008), Bhorat et al. (2021)) 9 Spatialfrictions aside, the literature has identified four key structural factors behind persistently high un- employment since the end of apartheid (e.g., Banerjee et al. (2008), Bhorat et al. (2021)). First, labor supply grew, particularly as more African women entered the workforce. Second, job losses in mining and agricul- ture reduced labor demand. Third, a skills mismatch persists—by 2019, unemployment was 25% among those without a high school education and 35% for those with only primary or some secondary education. Fourth, sluggish private sector growth has led to weak job creation, especially in the informal sector, with real GDP per capita in 2024 still at 2007 levels. 4 exhibit limited access to employment and services. Despite the formal end of apartheid, post-1994 housing policies have failed to reverse this legacy. Large-scale housing programs have reinforced urban sprawl, often placing low-income communities even farther from job-rich commercial centers. Limited afford- able transportation options further restrict economic mobility, contributing to persistently high unemployment and inequality. As a result, the spatial mismatches created during apartheid remain a major obstacle to inclusive growth and economic opportunity. For the poorest households, commuting costs can absorb up to 85% of daily income, once the opportunity cost of time is included. These regressive costs make geographic distance from jobs a structural barrier to economic inclusion and a persistent driver of inequality. If commuting were fast and affordable, living farther from business districts wouldn’t significantly hinder labor exchange. However, Shah and Sturzenegger (2022) find that commuting costs consume 17% of wage income – or 57% when factoring in commuting time costs. These costs are regressive. For the poorest households, direct costs reach 35% of income and, as mentioned above, up to 85% with opportunity costs included. South African cities offer public transport (Passenger Rail, Bus Rapid Transit), but public options are so poorly managed, and weakly integrated with other transporta- tion systems that only 2% use them. Instead, 77% rely on private minibus taxis, which are costly, unsafe, have uneven coverage, and lack good links to public transit. The analysis presented in this paper has three parts. First, we provide evidence of the unusual disconnect between population density and the concentration of economic activity in South Africa’s major cities. Using high-resolution data (i.e., 100-meter grids) from satellite imagery, night-time lights, and census records, we compare spatial pat- terns across South African cities and 20 global peers across five continents. This quantifies the “missing” population close to employment centers, revealing that population densi- ties near business districts are about 9,300 people per square kilometer lower than global norms—60% below the median density of comparator cities. Furthermore, these patterns have distributional consequences, as densely populated but remote areas tend to be asset- deprived. In Johannesburg, the lowest asset decile has a population density 274% higher than the wealthiest decile, while in Cape Town, it is 360% higher. Second and more critically, dense and poor areas are surrounded by weaker local labor markets, with lower employment levels and higher volatility. In Johannesburg, residents in the bottom asset decile are 87% less likely to have formal jobs than those in the top decile; in Cape Town, the gap is 61%. Third, for the three major cities, we develop an even higher resolution spatial diagnostic that identifies the mismatched hot-spots. This provides a bridge to analyses that would inform investments in housing and transport 5 infrastructure.10 Manysheva et al. (2025) provide a model that quantifies how education and spatial frictions may constrain income convergence in the post-apartheid era. Our analysis com- plements theirs and differs in three critical ways. First, we provide evidence of South Africa’s spatial distribution anomalies using 20 global cities as a benchmark. Second, we use spatial employment data across South Africa’s major cities to analyze spatial inequal- ities in local employment. Third, we delineate business districts and estimate measures of isolation across South Africa’s major cities.11 The paper is structured as follows. Section 2 presents evidence of the unusual sep- aration between economic and population density in South African cities compared to 20 cities across five continents. Section 3 combines granular census data and spatial tax records to estimate the relationship between asset deprivation, population density and local labor market outcomes. Section 4 identifies neighborhoods that drive these rela- tionships, and whose populations may therefore benefit from investments in transport infrastructure and housing, including opportunities to live closer to employment centers. We conclude by discussing transport, housing and other urban development policies that could stimulate agglomeration and ease mobility and accessibility to jobs within South Africa’s major cities. 2 Spatial Mismatch between Population Density and Jobs South African cities exhibit an inverted urban form. In most cities globally, population density declines with distance from business districts. In contrast, dense and poor neigh- borhoods in South African cities are often located 10–20 km away from commercial hubs. To quantify this divergence, we construct a benchmark from 20 cities across five conti- nents, and compare it to Johannesburg, Cape Town, Durban, Tshwane, and Nelson Man- dela Bay. We find that South African cities have, on average, 60% lower population den- sities within 5 km of BDs compared to the global benchmark, indicating a reversal of the expected urban gradient. 10 Our global comparative analysis in Section 2 uses great circle distance because we lack neighborhood boundary data for most cities, making road network-based travel time estimation computationally chal- lenging. In contrast, Section 4 focuses on just three cities, where neighborhood (i.e., sub-place) boundaries and centroids are available, allowing us to compute Origin-Destination matrices and estimate optimized (frictionless) travel times along road networks. For the same reason, we use Global Human Settlement Layer (GHSL) building volumes in Section 2 to delineate business districts, whereas Section 4 uses Open Street Map Point of Interest tags and Google Open Building vectors for finer granularity. A comparison of Figure 1 and Figure 7a shows that the results are comparable, capturing the same polycentric pattern for Johannesburg. 11 See Shilpi et al. (2018) and Bastos and Bottan (2023) for relevant analysis of the Homelands outside the cities. 6 In standard models, households weigh commuting costs against housing space, lead- ing to higher densities near business districts where job access is better (Alonso (1964) and Mills (1967)). While deviations occur in polycentric cities, most retain a negative density gradient with distance to economic hubs. We use this stylized fact to benchmark South African cities against 20 global peers. Our approach builds on earlier work (e.g., Bertaud and Malpezzi (2003)), who model population density as a declining exponential function of distance to CBD. Their study as- sumes monocentricity and treats the geometric city center as a proxy for economic activity. We relax both assumptions. First, we do not assume either a monocentric structure or that the geometric center of gravity coincides with the economic center of gravity. Instead, we use building im- agery (and granular census data on population density) to delineate business districts. Consequently, following (Kloosterman and Musterd (2001) and Burger et al. (2015)) we allow cities to be either monocentric (e.g., Nairobi) or approximately monocentric (e.g., New York City, which has jobs concentrated in and around Manhattan) or polycentric (e.g., Johannesburg and Cape Town). Second, due to data constraints, we group areas by proximity bands rather than precise distance. Since we cannot measure travel times and distances precisely, these bands serve as a practical approximation. Our analysis uses a linear functional form but captures non-linearities by allowing coefficients to vary across bands. Across five continents, we find similar spatial distributions of the population in most cities, but major South African cities stand out with patterns consistent with the more pronounced horizontal urban growth. 2.1 Estimating a Global Benchmark To benchmark urban structure, we compare five South African cities to a set of 20 global cities across five continents. These global peers include major cities such as New York, Tokyo, London, Lagos, Jakarta, Mumbai, Bogotá, Buenos Aires, and Nairobi. The five South African cities included in our analysis are Johannesburg, Cape Town, Durban, Tshwane, and Nelson Mandela Bay. Given that many cities in our sample are polycen- tric, we develop a data-driven method to empirically identify business districts—avoiding assumptions about geometric centers or relying on administrative boundaries. The anal- ysis proceeds in four steps. First, we overlay each city with a 100-meter grid and extract gridded population and building volume data from the Global Human Settlement Layer (Melchiorri et al. (2018)). Raster layers are cropped to the geographic extent of the city boundary and masked to retain only pixels that fall within the city limits.12 12 Croppingreduces the spatial extent of the raster to the bounding box of the city, while masking removes values outside the boundary shape. All raster and vector layers are then reprojected to a common co- 7 Second, we identify business districts using building volumes and population. Re- fer to Figure 1 which illustrates our methodology using Johannesburg as an example. A grid cell is flagged as a candidate BD cell if it exceeds a minimum building volume and contains fewer than 50 residents. These candidate cells are clustered using single-linkage hierarchical clustering (Friedman et al. (2001)). This method groups the grid cells into commercial agglomerations (i.e. Business Districts). The method starts with every grid cell being marked as being in its own cluster. It then proceeds sequentially to combine clusters and the algorithm stops when no grid cell outside a cluster lies within 2,000 me- ters of any grid cell inside that cluster.13 A cluster is designated as a business district if it contains at least 10 grid cells.14 Third, we calculate the distance from each populated grid cell to the nearest business district and assign cells to one of five distance bands: 0–5 km, 5–10 km, 10–15 km, 15–20 km, or beyond 20 km. Fourth, we compute average population density within each dis- tance band for the 20 global cities to produce a benchmark gradient. This gradient is then applied to the five South African cities to evaluate whether population is disproportion- ately concentrated farther from employment centers than expected. These findings serve as the basis for the regression analysis in Section 2.2, where we investigate the statistical significance of these patterns across cities. 2.2 Quantifying the Population Density Gaps Close to Employment Centers We quantify the ‘missing’ population densities close to business districts in South Africa’s major cities relative to the patterns observed in the global benchmark. We find that areas within 5 km of a business district have 9,300 fewer people per square kilometer than ex- pected, 60% lower than the median density seen in areas close to business districts in the global benchmark. ordinate reference system (EPSG:4326) to ensure spatial consistency across data sources. While this CRS records locations in degrees rather than meters, we compute great-circle (haversine) distances to account for the curvature of the Earth when measuring proximity. 13 Thereare two advantages to using this approach. First, this allows the data to determine the number of clusters (i.e. urban agglomerations). Second, the approach and the results are easily interpretable be- cause two different agglomerations are grouped based on a cut-off distance between the nearest neighbors between the groups. This is preferable to other approaches in our context, because commercial agglomer- ations often cluster along highways and other non-compact corridors. In other contexts, compactness of clusters is desirable, and the possibility of forming a chain is seen as a weakness. But chains are useful in our context (i.e. we allow for business corridors). 14 Thesevalues are consistent with a 3-story building covering the entire area, with fewer than 50 residents, but constitutes one possible choice of parameters. Therefore, we test the robustness of our main regression result by allowing the building volume cutoff and the cluster size to vary across cities and specifications, as detailed in Appendix Table A2. 8 Figure 2 shows that the South African cities differ from other major African cities and other major cities across the world. We limit each city to 1 km. grid cells within 20 km of the delineated business districts. We then place each grid cell in buckets that are between n-1 and n km from the nearest business district. We compute the total population, the average population density, the share of the total population, and the cumulative share of the population. While the cities outside Africa, and other African cities, exhibit negative gradients, with population density decreasing further away from economic density, South African cities are very different. The total population, population density, and the share of the population in each bin remain stable as the distance from the economic density in- creases. Consequently, there are far fewer people close to economic density, and relatively larger proportions of people further from the employment centers. Figures 3 and 4 illustrate that South Africa’s largest cities – Johannesburg, Cape Town, eThekwini (Durban), Tshwane (Pretoria), and Nelson Mandela Bay (Gqeberha) – exhibit smaller population densities near business districts, i.e. in areas that are less than 5 km away from the nearest business district. In contrast, larger proportions of their popula- tions live farther from business districts.15 The five major South African cities have the lowest population densities, along with Zurich, with a mean population density of 5,500 people per square kilometer (the median is 5,700).16 For the other 20 cities, the mean density is 17,200 people per square kilometer (the median is 15,200). Our approach to estimating the gap in population densities close to business districts involves comparing population gradients across the sample of 25 cities. Before discussing the regression estimates, it is worth examining Figure 5, which shows the average popu- lation density for each category in the distance typology. We also use NTL radiance (c.f. Elvidge et al. (2021)) as a proxy for economic activity, which declines with distance as expected—unlike population density in South Africa’s cities. Unlike most cities, several in South Africa show rising or flat population densities with distance—especially between 10–20 km from BDs.17 eThekwini (Durban) does not exhibit large anomalies, but the population densities in the 5-10 km and 10-15 km bands are comparable to the density in the 0-5 km band. Taken together, Figures 3, 4, and 5 suggest that there is a spatial mismatch between population and economic density that is particular to major cities in South Africa. We regress the population density (measured as people per square kilometer) Pd c in i distance category di in city c against the dummies for the categories in our distance typol- 15 Weexclude Mumbai in Figure 3, which has a population density above 25,000 people per square km and is an outlier. 16 There is also a low population density in areas further from business districts in Zurich. 17 Figure A1 in the appendix shows the difference in the decline subtracted from the global average. 9 ogy.18 To simplify notation, these dummies are denoted as di , where i is the upper limit of the distance range (i − 5 < d ≤ i), for i equal to 5, 10, and 15 km. c Pd i = ∑ β di d i + β 0 + ϵc ( di ) (1) i The first column in Table 1 reports the results from Equation (1). The equation esti- mates the typical fall in population density as the distance to the closest business district increases. As expected, across our sample of 25 cities, the average population density is largest for areas within 10 km of the nearest business district. Areas within 5 km of a business district have an average population density of 7,857 people per square kilometer higher than areas between 15 km and 20 km, while areas between 5 km and 10 km have an average population density of 7,341 people per square kilometer higher. The average population densities fall sharply for areas further than 10 km from the closest business district. We estimate separate coefficients for South African cities in the following specification ( Is is the indicator for South African cities): c Pd i = ∑ βdi di + Is + ∑ βsdi di Is + β0 + ϵc(di ) (2) i i The second column in Table 1 reports results from this regression. Globally, there are 10,484 additional people per square km within 0-5 km of the nearest business district compared to areas that are 15-20 km from the nearest business district. South African cities have approximately 2,400 fewer people (i.e. 12,816 - 10,484) per square kilometer within 5 km of the closest business district (relative to the population in the 15–20 km range). We next include city fixed effects in the regression (column 3). The results remain consistent. In South Africa, the coefficients are much smaller for areas within 5 km and between 5 km and 10 km of a business district, consistent with anomalously small relative population densities in areas close to business districts compared to the other cities in the sample. It is possible that South African cities have greater urban sprawl, with more dispersed economic activity that is not concentrated near traditional business districts. To assess this, we control for night-time light (NTL) intensity as a proxy for economic activity (column 4). In the global sample, the coefficients on the distance categories become smaller, and some lose statistical significance, suggesting that population density and NTL intensity are broadly aligned – denser areas also tend to emit more light. However, in South Africa, areas within 5 km of business districts still exhibit large negative anomalies in population 18 The 15 km–20 km range is the reference category. 10 density, even after accounting for luminosity. Specifically, using the band means, there are 9,300 fewer people per square kilometer within 5 km of a business district compared to the 5–10 km range — a reversal of the expected urban density gradient.19 If population densities between 15 km and 20 km of a business district were kept con- stant, the population density within 5 km should be 16,000 people per square kilometer, compared to just 5,500. This would bring it closer to the global mean of 17,200 people per square kilometer. This implies that the population densities within 5 km of business districts in South Africa should be almost three times as large as they were in 2020.2021 This analysis has focused on estimating the unusual divergence between economic and population density in South Africa’s major cities when benchmarked against 20 other cities across 5 continents. In the next section, we shift to analyzing a combination of tax administration data and census data in order to characterize the broader labor market context surrounding each neighborhood. In Section 4, we combine the analysis in sections 2 and 3 with high resolution building and road network data that is used to provide a spatial diagnostic for South Africa’s three largest cities. 3 Unequal Access to Employment Opportunities Section 2 showed that South African cities diverge from global urban structure norms with their densest neighborhoods located far from business districts (BDs). This section tests whether that divergence has consequences for local labor markets. Specifically, we show that these isolated, asset-deprived neighborhoods also experience thinner and more volatile formal local employment. This, in turn, leads to the identification of sub-places in which policy interventions may be most urgent, which we address in Section 4. Using 2011 census data at the sub-place level — an administrative unit roughly com- parable to a U.S. census tract in urban areas — we construct an asset index based on hous- ing, water and sanitation, energy access, durable goods ownership, and education. The index combines five components: Housing (share in modern dwellings), Energy (share us- ing modern lighting and cooking), Water and Sanitation (refuse collection, toilets, piped 19 Basedon descriptive sample means for South African cities, 5,500 people/sq. km within 0–5 km vs 14,800 within 5–10 km (the difference is about 9,300). This comparison is descriptive and does not come from the regression coefficients. 20 Table A1 presents similar results, using population density and night-time luminosity in natural logs. 21 Table A2 presents results from a sensitivity analysis, allowing the cluster size and building volumes to vary. The coefficients remain similar in magnitude across most parameter choices and are statistically significant in the majority of specifications. When the building cutoff is 110,000 cubic meters and the minimum cluster size is 15 100m grid cells, we lose four cities and statistical significance. Overall, Table A2 shows that the results in Table 1 are robust to reasonable changes in the key parameter values used to delineate business districts. 11 water), Durable Goods (ownership of TV, fridge, car, washing machine, computer), and Education (share with tertiary education). We answer this question in three steps. First, we use sub-place-level census data to measure neighborhood deprivation and link each neighborhood (i.e. sub-place) to the employment levels of its surrounding area using spatial tax records.22 Second, we exam- ine how formal employment varies across the asset distribution. Third, we test whether poor neighborhoods face not just lower access, but also greater instability. Across all mea- sures, we find that asset-deprived neighborhoods face thinner, more volatile formal labor markets. We focus on formal employment. Unlike in other developing countries, South Africa’s urban labor market has a small informal sector with weak ties to the formal economy (World Bank (2014), Bhorat et al. (2018)). Shah (2022) estimates that, in 2019, only around 4 percent of South Africa’s working-age population was engaged in own-account work. Based on relationships between GDP and informal employment across a large sample of countries, this number should have been closer to 20 percent in South Africa.23 The low engagement in informal work is largely attributed to a combination of factors, such as apartheid-era segregation policies, inadequate infrastructure, low purchasing power in poor areas, strict urban regulations that discourage informal trade, and high crime levels (e.g., Bhorat et al. (2018), Shah (2022)). 3.1 Data Sources I. Spatial Tax Data on Formal Employment To better understand formal employment in South African cities from a spatial perspec- tive, we use spatial tax data from the South African government (Nell and Visagie (2022)).24 Every city is divided into hexagons that are roughly 5 km square in area which is based on Uber’s hexagonal grid (at aperture 7).25 Local formal employment is estimated for every month from March 2013 until December 2021 for every hexagon using tax adminis- trative data. The intention is that the data represent job density in the relevant hexagon, 22 StatsSouth Africa defines a sub-place as "the second (lowest) level of the place name category, namely a suburb, section or zone of an (apartheid) township, smallholdings, village, sub-village, ward or informal settlement.” The 2022 census data is not considered to be representative at the sub-place level because of a 31 percent non-response rate. 23 Shah(2022) discusses many anecdotal pieces of evidence that may help explain these anomalies. For example, in Johannesburg some of the busiest parts of the wealthier areas do not permit retail trade even though these are sometimes areas that are close to the Townships. 24 https://spatialtaxdata.org.za/. The report written by the government provides insightful analysis. We complement their analysis with a greater focus on the relationship between employment and asset depri- vation. 25 https://www.uber.com/blog/h3/ 12 rather than density elsewhere, although this may be a source of error.26 This adminis- trative data captures formal sector employment, allowing us to spatially benchmark job availability across cities and income levels – an improvement over self-reported surveys or small non-representative samples. II. Census Data The latest census data available at the sub-place level across South Africa’s major cities is from 2011. The data variables we use constitute shares of individuals with certain at- tributes (e.g., the completion of tertiary education) or asset ownership (e.g., ownership of durable goods such as a washing machine) or housing (e.g., share of houses with modern sanitation). If there are too few houses in a sub-place these variables are listed as miss- ing. We link a sub-place to the hexagonal tax data on formal employment by matching a sub-place to a hexagon with which the distances between the centroids are minimized.27 This matters because spatial mismatch involves commuting frictions as well as local job scarcity. 3.2 Local Labor Markets, Asset-Deprivation, and Population Density Do poorer neighborhoods have systematically lower access to nearby formal employ- ment? To answer this we assign sub-places to asset index deciles within each city and link each sub-place to its nearest employment hexagon.28 We estimate Equation (3) below to quantify the magnitudes of these differences in av- erage local employment (Ei ) between the asset deciles (ai ).29 Table 2 provides the results, while Figure 6 presents these results as a graph, in which the magnitudes of the coeffi- cients are transformed to percent differences in local employment relative to the wealthi- est decile. 9 ln( Ei ) = ∑ β a d a d + β 0 + [ I − ρW ] − 1 ϵ i (3) d =1 In Johannesburg, the poorest decile has average local employment that is 87 percent lower than the wealthiest decile. 30 In Cape Town, this number is lower (61 percent), but still re- gressive. The patterns for eThekwini are smaller than Cape Town, while Tshwane exhibits 26 It is possible that a firm is in one hexagon, while it conducts operations that employ people in other hexagons. 27 These minimum distances range from 0 to 1.8 km. 28 More specifically, for every sub-place we find the hexagon with which the distance between both centroids is minimized, using the great circle distance metric to account for the Earth’s curvature. 29 The base category is the wealthiest decile. As Figures A2a–A2c show, asset indices are spatially correlated. We estimate a spatial error model with an inverse-distance spatial weights matrix following Kelejian and Prucha (2010); LeSage and Pace (2021). 30 Since the dependent variable is in logs, this is computed as 100 · [e−2.03 − 1]. 13 a much larger gradient, comparable to Johannesburg. The results in Nelson Mandela Bay (Gqeberha) are comparable to those in Cape Town in that the top six wealth deciles have average local employment numbers that are comparable and higher than the bottom four deciles. These results are based on employment data in 2019.31 It is also relevant to examine whether population density in itself is related to local employment in these cities. Table 3 reports the results from the equation below. 9 ln(PDi ) = ∑ β a d a d + β 0 + [ I − ρW ] − 1 ϵ i (4) d =1 In Johannesburg, the poorest asset decile has a population density that is 274 percent higher than the wealthiest decile. The bottom five asset deciles have population densities that are significantly larger than the wealthiest decile. In Cape Town, the relationship is even more stark. The lowest asset decile has a population density that is 360 percent larger than the wealthiest decile.32 Across all five cities, poorer neighborhoods are surrounded by thinner formal labor markets. In Johannesburg, residents of the poorest decile live near areas with 87% less formal employment than the wealthiest; the gap is 61% in Cape Town. This suggests that spatial mismatch is not only about commuting distances to city centers, but also about structurally thinner employment ecosystems around low-income neighborhoods. This pattern is widely acknowledged in South Africa’s urban policy literature and consistent with previous descriptive studies. Our estimates provide the first quantitative, multi-city confirmation using sub-place-level population and asset data. The analysis so far has concerned total formal employment within the area around a neighborhood. However, asset-deprived populations are often less educated and unlikely to get high-wage jobs even if they live in proximity to such employment. Therefore, in this section, we restrict the analysis to jobs that are at the bottom of the wage distribution, namely that pay wages less than 6,400 (2015 R) per month.33 This covers low-wage tier formal jobs that are still sought after but excludes higher- wage categories. We estimate Equation (5) to quantify the relationship between local em- ployment of lower-wage tier formal labor Ei (as a percentile, E p ) and wealth as measured by asset deciles ad : 31 However, analysis available from the authors shows that the patterns are stable across years. 32 Across all five metros, population density declines consistently with asset wealth, with the poorest deciles in Tshwane, eThekwini, and Nelson Mandela Bay also significantly denser than the wealthiest deciles. 33 Nell and Visagie (2022) state that the minimum wage was just above 3,200 (2015 R) per month, so including this wage-band (i.e., 3,200–6,400 R/month) would include minimum wage earners and those in the wage tier just above. 14 9 Ei = ∑ β a d a d + β 0 + [ I − ρW ] − 1 ϵ i (5) d =1 In Johannesburg, on average the most asset-deprived decile has lower-wage tier local employment that is 34.4 percentile points (pp) lower than the wealthiest decile (Table 4a). The relationship is steeper in Tshwane (43 pp) but slightly weaker in Cape Town and eThekwini (18-20 pp) and more muted in Nelson Mandela Bay. In eThekwini, local employment in the lower-wage tier declines steadily across asset deciles, including among the higher deciles. Even where formal jobs exist near poor neighborhoods, those jobs may be more volatile — creating uncertainty for residents and limiting their ability to develop skills while work- ing. To assess this, we model month-to-month variation in formal employment (control- ling for trend and seasonality), compute the standard deviation of the residual, and com- pute city-specific volatility percentiles. We then regress this volatility percentile against the asset decile. 3 2π f t s 2π f t s ∑ i f f ∆ ln( Et ) = γ0 + γS sin + γC cos + ϵit (6) f =1 12 12 Our measure of employment volatility (Vi ) is the standard deviation of the residual, p and Vi is the volatility in within-city percentiles. We estimate the relationship between local employment volatility (percentiles) and asset index using Equation (7). Table 4b reports the results. 9 ∑ β a d a d + β 0 + [ I − ρW ] − 1 ϵ i p Vi = (7) d =1 Employment is more volatile in poorer neighborhoods across all cities. In Johannes- burg, volatility is 26 percentile points higher in the poorest decile. Together, Tables 4a and 4b, provide evidence that asset-deprived populations across South Africa’s major cities live in neighborhoods in which lower-wage tier formal labor markets are thinner. Asset-deprived neighborhoods have lower levels of local employ- ment as well as higher levels of employment volatility. Both youths and other adults looking for formal employment opportunities would benefit from thicker local labor mar- kets to gain the requisite expertise to get the more sought-after jobs in the traditional downtowns. This section showed that spatial mismatch in South Africa is not limited to physical distance from economic density. Poor, dense neighborhoods — particularly former town- ships — are surrounded by weak local labor markets. They have fewer formal jobs nearby, greater job volatility, and fewer opportunities even in the low-wage tier. These spatial fric- 15 tions exacerbate inequality and reduce the capacity of cities to provide and absorb labor efficiently. While causality remains beyond the scope of this analysis, these results suggest the need for geographically targeted investments to thicken labor markets in under-served neighborhoods or significant improvements in transport infrastructure. To operationalize these insights for local planning, Section 4 introduces a diagnostic framework using high resolution data that identifies sub-places with high density and low employment access, allowing policymakers to spatially target transport infrastructure and labor market inter- ventions. 4 High-Resolution Spatial Mismatch Diagnostics This section identifies spatial mismatch in South Africa’s three largest metros, Johannes- burg, Cape Town, and Durban, using high-resolution data on building footprints, pop- ulation density, and network-based road access to business districts. Unlike Section 2’s global, coarser benchmark, here we switch to city-scale diagnostics with a finer grid, em- pirical business-district delineation from the built form, and OSRM-based road-network travel times. This complements Section 2’s international comparisons and Section 3’s em- ployment analysis by pinpointing densely populated sub-places that are remote, provid- ing evidence of the persistent spatial inequalities established under apartheid and identi- fying neighborhoods that would benefit most from targeted interventions. Specifically, we combine three spatial datasets. Google Open Buildings (GOB) pro- vides building footprint polygons estimated from satellite imagery. We aggregate these to 200-meter grid cells and compute two metrics – the number of buildings per cell and the average building size. Building count density serves as a proxy for overcrowding, particularly in informal settlements, and average building size serves as an indicator of formal development, including commercial activity or high-income housing. Second, the Global Human Settlement Layer (GHSL) provides estimates of population density at a 100-meter resolution. These are resampled to the 200-meter grid and used to measure population density. Third, Open Source Routing Machine (OSRM) is used to compute network-based travel times from each grid cell to the nearest business district (BD), using data from OpenStreetMap. Unlike as-the-crow-flies distance, OSRM estimates travel time along road networks, including gaps, detours, and bottlenecks, making it more suitable for our analysis.34 Business districts are defined using a combination of built form and economic activ- 34 OSRM (Open Source Routing Machine) is a routing engine based on OpenStreetMap road data. It allows us to calculate realistic travel times by car or public road network from any grid cell to a defined destina- tion, rather than relying on as-the-crow-flies distance. This makes it particularly useful in cities in which some neighborhoods may be isolated because of poor road connectivity. 16 ity indicators at a 200 m resolution. For each 200 m grid cell, we construct a business district (BD) score as the standardized sum of (i) the intensity of OpenStreetMap points of interest (POIs) tagged as “office,” “retail,” “commercial,” “government,” “shop,” or “amenity,” and (ii) the average building size, with a negligible (0.01) weight on total built area to avoid double-counting size. POI intensity captures commercial activity, while av- erage building size distinguishes large commercial buildings from densely packed small residential structures.35 4.1 Sub-Place Classification by Density and Access Each sub-place is categorized into five categories based on city-specific medians of pop- ulation density and travel time to business districts: High-Density Isolated (HDI), High- Density Connected (HDC), Low-Density Isolated (LDI), Low-Density Connected (LDC), and Sparsely Populated. For example, a High-Density Isolated sub-place is above the me- dian in population density and above the median in travel time to the closest business district. Figures 7a-c suggest that the spatial organization of South African cities continues to reflect the segregated urban structure designed by apartheid-era planners, which deliber- ately placed Black communities on the urban periphery with limited access to economic opportunity (Thompson and Berat (2014), Simpson (2021)). Soweto (Johannesburg), the Cape Flats (Cape Town), and Umlazi (Durban) are the largest of these former townships. Despite population growth and public investments in infrastructure since 1994, these ar- eas remain characterized by densely packed residential structures and disproportionately long travel times to business districts. In Soweto, average sub-place-level building density ranges from 200 to over 400 struc- tures per 200-meter grid, much higher than Johannesburg’s median of 128 structures. Building footprints remain small, typically between 45 and 65 square meters, compared to the citywide median of 101 square meters. Further, median travel times to business districts in Soweto are 1.5 to 2.5 times the citywide median. This confirms that Soweto remains both housing deprived and disconnected from the city’s employment centers. Similar patterns are identified in the Cape Flats. In sub-places such as Khayelitsha, Nyanga, and Philippi, grid-level building count densities exceed 220, often reaching 300 to 400 per grid, while average building sizes are between 30 and 60 square meters, well below the Cape Town median of 100 square meters. These sub-places also face travel 35 Allspatial layers are transformed to a custom South Africa Albers equal-area projection; distances, areas, and clustering thresholds are computed in meters. Candidate BD cells are those in the top decile of the BD-score distribution. These are clustered using single-linkage on grid-cell centroids with a 500 m cut, and clusters with at least 100 high-score cells are labeled as business districts. (WGS84 is used only for initial bounding-box filtering and ellipsoidal clipping.) 17 times that are 2 to 3 times longer than the city median. Umlazi, Durban’s largest former township, has a built environment that is similar to the other large former townships: high building count densities (typically over 250 per grid) and small average footprints (around 60 square meters). Together with the high population densities, these statistics reflect persistent overcrowding.36 Across all three metros, these township areas consistently rank in the upper quartile for building count density and in the lower quartiles for average building size. The consis- tent mismatch between population concentration and proximity to employment centers in these areas supports the interpretation that initial conditions imposed by apartheid plan- ning continue to shape urban spatial inequality. While this analysis does not make causal claims, the presence of these disparities, decades after apartheid ended, underscores the persistent spatial exclusion experienced by disadvantaged populations that live in these large cities. 4.2 Distance to Economic Opportunity and Asset-Deprivation Figures 8a and 8b illustrate these relationships using non-parametric (loess) plots of these relationships at the sub-place level. Figure 8a shows that the population density gradient is flat, and in fact rises slightly as sub-places move further away from economic density, in contrast to the negative gradients typical of other cities around the world. In contrast the asset index, the share of employed adults, and local labor market characteristics worsen as sub-places are further from business districts. The asset and employment patterns are more muted for Cape Town reflecting the influence of tourism and the presence of greater shares of wealthy retirees in some far flung sub-places. Figure 8b confirms the systemic disadvantages that deprived neighborhoods endure across all three of South Africa’s largest cities. The poorest neighborhoods are further from the nearest business district, are more densely populated, have lower fractions of their adult population employed, and have lower job densities and higher employment volatilities. Together, these relationships provide evidence of local poverty traps within South Africa’s largest and most dynamic cities.37 36 A million hectares of land were re-zoned between 1950 and 1991 to implement racial segregation policies (Christopher (1997)). South Africa’s post-apartheid policies have involved removing squatters living in central areas and moving them to the outskirts of cities. These removals were often justified using the apartheid-era National Building Regulations and Building Standards Act, under the pretext of reducing crime and health risks (Thompson and Berat (2014)). 37 Figures A2a-A2c show the same standardized variables used in Figure 8. Each panel reports within- city, population-weighted percentiles (0–100) for assets, population density, minimum travel time to the nearest business district, and employment, job-density, and job-volatility measures. “Sparse” sub-places are included for completeness and represent areas with population density below 1,000 people per square km. 18 These non-parametric plots also validate and connect our findings across sections 2, 3, and 4 by showing that the relationships consistently tell the same story, with differ- ent types and sources of data. First, we show that even with more accurate measures of distance, and data used to delineate business districts, there is no evidence of declining population densities further from economic density. In fact gradients are inverted from typical patterns. Second, we show that the low job density areas are systematically further away from the nearest business district. Third, we show that the building count densities (the last row in each figure) increase as we move away from the nearest business district, and decrease as the neighborhood’s asset index increases. This is consistent with clusters of densely packed informal, public, or traditional township housing in asset-deprived and isolated neighborhoods. Table 5a quantifies these relationships. In Johannesburg, a 10 percentile increase in distance to the nearest business district is associated with a 3.7 percentile decline in the asset index; in Cape Town and eThekwini, the declines are 4.0 and 4.4 percentiles respec- tively. These patterns hold after controlling for population density. We present similar regressions using the share of working-age adults employed, drawn from the 2011 census (Table 5b). In Johannesburg, employment shares drop by 4.9 percentage points per 10 per- centile increase in distance to the BD; in Cape Town and eThekwini, the effect is 3.7 and 6.5 points respectively. These results indicate that the most asset-deprived neighborhoods in Johannesburg, Cape Town, and eThekwini are both more remote from business districts and more densely populated. In standard urban models, distance from employment centers is typically compensated by lower housing costs or greater access to space and amenities. Across all three cities, the most disadvantaged neighborhoods are both spatially peripheral and economically deprived, inconsistent with a spatial equilibrium characterized by homoge- nous households and limited spatial frictions (Glaeser and Gottlieb, 2009). 4.3 Implications for Urban Investment The combination of building count densities, average building sizes, and relatively larger commute times, provides a practical basis for identifying neighborhoods that are disad- vantaged. High-Density Isolated neighborhoods with small average building sizes repre- sent clusters of informal or low-income housing that remain spatially disconnected. These sub-places would greatly benefit from targeted infrastructure investment, including last- mile public transport, road upgrades, and improved land use rights. In Johannesburg, the Meadowlands sub-places are clear candidates for transport in- vestments that link residential areas to employment centers in Braamfontein and Sand- ton. In Cape Town, infrastructure upgrades in Wesbank and Park Village, such as the 19 extension of BRT feeder routes or improvements to local road surfaces, could significantly reduce spatial barriers to employment. In Durban, investment in shared transport services or new route planning in Pinetown could address gaps in mobility. At the same time, Low-Density Connected neighborhoods with large building sizes and superior infrastructure offer potential sites for mixed housing initiatives, should these be politically feasible. Programs that allow for subdivision, backyard rental formalization, or medium-density infill could increase access without requiring large investments in new building construction. This diagnostic also supports implementation of the objectives of the Spatial Planning and Land Use Management Act (SPLUMA). It identifies neighbor- hoods that score poorly on spatial inequality (Section 7a), are unsustainable (7b), or rep- resent inefficient use of public infrastructure (7c). The typology allows national and local governments to align public investments with quantifiable spatial exclusion, consistent with larger net benefit ratios. The analysis presented in this section builds on the earlier sections by identifying the specific neighborhoods in which spatial disconnection is most pronounced. Unlike ag- gregate comparisons or national maps, the high-resolution mismatch typology developed here provides a bridge between diagnosis and intervention. This has the potential to pro- vide a low cost, but empirical, basis for more targeted planning inquiry and investment appraisal. 5 Conclusion South Africa is becoming increasingly urbanized, with nearly two-thirds of its popula- tion and three-quarters of its economic activity concentrated in urban areas. However, the country has not achieved the expected productivity gains that stem from the agglom- eration economies associated with well-functioning cities. Two factors may explain why urbanization in South Africa yields low dividends. First, urban development tends to be low-density and sprawls horizontally, resulting in cities with lower population densities compared to other global urban areas. Second, limited and costly public transportation exacerbates the distance to economic opportunities, especially for the 80% of the popula- tion without cars. This is particularly challenging for residents of remote, impoverished neighborhoods, who spend two to three hours commuting and half of their daily earnings on transportation. While this issue stems from apartheid-era policies, current efforts have not sufficiently addressed it. This paper presents quantitative evidence of the stark separation between population and economic density in South Africa’s major cities, emphasizing the disproportionate burden of these spatial inefficiencies on poor households. Using remote-sensed building 20 imagery, night-time light data, and census data, we find that population densities near employment centers, such as business districts, are 60% lower in South African cities com- pared to global norms based on the median density of 20 cities worldwide. This spatial mismatch has significant distributional effects, with poor households more likely to live in densely populated but peripheral areas that feature lower formal employment and higher job volatility. This contributes to South Africa’s high unemployment rate, with nearly two-thirds of working-age South Africans, especially those from low-income households, either unemployed or not actively seeking work. While we do not have access to granular housing price data, it is likely that zoning and regulatory constraints restrict the supply of lower-cost housing near employment centers, even in underpopulated areas.38 Therefore, the marginal resident trades off the increase in commuting costs against the decrease in housing costs, artificially inflated near employ- ment centers, and chooses to live further away from work. This consequently shifts the population distribution, raises equilibrium commuting costs, and results in fragmented cities characterized by greater inequality.39 This analysis complements the influential work of Lall et al. (2021), which emphasizes that urban sustainability and growth depend on shifting from “pancake” cities — low- rise horizontal expansion — to “pyramids” marked by vertical and infill development.40 We extend this framework by examining how access to economic density varies across neighborhood types, linking localized socioeconomic conditions to the urban built form. Unlike standard approaches, our multidimensional diagnostics uncover systematic spa- tial inequalities across South African cities and may be applied across other contexts. This paper highlights the urgent need to address spatial inefficiencies in South Africa, which continue to limit agglomeration benefits and reinforce deep-rooted inequalities. In response, post-apartheid governments have introduced urban transport strategies, such as the Integrated Rapid Public Transport Networks (IRPTNs), to expand commuter rail 38 See Jedwab et al. (2020) for a creative approach to measuring and benchmarking these restrictions in sparse data environments. 39 Duranton and Puga (2023) (also see Hsieh and Moretti (2019)) provide theory and evidence for the dead- weight losses from zoning regulations at the city level in the United States. Their analysis abstracts from secondary employment centers, although such polycentric patterns are common. Recent quantitative spatial-equilibrium models can replicate such polycentric urban structures within a unified framework (cf. Redding (2023)), yet these models typically assume well-functioning land, labor, and transport mar- kets. The mechanisms generating polycentricity in middle-income cities remain poorly understood. Our empirical findings highlight the need for richer theoretical frameworks that explicitly model polycentricity and spatial mismatch under such constraints. 40 Combes et al. (2025) find similar patterns in terms of greater horizontal development in Africa in their comprehensive and insightful examination of urban forms across the region. They find that, in 2015, monocentric patterns were more common, though large cities exhibited polycentric structures. However, they define monocentricity in terms of population density distributions, rather than the number of em- ployment centers. 21 and integrate it with bus networks. On the urban development side, policies have focused on increasing affordable housing for disadvantaged communities, typically located on the urban periphery where land is cheaper.41 Yet, after two decades, these strategies have not significantly reduced spatial inefficiencies or urban inequities – pointing to the need for more targeted and empirically grounded interventions. Our analysis does not claim causal identification. However, the spatial mismatch ob- served across three decades points to structural inertia. The high-resolution analysis con- firms that the legacy of apartheid continues to constrain opportunity, despite a shift to democratic governance and the ensuing greater public investments in these areas. This spatial diagnostic identifies neighborhoods most in need of interventions such as major roads and public transport hubs, thereby complementing causal evaluations and plan- ning processes. Transforming South African cities into more dynamic and equitable growth hubs re- quires a combination of policies and investments. Key actions include improving passen- ger rail services, enhancing the efficiency, affordability, and connectivity of private trans- port options like taxi-bus services, and adopting urban planning policies that promote denser, more integrated city centers. A back-of-the-envelope calculation suggests that moving from neighborhoods in the 90th percentile (more remote) to the 50th percentile (more central) in travel time to the nearest business district is associated with a percentile point increase in the share employed of 18 points in Johannesburg and 8 points in Cape Town.42 This methodology provides a scalable and standardized data-driven framework for complex urban systems with sparse data, helping to design and monitor policies that aim to reduce spatial disparities and advance South Africa’s progress towards more in- clusive labor markets. 41 Bertaud (2018) offers a conceptual explanation of how the Reconstruction and Development Programme shaped developer incentives. More broadly, he advocates for policy humility in complex urban systems, where planning efforts often struggle to overcome market frictions and produce intended outcomes. 42 These estimates are based on the coefficients in Table 5b (column 2): −0.46 × (−40) = +18.4 for Johan- nesburg and −0.21 × (−40) = +8.4 for Cape Town. They are approximate magnitudes derived from within-city percentile regressions and should be interpreted as suggestive rather than causal. 22 Figure 1: Construction of Distance Gradients in Johannesburg 2020 23 Figure 2: Population Density Gradients 24 Figure 3: Population Density within 5 km of a Business District Figure 4: Relative (Remote vs Proximate) Population Sizes 25 Figure 5: Declines in Population Density and Night-Time Lights away from Business Districts 26 Figure 6: Local Employment relative to the Wealthiest Decile 27 Figure 7a: Identifying Spatial Mismatch Hotspots: Johannesburg Notes: Travel times estimated using OpenStreetMap routing (OSRM). These reflect optimal routing and likely underestimate actual commute times, especially in congested or underserviced areas. However, relative accessibility patterns remain informative. 28 Figure 7b: Identifying Spatial Mismatch Hotspots: Cape Town Notes: Travel times estimated using OpenStreetMap routing (OSRM). These reflect optimal routing and likely underestimate actual commute times, especially in congested or underserviced areas. However, relative accessibility patterns remain informative. 29 Figure 7c: Identifying Spatial Mismatch Hotspots: Durban Notes: Travel times estimated using OpenStreetMap routing (OSRM). These reflect optimal routing and likely underestimate actual commute times, especially in congested or underserviced areas. However, relative accessibility patterns remain informative. 30 Figure 8a: Distance to the Nearest Business District Gradients Notes: Travel times estimated using OpenStreetMap routing (OSRM). These reflect optimal routing and likely underestimate actual commute times, especially in congested or underserviced areas. However, relative accessibility patterns remain informative. 31 Figure 8b: Asset Index Gradients Notes: Travel times estimated using OpenStreetMap routing (OSRM). These reflect optimal routing and likely underestimate actual commute times, especially in congested or underserviced areas. However, relative accessibility patterns remain informative. 32 Table 1: Population Density Declines by Distance to a Business District (1) (2) (3) (4) Population Density (people per km sq.) Avg. NTL Luminosity 134.286 [0.973] 0 km. < Distance to BD <= 5 km. 7857.008*** 10484.021*** 9357.701*** 6181.908 [3.450] [4.077] [4.650] [1.675] 5 km. < Distance to BD < 10 km. 7341.908*** 9809.740*** 8683.420*** 6886.949*** [3.113] [3.613] [4.381] [2.807] 10 km. < Distance to BD < 15 km. 3355.337** 4426.907*** 3300.588* 2058.327 [2.620] [2.877] [1.913] [1.022] South Africa 1274.844 [0.691] 0 km. < Distance to BD <= 5 km. X S. Africa -12816.354*** -11690.034*** -9356.396*** [-4.421] [-4.845] [-2.875] 5 km. < Distance to BD <= 10 km. X S. Africa -12020.447*** -10894.127*** -9181.891*** [-3.737] [-4.156] [-3.167] 10 km. < Distance to BD <= 15 km. X S. Africa -5039.140** -3912.820* -2800.047 [-2.536] [-1.836] [-1.245] Constant 7297.151*** 6978.440*** [7.628] [5.786] Country Dummy No No Yes Yes Observations 95 95 95 95 R-squared 0.13 0.29 0.83 0.84 Notes: *** p<0.01, ** p<0.05, * p<0.1, Robust t-statistics in brackets. Robust standard errors clustered by city. The base distance category is: 15 km < Distance to BD <=20 km 33 Table 2 : The Relationship between Local Employment and Asset Index Quantiles in 2019 Johannesburg Cape Town Tshwane eThekwini NM Bay VARIABLES Dep. Variable: Ln (Employment (FTE) in 2019) Lowest (Most Deprived) -2.03*** -0.95*** -1.84*** -0.56*** -0.86*** [-10.09] [-4.36] [-6.54] [-2.97] [-3.18] Second -2.06*** -0.58** -2.16*** -0.55** -0.69 [-9.90] [-2.27] [-7.79] [-2.42] [-1.56] Third -1.98*** -0.75*** -1.68*** -0.48** -1.37*** [-9.36] [-3.01] [-7.44] [-2.21] [-3.77] Fourth -1.57*** -0.49** -1.91*** -0.60*** -0.74* [-7.39] [-2.26] [-8.42] [-2.86] [-1.88] Fifth -1.52*** -0.43* -1.82*** -0.73*** -0.86*** [-7.22] [-1.89] [-8.49] [-3.54] [-2.73] Sixth -1.06*** -0.49*** -1.45*** -0.44** -0.94*** [-5.13] [-2.77] [-6.70] [-2.23] [-3.05] Seventh -1.08*** -0.20 -0.83*** -0.17 -0.72** [-5.91] [-1.11] [-4.04] [-0.90] [-2.56] Eighth -0.84*** 0.10 -0.72*** -0.44** -0.25 [-5.37] [0.71] [-3.96] [-2.46] [-1.00] Ninth -0.26* 0.39*** -0.23 -0.29* -0.37* [-1.83] [3.13] [-1.39] [-1.66] [-1.89] Spatial Error (lagged) 1.19*** 2.03*** 1.89*** 3.99*** 2.34*** [173.00] [27.67] [21.67] [6.21] [8.06] Constant 8.22*** 7.79*** 7.64*** 7.10*** 7.44*** [62.18] [75.77] [42.38] [64.29] [56.68] Observations 638 687 360 389 153 *** p<0.01, ** p<0.05, * p<0.1, z-statistics in brackets spatially lagged standard errors. The spatial unit is the sub-place (admin 5). The asset index quantiles are computed separately for each city. 34 Table 3 : The Relationship between Population Density and Asset Index Quantiles Johannesburg Cape Town Tshwane eThekwini NM Bay Asset Index (Within-City) quantiles Dep. Variable: Ln (Population Density) Lowest (Most Deprived) 1.32*** 1.53*** 1.16*** 0.63*** 1.08*** [12.21] [12.93] [8.06] [5.78] [7.50] Second 1.72*** 1.62*** 0.94*** 0.65*** 1.49*** [15.01] [11.89] [6.89] [5.02] [6.62] Third 1.72*** 1.74*** 1.22*** 0.89*** 1.35*** [15.26] [12.75] [9.87] [6.97] [6.80] Fourth 1.49*** 1.35*** 1.18*** 0.69*** 1.26*** [13.20] [11.15] [9.46] [5.38] [6.46] Fifth 1.37*** 1.28*** 1.03*** 0.95*** 1.16*** [12.11] [10.29] [8.35] [7.48] [6.81] Sixth 1.33*** 1.09*** 0.81*** 0.77*** 1.35*** [11.68] [11.14] [6.47] [6.05] [8.06] Seventh 1.07*** 0.84*** 0.71*** 0.67*** 1.14*** [10.67] [8.52] [6.05] [5.40] [7.40] Eighth 0.50*** 0.55*** 0.34*** 0.64*** 0.92*** [5.85] [7.24] [3.19] [5.44] [6.78] Ninth 0.04 0.30*** 0.09 0.24** 0.48*** [0.45] [4.48] [0.96] [2.12] [4.25] Spatial Error (lagged) 1.08*** 2.51*** 1.93*** 1.51*** 1.36*** [42.67] [5.96] [14.80] [13.51] [6.78] Constant 3.17*** 3.35*** 2.89*** 3.04*** 2.80*** [43.68] [69.53] [25.28] [34.49] [22.51] Observations 658 725 407 446 189 *** p<0.01, ** p<0.05, * p<0.1, z-statistics in brackets spatially lagged standard errors. The spatial unit is the sub-place (admin 5). The asset index quantiles are computed separately for each city. 35 Table 4a : The Relationship b/w Low-Wage Employment and Asset Index Quantiles in 2019 Johannesburg Cape Town Tshwane eThekwini NM Bay Asset Index (Within- Dep. Variable: Employment FTE Percentiles: Low-Wage Tier in City) quantiles 2019) Lowest (Most Deprived) -34.34*** -19.89*** -42.71*** -19.53*** -12.46* [-10.34] [-4.57] [-8.71] [-5.13] [-1.80] Second -33.75*** -9.62* -44.75*** -25.31*** -8.29 [-9.74] [-1.86] [-9.47] [-5.72] [-0.84] Third -35.17*** -15.17*** -35.66*** -16.60*** -23.91** [-10.06] [-2.99] [-9.05] [-3.57] [-2.40] Fourth -30.05*** -8.12* -42.31*** -20.18*** -11.64 [-8.53] [-1.82] [-10.68] [-4.58] [-1.09] Fifth -31.91*** -8.47* -41.87*** -24.02*** -11.17 [-9.19] [-1.86] [-11.19] [-5.46] [-1.32] Sixth -21.54*** -9.80*** -36.42*** -12.93*** -15.18* [-6.24] [-2.75] [-9.63] [-2.97] [-1.96] Seventh -19.10*** -3.52 -21.84*** -12.07*** -11.80* [-6.28] [-0.97] [-6.12] [-2.90] [-1.66] Eighth -14.86*** 0.30 -16.31*** -7.30* -6.71 [-5.69] [0.11] [-5.14] [-1.84] [-1.09] Ninth -2.27 7.92*** -3.34 -8.03** -3.44 [-0.94] [3.17] [-1.14] [-2.12] [-0.67] Spatial Lagged Error 1.18*** 1.38*** 1.74*** 3.32*** 2.50*** [155.33] [27.10] [34.15] [8.97] [12.81] Constant 62.92*** 55.50*** 63.91*** 62.45*** 70.83*** [28.58] [23.61] [21.00] [24.98] [18.62] Observations 645 715 362 431 171 *** p<0.01, ** p<0.05, * p<0.1, z-statistics in brackets spatially lagged standard errors. The spatial unit is the sub-place (admin 5). The asset index quantiles are computed separately for each city. 36 Table 4b: The Relationship b/w Local Low-Wage Employment Volatility and Asset Index Quantiles in 2019 Johannesburg Cape Town Tshwane eThekwini NM Bay Asset Index (Within-City) Dep. Variable: Employment Volatility Percentiles: Low-Wage quantiles Tier in 2019 Lowest (Most Deprived) 25.75*** 18.16*** 41.63*** 13.19*** 19.99** [6.24] [4.33] [7.80] [3.10] [2.54] Second 28.35*** 4.05 25.72*** 13.49*** 29.25** [6.59] [0.82] [5.13] [2.68] [2.19] Third 27.37*** 11.47** 31.45*** 13.03*** 33.91** [6.31] [2.37] [7.44] [2.59] [2.57] Fourth 21.81*** 2.99 32.12*** 14.33*** 25.06* [5.11] [0.70] [7.10] [2.92] [1.87] Fifth 25.97*** 1.03 32.78*** 15.63*** 27.60*** [6.07] [0.24] [7.73] [3.22] [2.76] Sixth 16.89*** -0.76 26.27*** 11.66** 23.13** [3.94] [-0.22] [6.22] [2.44] [2.57] Seventh 7.10* -2.49 15.63*** 9.58** 20.45** [1.87] [-0.71] [3.86] [2.08] [2.44] Eighth 7.71** -5.85** 5.87 2.11 7.78 [2.38] [-2.16] [1.63] [0.48] [1.06] Ninth -4.79 -8.97*** 0.30 9.65** 17.14*** [-1.60] [-3.74] [0.09] [2.33] [2.96] Spatial Lagged Error 1.17*** 2.50*** 2.00*** 3.91*** 2.53*** [81.04] [20.79] [7.54] [5.72] [3.94] Constant 41.35*** 41.45*** 36.31*** 38.80*** 29.26*** [15.04] [22.71] [9.93] [14.50] [7.50] Observations 648 716 381 424 156 *** p<0.01, ** p<0.05, * p<0.1, z-statistics in brackets spatially lagged standard errors. The spatial unit is the sub-place (admin 5). The asset index quantiles are computed separately for each city. 37 Table 5a : Spatial Correlates of the Asset Index Johannesburg Cape Town eThekwini Travel Time -0.37*** -0.32*** -0.40** -0.19* -0.44*** -0.44*** (-3.98) (-4.82) (-2.47) (-1.71) (-6.76) (-6.89) Population Density -0.64*** -0.72*** -0.26*** (-7.64) (-11.97) (-4.53) Constant 68.92*** 99.32*** 69.89*** 95.61*** 71.97*** 84.61*** (10.39) (25.52) (9.08) (27.94) (15.01) (18.63) Observations 674 674 749 749 477 477 R-squared 0.14 0.55 0.16 0.63 0.19 0.26 *** p<0.01, ** p<0.05, * p<0.1, Robust standard errors. The spatial unit is the sub-place (i.e. admin 5 units). Each regression uses sub-place level population weights, and clusters standard errors by the main place (admin 4 units). Travel Time is the optimized, friction-less, travel (based on OSRM), to the nearest business district. Table 5b : Spatial Correlates of Share of the Adult Population Employed Johannesburg Cape Town eThekwini Travel Time -0.49*** -0.46*** -0.37*** -0.21*** -0.65*** -0.65*** (-4.44) (-5.43) (-4.38) (-3.46) (-9.80) (-10.59) Population -0.40*** -0.55*** -0.25*** Density (-3.13) (-7.24) (-5.24) Constant 74.88*** 93.65*** 68.34*** 88.00*** 82.55*** 94.97*** (12.21) (12.78) (11.36) (26.04) (18.00) (20.21) Observations 676 676 762 762 477 477 R-squared 0.24 0.39 0.13 0.41 0.42 0.49 *** p<0.01, ** p<0.05, * p<0.1, Robust standard errors. The spatial unit is the sub-place (i.e. admin 5 units). Each regression uses sub-place level population weights, and clusters standard errors by the main place (admin 4 units). Travel Time is the optimized, friction-less, travel (based on OSRM), to the nearest business district. 38 References Alonso, William, Location and land use: toward a general theory of land rent, Harvard university press, 1964. Anas, Alex, Richard Arnott, and Kenneth A Small, “Urban spatial structure,” Journal of economic literature, 1998, 36 (3), 1426–1464. Arribas-Bel, Daniel, M-À Garcia-López, and Elisabet Viladecans-Marsal, “Building (s and) cities: Delineating urban areas with a machine learning algorithm,” Journal of Urban Economics, 2021, 125, 103217. Banerjee, Abhijit, Sebastian Galiani, Jim Levinsohn, Zoe McLaren, and Ingrid Woolard, “Why has unemployment risen in the new South Africa?,” Economics of Transition, 2008, 16 (4), 715–740. Barza, Radu, Edward L Glaeser, César A Hidalgo, and Martina Viarengo, “Cities as engines of opportunities: evidence from Brazil,” Technical Report, National Bureau of Economic Research 2024. Bastos, Paulo and Nicolas Bottan, “Resource rents, coercion, and local development: Evidence from post-apartheid South Africa,” World Development, 2023, 169, 106289. Bertaud, Alain, Order Without Design: How Markets Shape Cities, MIT Press, 2018. and Stephen Malpezzi, “The spatial distribution of population in 48 world cities: Implications for economies in transition,” Center for urban land economics research, University of Wisconsin, 2003, 32 (1), 54–55. Bhorat, Haroon, Ben Stanwix, and Amy Thornton, “Changing Dynamics in the South African Labour Market,” 2021. , Zaakhir Asmal, Kezia Lilenstein, and Kirsten Van der Zee, SMMEs in South Africa: Under- standing the constraints on growth and performance, DPRU, University of Cape Town Cape Town, 2018. Bryan, Gharad, Edward Glaeser, and Nick Tsivanidis, “Cities in the developing world,” Annual Review of Economics, 2020, 12 (1), 273–297. Burger, Martijn J, Evert J Meijers, Marloes M Hoogerbrugge, and Jaume Masip Tresserra, “Bor- rowed size, agglomeration shadows and cultural amenities in North-West Europe,” European planning studies, 2015, 23 (6), 1090–1109. Christopher, Anthony John, “Racial land zoning in urban South Africa,” Land Use Policy, 1997, 14 (4), 311–323. Combes, Pierre-Philippe and Laurent Gobillon, “The empirics of agglomeration economies,” in “Handbook of regional and urban economics,” Vol. 5, Elsevier, 2015, pp. 247–348. , Clément Gorin, Shohei Nakamura, and Mark Roberts, “An anatomy of urbanization in sub- Saharan Africa,” Regional Science and Urban Economics, 2025, 115, 104152. Donaldson, Dave and Adam Storeygard, “The view from above: Applications of satellite data in economics,” Journal of Economic Perspectives, 2016, 30 (4), 171–98. 39 Duranton, Gilles and Diego Puga, “Micro-foundations of urban agglomeration economies,” in “Handbook of regional and urban economics,” Vol. 4, Elsevier, 2004, pp. 2063–2117. and , “Urban growth and its aggregate implications,” Econometrica, 2023, 91 (6), 2219–2259. Duval, Romain, Yi Ji, and Ippei Shibata, “Labor market reform options to boost employment in South Africa,” International Monetary Fund Working Paper WP/21/165, 2021. Elvidge, Christopher D, Mikhail Zhizhin, Tilottama Ghosh, Feng-Chi Hsu, and Jay Taneja, “An- nual time series of global VIIRS nighttime lights derived from monthly averages: 2012 to 2019,” Remote Sensing, 2021, 13 (5), 922. Fay, Marianne and Charlotte Opal, Urbanization without growth: A not so uncommon phenomenon, Vol. 2412, World Bank Publications, 2000. Friedman, Jerome, Trevor Hastie, and Robert Tibshirani, The elements of statistical learning 2nd Ed., second ed., Springer series in statistics Springer 2012 reprint, Berlin, 2001. Glaeser, Edward L and Joshua D Gottlieb, “The economics of place-making policies,” Technical Report, National Bureau of Economic Research 2008. and , “The wealth of cities: Agglomeration economies and spatial equilibrium in the United States,” Journal of economic literature, 2009, 47 (4), 983–1028. Gobillon, Laurent and Harris Selod, “Spatial mismatch, poverty, and vulnerable populations,” Handbook of regional science, 2021, pp. 573–588. , , and Yves Zenou, “The mechanisms of spatial mismatch,” Urban studies, 2007, 44 (12), 2401– 2427. Hausmann, Ricardo, Andrés Fortunato, Alexia Lochmann, Kishan Shah, Lucila Venturi et al., “Growth Through Inclusion in South Africa,” Technical Report 2023. Henderson, J Vernon and Matthew A Turner, “Urbanization in the developing world: too early or too slow?,” Journal of Economic Perspectives, 2020, 34 (3), 150–173. Hsieh, Chang-Tai and Enrico Moretti, “Housing constraints and spatial misallocation,” American economic journal: macroeconomics, 2019, 11 (2), 1–39. Jedwab, Remi and Dietrich Vollrath, “Urbanization without growth in historical perspective,” Explorations in Economic History, 2015, 58, 1–21. , Jason Barr, and Jan K Brueckner, “Cities without skylines: Worldwide building-height gaps and their implications,” Technical Report, CESifo Working Paper 2020. Kain, John F, “Housing segregation, negro employment, and metropolitan decentralization,” The quarterly journal of economics, 1968, 82 (2), 175–197. Kelejian, Harry H and Ingmar R Prucha, “Specification and estimation of spatial autoregressive models with autoregressive and heteroskedastic disturbances,” Journal of econometrics, 2010, 157 (1), 53–67. Kloosterman, Robert C and Sako Musterd, “The polycentric urban region: towards a research agenda,” Urban studies, 2001, 38 (4), 623–633. 40 Lall, Somik, Mathilde Lebrand, Hogeun Park, Daniel Sturm, and Anthony Venables, “Pancakes to pyramids: City form to promote sustainable growth,” 2021. LeSage, James P and R Kelley Pace, “Interpreting spatial econometric models,” in “Handbook of regional science,” Springer, 2021, pp. 2201–2218. Manysheva, K, M Mestieri, and J Schauer, “End of Apartheid, Not of Inequality: The Slow Tran- sition in a Segregated Economy,” CEPR Discussion Paper No. 20472, 2025. Marx, Benjamin, Thomas Stoker, and Tavneet Suri, “The economics of slums in the developing world,” Journal of Economic perspectives, 2013, 27 (4), 187–210. Melchiorri, Michele, Aneta J Florczyk, Sergio Freire, Marcello Schiavina, Martino Pesaresi, and Thomas Kemper, “Unveiling 25 years of planetary urbanization with remote sensing: Perspec- tives from the global human settlement layer,” Remote Sensing, 2018, 10 (5), 768. Mills, Edwin S, “An aggregative model of resource allocation in a metropolitan area,” The Ameri- can Economic Review, 1967, 57 (2), 197–210. Moretti, Enrico, “Local labor markets,” National Bureau of Economic Research Working Paper 15947, 2010. Nell, A and J Visagie, “Spatial Tax Panel 2014–2022 [dataset],” 2022. Redding, Stephen J, “Quantitative urban models: from theory to data,” Journal of Economic Per- spectives, 2023, 37 (2), 75–98. Shah, Kishan, “Diagnosing South Africa’s High Unemployment and Low Informality,” CID Re- search Fellow and Graduate Student Working Paper Series, 2022. and Federico Sturzenegger, “Search, transport costs, and labor markets in South Africa,” CID Research Fellow and Graduate Student Working Paper Series, 2022. Shilpi, Forhad, Lu Xu, Raisa Behal, and Brian Blankespoor, “People on the move: Spatial mis- match and migration in post-apartheid South Africa,” World Bank Urbanisation Review Paper, 2018, 1. Simpson, Thula, History of South Africa: From 1902 to the present, Penguin Random House South Africa, 2021. Taubenböck, Hannes, Ines Standfuß, Michael Wurm, Angelika Krehl, and Stefan Siedentop, “Measuring morphological polycentricity-A comparative analysis of urban mass concentrations using remote sensing data,” Computers, Environment and Urban Systems, 2017, 64, 42–56. Thompson, L and L Berat, “A history of South Africa,” A History of South Africa, Fourth Edition, 2014, pp. 1–437. Thompson, Leonard Monteath, A history of South Africa, Yale University Press, 2001. World Bank, Economics of South African townships: special focus on Diepsloot, World Bank, 2014. 41 Online Appendix 42 Figure A1: Anomalies in the Declines in Population Density and Night-Time Lights 43 Figure A2a: Spatial Distributions of Wealth, Jobs, and Population: Johannesburg 44 Figure A2b: Spatial Distributions of Wealth, Jobs, and Population: Cape Town 45 Figure A2c: Spatial Distributions of Wealth, Jobs, and Population: eThekwini (Durban) 46 Table A1: Population Density Declines by Distance to a Business District (natural logs) (1) (2) (3) (4) VARIABLES Population Density (Natural Logs) Avg. NTL Luminosity (Natural Logs) 0.015 [1.212] 0 km. < Distance to BD <= 5 km. 0.762*** 1.042*** 1.010*** 0.660 [3.556] [4.388] [4.414] [1.605] 5 km. < Distance to BD < 10 km. 0.690*** 0.961*** 0.928*** 0.730** [3.187] [4.014] [4.075] [2.367] 10 km. < Distance to BD < 15 km. 0.453** 0.601*** 0.569*** 0.432* [2.765] [2.990] [2.906] [1.958] South Africa 0.356 [1.332] 0 km. < Distance to BD <= 5 km. X South Africa -1.315*** -1.282*** -1.025** [-4.596] [-4.607] [-2.715] 5 km. < Distance to BD <= 10 km. X South Africa -1.264*** -1.231*** -1.043** [-3.737] [-3.741] [-2.768] 10 km. < Distance to BD <= 15 km. X South Africa -0.653** -0.621** -0.498* [-2.650] [-2.569] [-1.992] Constant 8.689*** 8.600*** [53.605] [41.293] Country Dummy No No Yes Yes Observations 95 95 95 95 R-squared 0.17 0.35 0.78 0.80 *** p<0.01, ** p<0.05, * p<0.1, Robust standard errors clustered by city. The base distance category is: 15 km. < Distance to BD <=20 km. 47 Table A2: Sensitivity Analysis (Robustness tests for main parameter in Table 1) 0 km. < Distance to BD <= 5 km. X S. Africa Cluster Size threshold Building Cut-off Coefficient t-stat 5 80000 -6597.387*** [-2.878] 10 80000 -7410.472*** [-2.968] 15 80000 -8261.381*** [-2.800] 5 90000 -9639.942** [-2.716] 10 90000 -9903.186*** [-3.035] 15 90000 -10281.945*** [-3.679] 5 100000 -10340.015*** [-2.942] 10 100000 -9356.396*** [-2.875] 15 100000 -7928.111*** [-3.088] 5 110000 -7067.521** [-2.783] 10 110000 -6237.549** [-2.542] 15 110000 -3103.205 [-1.096] Notes: *** p<0.01, ** p<0.05, * p<0.1, Robust standard errors clustered by city. The base distance category is: 15 km. < Distance to BD <=20 km. 48