WPS7994 Policy Research Working Paper 7994 For India’s Rural Poor, Growing Towns Matter More than Growing Cities John Gibson Gaurav Datt Rinku Murgai Martin Ravallion Poverty and Equity Global Practice Group March 2017 Policy Research Working Paper 7994 Abstract It is theoretically ambiguous whether growth of cities mat- poverty measures for rural India, calibrated to a panel ters more to the rural poor than growth of towns. This of 59 regions observed four times over 1993–2012. The paper empirically examines whether growth of India’s expansion of lit area had greater effect on the rural poverty secondary towns or big cities mattered more to recent measures than did intensive margin growth in the bright- rural poverty reduction, noting that data deficiencies ness of light from urban areas. For India’s current stage of have made this a difficult question to answer previously. development, growth of secondary towns may do more to Satellite observations of night lights are used to measure reduce rural poverty than big city growth, although the urban growth on the extensive and intensive margins theoretical model suggests that cities may eventually take in the context of a spatial Durbin fixed-effects model of over from towns as the drivers of rural poverty reduction. This paper is a product of the Poverty and Equity Global Practice Group. 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://econ.worldbank.org. The authors may be contacted at rmurgai@worldbank.org. 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 For India’s Rural Poor, Growing Towns Matter More than Growing Cities John Gibson, Gaurav Datt, Rinku Murgai and Martin Ravallion1 JEL Codes: O15, R12 Keywords: Cities, luminosity, poverty, secondary towns, spatial Durbin model, India 1 Gibson is with the Department of Economics, University of Waikato, Datt is with the Department of Economics, Monash University, Murgai is with the World Bank and Ravallion is with the Department of Economics, Georgetown University. These are the views of the authors and need not reflect those of their employers. We are grateful to participants in the Secondary Towns, Jobs and Poverty Reduction Conference at the World Bank for their helpful comments, and to Geua Boe-Gibson for the preparation of the maps. I. Introduction There appears to be a broad consensus among development economists that agricultural growth, and rural development more broadly, is good for rural poverty reduction (although this was not always widely accepted).2 Models of the development process have also attached importance to the scope for rural poverty reduction through urban economic growth, and some observers have seen this as the more important channel for rural poverty reduction.3 Urban economic growth is expected to contribute to reducing rural poverty through two main channels: (i) Labor absorption: an expanding urban economy will benefit the rural poor by either absorbing surplus rural labor, as in the classic Lewis (1954) model, or by tightening rural labor markets (leading to higher wage rates). (ii) Backward linkages: growth in the urban economy increases public or private resources that benefit the rural poor; for example, greater urban demand for rural products may increase rural incomes or labor-augmenting technical progress in urban areas may increase the remittances sent back to rural families. The strength of these channels has been an important issue for setting development priorities for India, as elsewhere. The evidence suggests that India’s urban economic growth in the post-Independence period up to around 1990 did rather little to reduce rural poverty, although urban growth had reduced urban poverty, and rural poverty was primarily driven by rural growth (Ravallion and Datt, 1996). Since economic reforms began in earnest in India in the early 1990s, there has been considerable progress in reducing poverty, with trend rates of decline that are higher than in the pre-reform era (Datt, Ravallion and Murgai, 2016). The indications are that urban economic growth since the early 1990s has been more poverty reducing, and that this has come with larger gains to India’s rural poor (Datt and Ravallion, 2011; Datt et al., 2016). Lanjouw and Murgai (2014) conjecture that the link from urban development to rural poverty reduction is stronger if urban economic growth stems from India’s secondary towns rather than from the big cities. The secondary towns may be more tightly connected to the surrounding rural hinterland than are the cities, so growth in small towns may have more effect on rural poverty. Yet it is the big cities, defined as those with population above one 2 For an overview of past debates on the relationship between economic growth (including by sector) and poverty reduction see Ravallion (2016, Chapter 8). 3 See, for example, Collier (2009), with reference to Africa. 2 million, that have the lowest poverty rates and that appear to be growing faster than the smaller statutory towns, with the share of the urban population in the big cities rising from 38 percent in 2001 to 42 percent in 2011 (Tripathi, 2013). Higher wage rates in larger cities will to some degree spill over to the towns and rural areas both through labor market adjustment and because they may also generate larger trade and remittance flows. Thus, it is theoretically ambiguous whether larger cities generate larger gains to the rural poor. The Lanjouw and Murgai hypothesis that India may have experienced faster poverty reduction if smaller towns had grown as fast as the cities is consistent with evidence from other countries on the relationship between poverty and city size (Ferré, Ferreira and Lanjouw, 2012). However, it is difficult to test this hypothesis for India, or more generally to relate variation in growth of different types of cities to variation in rural poverty reduction. One difficulty arises because it is only once every ten years that city growth (in terms of population rather than economic output) is measured in India, using the census. A lot of the variation in rural poverty reduction occurs within a ten-year censual period and so would be missed by studies that rely on the census data to measure urban growth. Another difficulty is the absence of timely and spatially detailed (e.g. at city level) economic statistics. This study uses a proxy for urban growth to test the hypothesis that it is the growth rates of India’s secondary towns, rather than the big cities, that matter most for rural poverty reduction. We use nighttime lights, as detected by satellite, to measure urban growth, distinguishing between growth on the extensive and intensive margins, and between the growth of cities and of secondary towns. These new measures of urban growth are econometrically related to sub-national poverty estimates that are formed at a finer spatial resolution than in the existing literature. Specifically, we use a division of India into 59 National Sample Survey (NSS) regions that are more finely-grained than the usual division into states and union territories. Our study covers four observations for each of these regions between 1993/94 and 2011/12, based on NSS “thick” rounds (with larger sample sizes such that the survey is representative at the NSS regional level). We also account for the spatial autocorrelation that is increasingly apparent in patterns of rural poverty in India. The following section provides a simple theoretical model of a three-sector labor market in which one of the sectors—the “big city”—has a labor market distortion, but wages are flexible in the other two sectors, the secondary towns and the rural hinterland, with workers free to move between the two. For this model, we derive conditions under which technical progress in the big cities has less impact on the rural wage rate than does technical progress in the secondary towns. However, this is only one possible outcome. Even in this 3 simple model, city growth could more effectively “trickle down” to the rural poor. It is an empirical question as to which type of urban growth is better for the rural poor. Section III describes our data for addressing that question, in which we have formed a regional panel data set, combining results from household surveys with data on the extent of night light. Section IV explains our econometric model, which is calibrated to the panel data. Alternative models are also described, and shown to be testable restrictions on our preferred (encompassing) specification. Our results are then presented in Section V, which provide strong support for the hypothesis that economic growth in secondary towns has more impact on rural poverty than does growth of the big cities. Section VI concludes. II. A Simple Theoretical Model The purpose of the following model is to illustrate one source of urban-rural linkage, namely through the labor market, for which urban economic growth emanating from cities brings different gains to the rural poor than growth in towns. We suppose that the urban economy comprises two sectors, a town and a city. In the spirit of the classic Harris-Todaro (1970) model of rural-urban migration in the presence of an urban labor-market distortion, we assume that the wage rate in the city is fixed above the market clearing level, but the wages in the town and the rural economy are fully flexible, and come into parity. Technical progress in the town is assumed to increase the marginal product of labor there, which will lead to higher wages in the town, and also in the rural hinterland due to the integrated labor markets; indeed, in equilibrium the wage gains will be the same. Technical progress in the cities will increase employment there, which will attract workers out of unemployment and from both the town and rural areas. This will bring gains to wages in the latter sectors. In the context of this model, we ask whether technical progress in the town has more impact on poverty than that in the city. There is no inequality within sectors (although this can be relaxed to assume an inequality-neutral growth processes). The poor in this model are taken to be all workers except those who get a job in the city (i.e., rural plus township workers, plus the urban unemployed). In other words, the poverty line is below the city wage rate but above the rural and town wage rates. In more formal terms, the model is as follows: The production functions are Ai Fi ( N ie ) (i=c,t,r) for the city, town and rural areas respectively, where N ie denotes employment in sector i and Ai is the shift parameter for (Hicks-neutral) technical progress; we will only 4 consider the comparative static effects of changes in Ac and At so we set Ar  1 . All three production functions are strictly increasing and strictly concave. The respective wage rates are Wi for i=c,t,r. These are all taken to be normalized by the poverty line. The town and rural wage rates are flexible, such that all those who want work can find it; in equilibrium, Wt  Wr . The city wage rate is fixed, such that N c  N ce are left unemployed and they are assumed to earn nothing (where N c is the city workforce, including the unemployed). In equilibrium, the rural wage rate is equated with the expected wage rate in the city (the probability of getting a city job times the city wage rate), Wr  ( N ce / N c )Wc . Firms maximize profits, requiring that wage rates equate with marginal products, Wi  Ai Fi( N ie ) (i=c,t,r), with (variable) wage elasticities of labor demand denoted i ( 0) . Total population is normalized at unity ( N c  N te  N re  1 ). We now consider the effects on rural poverty of technical progress in the town versus the city. Since all rural workers are taken to be poor, we will only consider impacts on the rural poverty gap index (PG), which is the mean distance of the rural wage rate below the poverty line. Since PG  1  Wr in our model, we focus solely on the rural wage rate.4 Proposition 1: Technical progress in the town will have a larger (smaller) proportionate impact on the rural wage rate if the ratio of the city workforce (employed plus unemployed) to the town’s workforce is greater than (less than) the ratio of the wage elasticity of the town’s labor demand to that of city labor demand. To verify this claim, consider the effects on the rural wage rate of an increase in Ac and compare this to the effect of an increase in At . On log differentiating and solving (invoking the usual implicit function theorem) we obtain:5  ln Wr t N te  0 (1.1)  ln At  r N re  t N te  N c  ln Wr c N c  0 (1.2)  ln Ac r N r  t N te  N c e For the economy as a whole, PG  (1  W )(1  N c ) where W  Wr ( N r  Nt ) /(1  N c ) is the mean wage 4 e e e e rate of the poor in the economy as a whole. As long as the workforce-weighted mean labor demand elasticity in the town and rural sectors is less than unity, the direction of change in the overall poverty gap also depends on the change in the rural wage rate. 5 Note that the following derivatives are not symmetric since it is the total city workforce (including the unemployed) in the numerator of (1.2). This feature stems from the fixity of the city wage rate. An addendum is available from the authors with more detail on the derivation. 5 Growth in either urban sector reduces the rural poverty gap. The ratio of the two proportionate effects on rural wages is:  ln Wr  ln Wr c N c  (2)  ln Ac  ln At t N te Thus Proposition 1 follows. So even in this simple model we cannot predict which urban sector will be the most rural-poverty reducing. When city size is small ( N c approaches zero) technical progress in that sector will bring negligible gains to the rural poor via the labor market. Alternatively, even when the city workforce is as large as that in the secondary towns, economic growth in the cities will bring less gain to the rural poor if labor demand is less wage elastic in the cities than the towns. But these are special cases, and one can readily find counter-examples, such as when the city workforce is either relatively large or its labor demand is more wage elastic. Proposition 1 also suggests that we may well find that town growth tends to matter more to the rural poor at early stages of overall economic development, assuming that this entails that cities expand (in number or size). But in due course, this will change, such that cities become the stronger driver of rural poverty reduction. This is implied by the above model if the wage elasticities of labor demand change little as the share of the workforce living in cities grows, with that in towns falling. III. Data and Descriptive Statistics 3.1 Poverty Data Our poverty data are based on estimates of real household consumption that are measured in four “thick” rounds of household surveys conducted by the National Sample Survey Organization (NSSO). These “thick” rounds each have a sample size of over 100,000 households as opposed to the more frequent “thin” rounds. We use data from 1993/94 (50th round), 2004/05 (61st round), 2009/10 (66th round), and 2011/12 (68th round). The NSSO surveys began in the early 1950s and in terms of the international norms for consistency of expenditure or consumption surveys they have maintained a high degree of comparability over time, with the exception of the late 1990s (Datt et al, 2016).6 For the four surveys that we use, a mixed-recall period (MRP) is used, with a 30-day recall for food and frequently 6 There is debate about the consumption and poverty estimates from the 55th round (1999/00) that is described in detail by Deaton and Kozel (2005). We do not use the results from this round. 6 consumed items, and a one-year recall for clothing, footwear, health, education, and durable goods. The monthly per capita consumption estimates are put in spatially and temporally real terms by using India’s official poverty lines as deflators.7 We use two poverty measures. The first is the headcount index, given by the percentage of the population living in households with consumption per capita less than the poverty line. The second is the poverty gap index, given by the mean distance below the poverty line expressed as a proportion of that line, where the mean is formed over the entire population (counting the non-poor as having zero poverty gap). In keeping with our focus on how urban economic growth has impacted rural poverty we confine attention to the rural poverty measures.8 Much of the research on poverty in India is based on a disaggregation into the urban and rural sectors of each state or union territory (UT). Here we use instead more finely grained data, based on “NSS Regions”. These regions have been used by the NSSO since the 4th round, where disaggregation below the state level was first available. Initially, 52 regions were formed, where a region was defined as a group of contiguous districts from within the same state that had similar geographical features and had similar population densities in the 1951 population census. With the subsequent splitting of states and districts, the number of NSS regions grew to 88 by the 68th round survey in 2011/12. We only use data for 19 major states, and we also control for state- and district-level splitting by anchoring our panel to the administrative geography of India as of the 55th round survey in 1999/00. Consequently, the number of regions we have poverty estimates for is 59, with these available for four years, giving a balanced panel of N=236. The pattern of change in rural poverty over time is shown in Table 1. The headcount index halved over the two decades considered, falling from 50 percent in 1993/94 to 25 percent in 2011/12. The poverty gap index declined even faster, with a fall from 0.123 in 1993/94 to 0.045 in 2011/12. The reductions in poverty rates are spread fairly evenly between the four survey years, but since there is a ten-year gap between the first two surveys and then gaps of four years and two years, it is clear that the pace of poverty reduction is increasing. 7 These are the “Tendulkar” poverty lines, after Suresh Tendulkar who headed an expert group that revised India’s official poverty lines. Among several revisions were a smaller rural-urban cost of living differential than the original Planning Commission poverty lines. At 2011 purchasing power parity (PPP) the Tendulkar poverty lines have a value of $2.09 per day. 8 The NSSO follows the Census in defining areas as urban, based on criteria in terms of minimum population, minimum population density, and the percentage of the population working outside of agriculture, or whether an area is a Statutory town (e.g., has a Municipality, Corporation, Cantonment Board, etc.). 7 The NSS is a weighted survey but the weights do not affect the estimated downward trends in poverty rates in Table 1. This is reassuring because the spatial panel estimators that we use in the econometric analyses below do not handle variable sampling weights. We therefore rely on the unweighted data; based on Table 1 this should provide similar patterns to the weighted data. The other pattern in Table 1 is the regional divergence in poverty rates; the coefficient of variation of rural poverty rates based on the NSS regions rises monotonically over time, for both poverty measures and whether using sampling weights or not. In other words, even though India has been escaping from mass absolute poverty it is doing so at an uneven rate over space. The spatial pattern of rural poverty at baseline is shown in Figure 1, for the headcount index (left panel) and poverty gap index (right panel). The spatial autocorrelation in the data is apparent, with regions having high poverty likely to be neighbors with other high poverty regions and vice versa for regions with low poverty rates. Roughly speaking, if one drew a straight line from Nashik in Maharashtra (which is about 100 miles northwest of Mumbai) that went at a 50 degree angle to pass west of Kanpur in Uttar Pradesh (UP), all but three regions to the southeast of that line would be in the highest three poverty classes, which had headcount poverty rates of 42 percent or more in 1993/94.9 Conversely, almost all regions along the Arabian Sea coast, and in northwest India were in the lowest classes for poverty. A common test for spatial autocorrelation is Moran’s I test, which has some parallels with the Durbin-Watson statistic.10 This test reveals significant positive spatial autocorrelation with I=0.30 for the headcount index in 1993/94 and I=0.26 for the poverty gap index. The spatial pattern of rural poverty at the endpoint of our panel, in 2011/12, is even more apparent than it was a baseline. In the left panel of Figure 2 a belt of high headcount poverty is apparent, which extends from the northeast in Assam and comes west across India at a slight downward angle to reach the Eastern Gujarat region, located on the Arabian Sea coast about midway between Mumbai and Ahmedabad. The Moran statistic for the headcount index in 2011/12 is 0.43 while for the poverty gap index it is 0.38. These increases in the Moran statistic from the baseline show that rural poverty has become increasingly spatially 9 The three exceptions were the Central Plains of West Bengal (WB) which surround Kolkata, Coastal Tamil Nadu (TN) and Inland Southern Andhra Pradesh (AP). 10 For any variable z in deviation from mean form and spatial weights matrix W, Moran’s I is equivalent to the slope coefficient in a linear regression of Wz on z (Anselin, 1988). In other words, it examines the strength of the relationship between one observation and the spatially weighted average of its neighboring observations. The weights matrix used here is based on contiguity; any region sharing any boundary point with another region is considered its neighbor. These are known as Queen weights, based on where pieces move on a chessboard. 8 concentrated. This growing inequality over space in the poverty rate also featured in the rise over time in the coefficients of variation for the poverty statistics in Table 1. The description of poverty patterns over time and space is completed by Figure 3, which shows the proportion of baseline poverty eliminated by 2011/12. The spatial patterns are especially clear when considering the poverty gap index in the right panel of Figure 3. The southern states of Kerala, Karnataka, TN, and AP, and all of Maharashtra except the Coastal region (where the poverty gap index increased) eliminated from 60-95 percent of the baseline poverty gap index in the rural sector and similar progress is seen in much of northwest India. However, in Assam, Bihar, Chhattisgarh, Jharkhand, Orissa, and parts of Madhya Pradesh, a much smaller proportion of rural poverty was eliminated. Thus, the rate of reduction in poverty is also spatially auto-correlated, with Moran statistics of 0.30 for the proportion of the rural poverty gap eliminated and 0.15 for the proportion of the rural headcount poverty that was eliminated. In other words, rural poverty started out as spatially concentrated, the poverty reduction process has also been spatially uneven, and so rural poverty at the end-line for our data is even more spatially concentrated than it was at the start. 3.2 Night Lights Data If India were China it would be fairly easy to examine how growth of different types of cities affects rural areas. China has annual GDP data for prefectures, a sub-provincial unit a little more disaggregated than NSS regions are with respect to India’s states, and also for each city and county. One could compare how rural poverty is affected by growth in big cities versus growth in smaller ones, such as county-level cities.11 However, the available sub-national economic statistics are much more limited for India. For example, annual data on Gross District Domestic Product are reported by the Planning Commission, but exclude some big states like Gujarat, and for many states the data end in 2005/06 and so miss the recent period of rapid poverty reduction. The Population Census for 1991 and 2011 roughly lines up with the 50th and 66th or 68th round NSS data but not with the 61st round, and does not inform about economic growth. The lack of agglomeration-level boundaries, like the Metropolitan Statistical Areas (MSA) in other countries, also means that India’s census data correspond to administrative rather than to economic definitions of cities. 11 Deng et al (2010) are an example of studying patterns of urban change in China using these spatially detailed GDP data. 9 To measure urban growth, we therefore rely on data that are not limited by India’s statistical infrastructure, namely night-time lights, as detected by satellite. These data are originally from the Defence Meteorological Satellite Program (DMSP) and are processed by the National Oceanic and Atmospheric Administration. They provide annual estimates over 1992-2012 at a one square kilometer resolution (in some years two satellites are in orbit so there are 33 satellite-year observations). Gibson et al (2015) use these data to estimate rates of area expansion over 1992-2012 for 47 Indian agglomerations with a population of at least one million in the 2011 census.12 These agglomerations are shown in Figure 4, in terms of their lit area in 2012. We treat these 47 agglomerations as India’s “big cities” and contrast the effects of their growth with the effects of the growth of smaller towns.13 The threshold of one million population is also used by Chatterjee et al (2015) and Li and Rama (2015) to demarcate between India’s large and small urban areas. Figure 4 also shows the NSS regions; 39 of 59 have one or more of the big cities in them. For the others, the city effects come indirectly, with these spatial spillovers captured by our econometric model. We consider city growth on the extensive and intensive margins, since it is a priori unclear which matters more to poverty reduction. For example, extensive margin growth may offer easier commuting from rural areas but more sprawling cities may be less efficient. Specifically, we use the night lights to both demarcate the edge of cities for examining effects of expanded areas (the extensive margin), and to calculate the intensity of light—the “digital number” (DN)— emanating from within these boundaries (the intensive margin).14 There is also a measurement reason for distinguishing between lit area and light output within an area. Some DMSP satellites distribute measured light less intensely over a wider area and others focus it into a smaller area.15 For example, Liu et al (2012) note that 12 The census lists 53 million-plus cities but two are within the Delhi agglomeration (Faridabad and Ghaziabad) and four others (Malappurum, Srinigar, Kollam and Kannur) were too dim to be detected at the 50% threshold used for the other cities. At the 20% luminosity threshhold we use to distinguish towns from rural areas, these four cities contribute just 0.4% of total lit area (and 1.3% of big city lit area, if measured at that threshold). 13 The area expansion rate of the 23 cities that were million plus in 1991 was not significantly different from the others that crossed the million threshhold between 1991 and 2011, indicating that there is no evident selection or survivorship bias by defining the sample of big cities by their meeting a criteria set at the end of the period. 14 Technically, we measure areas by starting at the center of each big city, where lights should be brightest, and as the algorithm moves outwards and comes across pixels less illuminated than the brightness threshold it searches in a different direction. If the algorithm finds no contiguous pixels with light above the threshold except those closer to the city center that it has already scanned over, it sets a boundary. 15 To briefly digress, DMSP was to detect clouds, for daily weather forecasts to help Air Force pilots, rather than to detect ground level activity to help economists. Photons from the observed area enter a scanner and create a pulse of electrons that map to a DN ranging from 0-63; this range was because on-board data holding could spare just six bits of memory (26=64) for each datum (Abrahams et al 2015). Given the original purpose, sensors on board DMSP satellites were adjusted over the lunar cycle to keep the brightness of cloud tops the same, with 10 satellite F15 in orbit in 2008 gave an average DN value of 16.6 across 0.9 million lit pixels in China, while F16 in orbit in the same year gave an average DN value of only 10.7 but spread over 1.7 million lit pixels. Similarly, Gibson et al (2015) show that satellites F14 and F15 differ by 85 km2 (11 percent of the two-satellite mean), when estimating area for Bangalore in 2001. A related issue is that night-lit activity with a particular DN value on one day may not map to the same DN on another day due to sensor adjustments and inter-satellite differences. We respond to these measurement issues in four ways.16 First, we use an annual composite of non-ephemeral lights so unrecorded adjustments may wash out over the course of the year as repeated observations converge to some average amplification level. Second, we average over all satellite-year observations in a two-year window centered on the timing of the NSS rounds to ameliorate errors due to inter-satellite differences. Third, we measure smaller cities using two different luminosity thresholds, of 20 percent and 30 percent of the maximum (equivalent to DN values of 13 and 19), to check that the patterns we find are not sensitive to a particular choice of DN values.17 Finally, we use the decomposition of night lights into lit area effects and average DN values within lit areas. We note in passing that most studies using DMSP data are not from economics and are typically concerned with urban area rather than with average DN values, while studies in economics typically just use average DN values. In order to distinguish between cities and towns, and between towns and rural areas, we need luminosity thresholds. Several are used in practice, with examples of some close to what we use including Small and Elvidge (2013), who use DN12 for their Asia-wide study of urban areas, and Alvarez-Berrios et al (2013), who use DN20 for intermediate urban areas in Latin America. Lower thresholds are needed for towns than are used for cities because even some million-plus agglomerations were undetectable at a 50 percent threshold (DN>31). Conversely, a low threshold for all urban areas clumps big cities together along lit no record kept of these adjustments and produce data that are top-coded at DN=63. Also, comparability between satellites and sensor decay were less important for daily cloud monitoring than for measuring long-term changes in ground-level lights. 16 Various ‘calibrated’ night lights series are also available but these typically involve extrapolations from spatially and temporally limited episodes when amplification levels on the sensors were adjusted to suit the needs of researchers rather than of Air Force pilots (Hsu et al, 2015) or extrapolations from areas such as Sicily that were assumed to have stable lights (Elvidge et al, 2009). These assumptions may not be broadly acceptable. 17 We refer to percentages of the maximum DN value to make the point that it is relative luminosity that is being measured, rather than the DN being a constant unit of measurement like a kilometer. For big cities, we use a single threshold of 50 percent because Gibson et al (2015) show the robustness of temporal and spatial patterns to using alternative thresholds of 40 and 60 percent. Cross-validation exercises also supported results using the 50 percent threshold, which contrasts with findings from Abrahams et al (2015) who found greatly exaggerated city sizes if night lights are used without correcting for various features causing blurring. 11 corridors; for example, with a 20 percent threshold one big agglomeration extends 500 km along India’s National Highway 1 from Delhi through Chandigarh and Ludhiana all the way to Amritsar. This clumping results in implausibly large cities and reduces the number of them separately distinguishable, yielding a smaller sample for examining variation in different types of urban growth. Our approach to distinguishing growth in cities from towns is shown in Figure 5, using the example of three of the big cities in southern India. In 1993, the Inland Southern region of Karnataka, which is where Bangalore is located, had about 900 km2 of urbanized area, if using a luminosity threshold of 30 percent.18 This was split roughly equally between areas of Bangalore that were at or above 50 percent of the maximum brightness, and areas elsewhere in the region that were between 30 and 50 percent of the maximum. Some of these less bright areas include the outskirts of Bangalore and we discuss our treatment of fringe areas of big cities below. The map on the right of Figure 5 shows that by 2011 the area of Bangalore had almost quadrupled, using the 50 percent threshold to demarcate this big city from other urban areas. The rate of expansion of the secondary towns in this NSS region was even faster, so the total urban area of the region (using the 30 percent threshold to separate urban from non-urban) was almost five times what it was in 1993. In the adjacent region of Inland Tamil Nadu, the city of Coimbatore grew almost as fast as Bangalore; the annual trend expansion rate was 4.6 percent compared to 4.9 percent for Bangalore (Gibson et al, 2015). However, the growth in less brightly lit areas of Inland Tamil Nadu was even faster, and so between 1993 and 2011 Coimbatore contributed just 12 percent of the total urban area expansion for the Inland Tamil Nadu region. In contrast, Bangalore contributed 35 percent of the total urban area expansion for the Inland Southern Karnataka region. At the heart of our empirical strategy is the fact that for some regions the urban growth has mainly come from a brightly lit big city like Bangalore while for other regions the urban growth has come mainly from the expansion of less brightly lit secondary towns, as in Inland Tamil Nadu. Across all the regions we study, the typical pattern is more like Inland Tamil Nadu than like Inland Southern Karnataka; a larger share of the growth in total lit area was due to secondary towns, which contributed about two-thirds of the growth from 1993/94 to 2004/05 and about four-fifths thereafter. Finally, note that a less rapidly growing big city, such as Thissur in Kerala, provides another contrast.19 While the limited 18 The map in the left-hand panel of Figure 5 truncates part of the Inland Southern region, to focus on big cities. 19 Across all 47 big cities the trend annual expansion rates ranged from -1.6 to 8 percent and the coefficient of variation over cities in expansion rates was 0.7 (Gibson et al, 2015) so there is a lot of variability to exploit. 12 expansion of this city is apparent in Figure 5 it is part of an NSS region (Southern Kerala) that has three big cities (Kochi and Thiruvananthapuram are the other two) and so provides one of the few examples of big city multi-polar growth. The approach to splitting big city growth from smaller town growth in Figure 5 raises a question of whether the dimmer fringe of a city should be considered as part of that city or as a separate, smaller, town. Such areas could be existing, distinct, towns that get engulfed by the expanding big city or they may be new parts of the city built on virgin land that are yet to be as brightly lit as the center of the city. To check if our method of putting dim areas on the edge of the city into the secondary towns component of total urban growth affects the results, we created another data set by using a ‘fixed mask’ approach. (This terminology is because our algorithm masks the area covered by the big cities when it calculates statistics on urban area and average DN values for the secondary towns within an NSS region.) Specifically, we take the area covered by each city when it is at its maximum extent, which is in 2012, and exclude that area when smaller town area and the average DN values in smaller towns are calculated for any of the satellite-years. In other words, land that eventually gets engulfed in the city is treated as never being part of the area that small towns could occupy and is reserved as only available for the city even if the city has not yet grown into that area at the time of a particular satellite-year observation. In contrast, our main results use what we call a ‘variable mask’ where the boundary between big cities and smaller towns changes satellite- year by satellite-year. IV. The Econometric Model Spatial autocorrelation in levels of, and changes in, poverty was noted in Section 3.1, and our econometric modeling recognizes this by using the spatial panel estimator of Belotti et al (2014). The regional N-vector of poverty measures for date t (=1,…,T) is denoted Pt and the matrix of explanatory variables is X t which includes our measures of urban economic activity as indicated by the night lights data. Our spatial Durbin model (SDM) can be written as:20 Pt  WPt  X t 1  WX t  2    vt (i=1,…,N; t=1,…,T) (3) 20 For an excellent overview of the spatial Durbin model see Elhorst (2012). Note that the model is static. A temporal lag of Pt can also be included. We found that such a temporal lag of the NSS region poverty rate was statistically insignificant so we dropped it from the model. This allows us to keep the first year of our data. 13 Here the spatial weighting matrix W describes the structure of spatial relationships between the NSS regions. The W matrix has zeros along the main diagonal, given that no region is its own neighbor, while (to allow for geographic spillover effects) the off diagonals are set to unity for immediate neighbors and zero otherwise. This model allows for changes in an explanatory variable in a particular region to not only affect the poverty rate in that region, but also in surrounding regions. The error term in (3) has two components. The term  represents time-invariant regional fixed effects. The second error term, vt , is assumed to be a white-noise process, orthogonal to the X’s. The fact that we have regional fixed effects gives us greater confidence in assuming exogeneity in that threats to validity could be expected to come in large part from latent factors that are relatively constant over time. For example, higher latent agricultural productivity in a region could simultaneously cause lower rural poverty and higher urban living standards. This could happen through urban commercial activity responding to rural demand induced by higher agricultural productivity. The regional fixed effect should soak up the bulk of the variance due to such factors. However, we acknowledge that the exogeneity assumption can still be questioned. Probably the most plausible threat is that temporal (positive or negative) shocks to local agricultural productivity may induce changes in urban activity levels, making urban areas more or less bright. We will maintain the exogeneity assumption but make some observations that offer some support. Two other possible specifications are the spatial autoregressive model (SAR) and the spatial error model (SEM). Both are nested within the SDM. The SAR model is:21 Pt  WPt  X t     vt (4) The SAR is obtained from the SDM under the testable restriction on equation (3) that β2=0. The restrictions to get the SEM from the SDM are analogous to the common factor (COMFAC) restrictions of Hendry and Mizon (1978), where a static model with AR(1) errors is a restricted version of an autoregressive, distributed lag model ADL(1,1). The SEM takes the form: Pt  X t     vt with vt  Wv t   t (5) This can be derived from the SDM under the nonlinear parameter restriction that 2   1 . The SDM will give unbiased coefficient estimates even if the true data-generation process is a SAR or SEM but the reverse is not true; for example, imposing data-inconsistent COMFAC 21 The SAR model may also have autoregressive disturbances, but these may be an artefact of mis-specifying the spatial lags by omitting the lagged X variables in equation (3). 14 restrictions by estimating a SEM when the true model is SDM involves omitting relevant variables. For either poverty measure, and irrespective of using either lit area or the sum of lights (lit area times the average brightness within lit areas), the two sets of restrictions to derive the SAR and SEM from the SDM are rejected at the p<0.01 level. So, our results focus on the more general SDM specification in (3). A feature of the SDM is that the total effect of changes in an X variable – such as expansion in cities or towns – may be quite different to what is shown by ˆ since a local change in poverty rate affects the poverty rates of neighbors which, in turn, affects the poverty rate of their neighbors, including the original region. These spillover and feedback effects let us decompose effects of urban growth on rural poverty into direct and indirect components. To see how, note first that equation (3) can also be written as (in matrix notation and dropping t subscripts): P  ( I  W ) 1 ( X1  WX 2 )  ( I  W ) 1 v (6) Following Elhorst (2012), the partial derivatives w.r.t. the kth explanatory variable can then be written as (noting that the diagonal elements of W are zero): P  ( I  W ) 1 ( 1k I N   2kW ) (7) X k (Here 1k is the kth element of the vector 1 and similarly for  2 k .) The first RHS term is the direct effect and the second is the indirect effect. The indirect effects are of two types: local effects arise from a region’s neighbors and occur as long as  2 k  0, while global effects need not arise just from neighbors, and occur as long as   0. The direct and indirect effects can vary over regions, so the spatial panel estimator that we use follows LeSage and Pace (2009) in reporting a single direct effect, that averages the diagonal elements of the matrix in (7) and a single indirect effect that averages the row sums of the non-diagonal elements of that matrix. The indirect effects are important for three reasons. First, influences on rural poverty may be stronger from surrounding regions than the own-region, depending on the geography of labor market linkages and the variation over space in stages of urban development. Also, the indirect pathway allows big cities to have a potential effect on poverty of NSS regions that lack a big city. Finally, the decomposition provides a straightforward way to deal with big cities that sprawl over region boundaries; for example, in Figure 5, Bangalore extends across the boundary into Inland Tamil Nadu in 2011. The effect of big city lights within the 15 NSS region are part of the direct effect and the lights from the part of a city that has spilled over the border are part of the indirect effect that comes from neighbors.22 V. Types of Urban Growth and Rural Poverty 5.1 Effects of Unmasked Regional Lights on Rural Poverty We start by examining whether night lights have any effect at all, before we compare effects of lights coming from big cities with those from smaller towns. We refer to these results as ‘unmasked’ since we consider the entire area of each NSS region without first masking the pixels lit by the big cities. Table 2 reports the estimates of the SDM for the headcount index, H and the poverty gap index, PG that consider four different ways to use the night lights data. In columns (1) and (5), the total lit area for each NSS region, for any clusters of lights exceeding 20 percent of maximum brightness (DN=13) is used as the proxy for urban development. Variables in these models are in logarithms for the convenience of using elasticities.23 The results suggest an absolute elasticity of rural poverty with respect to own-region lit urban area of about 0.1 to 0.2, with a slightly more precisely estimated elasticity for H than for PG. The elasticities with respect to the spatially weighted average of lit area in neighboring regions are larger, at between 0.3 and 0.4, and the spatial lag of the poverty rate has an elasticity of 0.3. Allowing for the spillover to neighbors and the feedback effects shown in equation (7), the total effect of growth in lit area on H has an elasticity of -0.6 and on PG has an elasticity of almost -0.9. These indirect effects will have both local and global components, given that  2 k  0 and   0. When the average DN value is used, either solely in columns (2) and (6), or jointly with lit area in columns (3) and (7), it has no statistically significant association with the poverty gap index, and a positive relationship with rural headcount poverty (albeit one that is not significant when lit area is also in the model). In contrast to the mixed and imprecise 22 Chandigarh and Delhi lit area includes Union Territory (UT) area that is not part of our data set of 19 major states. The omitted area averages 7.6% of all big city area. In sensitivity analyses we apportion the UT area into the neighboring NSS regions, according to each region’s share of the Delhi or Chandigarh lit area agglomeration in the particular year. The differences in effect of big city growth and smaller town growth on rural poverty are somewhat more precisely estimated in these analyses, while the qualitative pattern of results is unchanged. 23 When urban growth is split into big city and smaller town components not all regions have a big city and the logarithm of zero is undefined. To address this problem, we use the inverse hyperbolic sine transformation whereby ihs( x )  ln( x  ( x  1) ) which is shown in Appendix Table 1 to give identical results to using 2 0.5 logarithms. 16 evidence for average DN values, the significant negative effect of lit area on rural poverty is almost the same regardless of whether average DN values are included in the model. As noted in Section IV, our assumption that night lights are exogenous to rural poverty can be questioned. Recall that long-run cross-sectional differences that would otherwise be a threat to identification will be picked up by the fixed effect. The main remaining threat to identification is likely to be shocks to agricultural productivity that spill over into urban areas. However, one would expect to see brightness being more closely associated with rural poverty reduction than is lit area. Yet we find the opposite.24 The last results reported in Table 2 (columns (4) and (8)) use the sum of lights within each region, by multiplying lit area by the average DN value within lit areas; by comparing results in column (4) with (1) and column (8) with (5), it is clear that just using lit area is sufficient to capture this combined effect. In other words, it is the expansion of India’s cities and other urban areas on their extensive margin that seems to have the most significant relationship with rural poverty reduction. We use this finding to guide our specifications for comparing the effects of growth of the big cities with that of the smaller towns, which we proxy for by using either the lit area or the sum of lights.25 5.2 Direct Effects of City and Town Expansion on Rural Poverty The results in Table 2 consider the lit area (or the sum of lights) within an NSS region, irrespective of whether it is part of a big city or smaller town. In Table 3 we distinguish growth in lit area (or in the sum of lights) into the component from big cities and the component from smaller towns. The instantaneous elasticity of the rural headcount index with respect to the lit area of smaller towns within an NSS region is -0.13, and for big cities it is -0.04. Total elasticities are higher, both direct and indirect, with the latter being larger (in absolute value). Specifically, the total effect shows an elasticity of H with respect to smaller town lit area of -0.55 (with a standard error of 0.09) while the elasticity with respect to big city area is a statistically insignificant -0.09. For the poverty gap index the elasticity w.r.t. secondary town area is even larger, at -0.85, than for H, reflecting gains below the poverty line, while the elasticity w.r.t. big city area is insignificantly different from zero. A statistically significantly larger town 24 A related sensitivity analysis used time fixed effects, which reduced the precision of the results somewhat while not causing much change in the magnitude of the elasticities. 25 The same pattern is also clear, although a little less precisely estimated, if the 30 percent threshold is used to distinguish urban areas from non-urban areas (Appendix Table 1). 17 effect than big city effect is apparent in the test statistics reported at the foot of Table 3, especially for the poverty gap index and in the indirect and total effects. These differences between big city effects and smaller town effects are not an artefact of less brightly lit fringe areas of big cities being counted as part of the smaller, more dimly lit, towns. If a fixed mask approach is used that excludes areas that ultimately become part of the big city from being ever counted as part of small towns, the results are largely the same (columns (3) and (4)). The elasticities have slightly smaller magnitude than under the variable mask approach in columns (1) and (2) but the dominance of small town over big city effects is just as apparent. Another sensitivity analysis is to use the sum of lights approach, by multiplying lit area by the average DN within that area and the results in columns (5) to (8) confirm significantly larger secondary town effects. The results also hold if the distinction between smaller towns and rural areas is based on a 30% luminosity threshold rather than the 20% threshold used in Table 3 (Appendix Table 2 has these sensitivity analyses). It might be argued that the “town effect” we identify may be in large part an indirect “city effect,” i.e., that city growth drives town growth, which then helps drive rural poverty reduction. Table 4 takes this argument seriously by providing analogous results to Table 3 but dropping the variables for towns and only reporting results for H since the results have similar patterns to those for PG. (For balance, the table also gives the results when we drop the city lights variables.) We find a total effect of the area of city lights of -0.715. Combined with the results of Table 3, this suggests that 87% of the total big city effect is via induced town growth, and 13% is direct. However, it should also be noted that the R2 has fallen from 0.631 to 0.546 in comparing columns 1 of Tables 3 and 4. This suggests there is a sizeable share of the variance that may be a direct effect of town growth, independently of city growth. Moreover, the tests of the nesting restrictions needed to derive the ‘big city only’ or ‘small town only’ models show that the models that drop the variables for smaller towns are not consistent with the data, whereas one would not reject the restrictions needed for dropping the big city variables. To summarize, our key finding is that the lights data indicate that the effects of urban economic growth on rural poverty are almost entirely attributable to towns rather than cities. 5.3 Distributional Effects of City and Town Expansion Do the effects we have seen on rural poverty measures entail significant systematic changes in the extent of inequality within rural areas? A simple test for such effects is to 18 replace the dependent variable in our SDM in equation (3) by estimates of the Gini index of consumption inequality from the same NSS data. Table 5 provides the results. Whether we use illuminated area or total light, or variable or fixed masks, we find no sign of significant effects on inequality. The effects of urban growth on rural poverty appear to be transmitted entirely via growth in mean rural consumption. This result does not differ between secondary towns or cities. VI. Conclusions The scope for escaping rural poverty through urban economic growth has been a longstanding development issue, going back to the classic model of Lewis (1954). We have revisited this issue, focusing specifically on the question of whether cities or towns are better generators of rural-poverty reducing growth in India, using data on 59 regions observed four times from 1993/94 to 2011/12. The rural headcount poverty rate fell by half in this period and the poverty gap index fell by two-thirds. The pace of poverty reduction was spatially uneven. Using night lights as the indicator of urban economic activity, we find that the geographic expansion of urban areas is associated with lower rural poverty measures; these effects are consistently found using different specifications, while there is little apparent effect on the rural sector from the brightness of lights coming from urban areas. It is urban growth on the extensive rather than the intensive margin that matters for rural poverty reduction. When the effects of big-city growth on rural poverty are compared to those of secondary towns, some consistent associations emerge, indicating that the growth of towns matters far more than cities to reducing rural poverty in India. This effect is close to inequality-neutral, in that it was not associated with higher, or lower, inequality within rural areas. In expectation, the poverty effect is transmitted through mean consumption. We remind readers that here we have only studied this question for India, at its stage of economic development. The pattern we have uncovered may not hold at all stages of development. Indeed, our theoretical model suggests that cities may eventually take over from towns as the drivers of rural poverty reduction. For now, however, India should look to its towns, not its cities, to help reduce rural poverty, on top of promoting agricultural and rural development. 19 References Abrahams, Alexei, Nancy Lozano-Gracia, and Christopher Oram (2015) Correcting Over- glow in Night-time Lights Data. Unpublished manuscript. Accessed at: https://sites.google.com/site/alexeiabrahams2/ Álvarez-Berríos, Nora, Isabel Parés-Ramos, and Mitchell Aide (2013) Contrasting patterns of urban expansion in Colombia, Ecuador, Peru, and Bolivia between 1992 and 2009. Ambio 42(1): 29-40. Anselin, L. (1988) Spatial Econometrics: Methods and Models Dordrecht: Kluwer Academic Publishers. Belotti, Federico, Gordon Hughes, and Andrea Piano Mortari (2014) XSMLE: Stata module for spatial panel data models estimation. Statistical Software Components Boston College. https://ideas.repec.org/c/boc/bocode/s457610.html Chatterjee, Urmila, Rinku Murgai, and Martin Rama (2015) Job opportunities along the rural- urban gradation and female labor force participation in India. World Bank Policy Research Working Paper No. 7412. Collier, Paul (2009) Africa’s Organic Peasantry. Harvard International Review October 26, http://hir.harvard.edu/agricultureafrica-s-organic-peasantry/ Datt, Gaurav and Martin Ravallion (2011) Has India’s economic growth become more pro-poor in the wake of economic reforms? World Bank Economic Review, 25(2): 157-189. Datt, Gaurav, Martin Ravallion, and Rinku Murgai (2016) Growth, Urbanization and Poverty Reduction in India. National Bureau of Economic Research Working Paper No. w21983. Deaton, Angus, and Valerie Kozel (2005) Data and dogma: the great Indian poverty debate. The World Bank Research Observer 20(2): 177-199. Deng, Xiangzheng, Jikun Huang, Scott Rozelle, and Emi Uchida (2010) Economic growth and the expansion of urban land in China. Urban Studies 47(4): 813-843. Elhorst, J. Paul (2012) Dynamic spatial panels: models, methods, and inferences. Journal of Geographic Systems 14: 5–28. Elvidge, Christopher, Daniel Ziskin, Kimberly Baugh, Benjamin Tuttle, Tilottama Ghosh, Dee Pack, Edward Erwin, and Mikhail Zhizhin (2009) A fifteen year record of global natural gas flaring derived from satellite data. Energies 2(3): 595-622. 20 Ferré, Celine, Francisco Ferreira and Peter Lanjouw (2012) Is There a Metropolitan Bias? The relationship between poverty and city size in a selection of developing countries. World Bank Economic Review 26(3): 351-382. Gibson, John, Geua Boe-Gibson and Glen Stichbury (2015) Urban land expansion in India 1992-2012. Food Policy 56(1): 100-113. Harris, John, and Michael Todaro (1970) Migration, Unemployment and Development: A Two Sector Analysis. American Economic Review 40: 126-142. Hendry, David, and Grayham Mizon (1978) Serial Correlation as a Convenient Simplification, Not a Nuisance: A Comment on a Study of the Demand for Money by the Bank of England. Economic Journal 88(351): 549-563 Hsu, Feng-Chi, Kimberly Baugh, Tilottama Ghosh, Mikhail Zhizhin, and Christopher Elvidge (2015) DMSP-OLS Radiance Calibrated Night-time Lights Time Series with Intercalibration. Remote Sensing 7(2): 1855-1876. Lanjouw, Peter and Rinku Murgai (2014) Urban growth and rural poverty in India: 1983-2005. in N. Hope, A. Kochar, R. Noll, and T.N. Srinivasin (editors) Economic Reform in India: Challenges, Prospects, and Lessons pp. 371-404, Cambridge University Press, New York. LeSage, James and Robert Kelley Pace (2009) Introduction to Spatial Econometrics Chapman and Hall and CRC Press, Boca Raton. Lewis, Arthur (1954) Economic Development with Unlimited Supplies of Labor, Manchester School of Economic and Social Studies 22: 139-191. Li, Yue and Martin Rama (2015) Households or locations? Cities, catchment areas and prosperity in India. World Bank Policy Working Paper No. 7473. Liu, Zhifeng, Chunyang He, Qiaofeng Zhang, Qingxu Huang, and Yang Yang (2012) Extracting the dynamics of urban expansion in China using DMSP-OLS nighttime light data from 1992 to 2008. Landscape and Urban Planning 106(1): 62-72. Ravallion, Martin (2016) The Economics of Poverty: History, Measurement, and Policy. New York: Oxford University Press. Ravallion, Martin and Gaurav Datt (1996) How Important to India's Poor is the Sectoral Composition of Economic Growth? World Bank Economic Review 10:1-26. Small, Christopher, and Christopher Elvidge (2013) Night on Earth: Mapping decadal changes of anthropogenic night light in Asia. International Journal of Applied Earth Observation and Geoinformation 22(1): 40-52. 21 Tripathi, Sabya (2013) Do large agglomerations lead to economic growth? Evidence from urban India. Review of Urban and Regional Development Studies 25(2) 176-200. Van de Walle, Dominique, Martin Ravallion, Vibhuti Mendiratta and Gayatri Koolwal (2016) Long-Term Gains from Electrification in Rural India, World Bank Economic Review. 22 Table 1: Trends in Rural Poverty Rates, India 1993/94 to 2011/12 Unweighted Weighted By Rural Population Headcount Poverty Headcount Poverty Gap Rate Gap Index Rate Index 1993/94 0.495 0.123 0.504 0.125 0.29 0.44 0.26 0.39 2004/05 0.417 0.094 0.422 0.094 0.36 0.56 0.33 0.48 2009/10 0.310 0.062 0.337 0.069 0.52 0.70 0.45 0.59 2011/12 0.245 0.045 0.256 0.047 0.58 0.77 0.51 0.66 Note: Based on averages for 59 NSS regions from 19 major States, using the Tendulkur poverty line. Coefficient of variation in italics. 23 Table 2: Effects of Unmasked Regional Lights (at 20% threshold) on Rural Poverty Rates for 59 NSS Regions from 1993/94 to 2011/12 ---------------Headcount Poverty Rate--------------- ------------------Poverty Gap Index------------------ (1) (2) (3) (4) (5) (6) (7) (8) Lit area in the NSS region -0.149 -0.151 -0.188 -0.192 (1.75)* (1.76)* (1.59) (1.61) Average DN value within lit areas 0.070 -0.129 -0.350 -0.687 (0.16) (0.32) (0.59) (1.21) Sum of lights (lit area × average DN) -0.154 -0.216 (1.88)* (1.91)* W × Lit area in the NSS region -0.267 -0.271 -0.391 -0.413 (2.79)*** (2.82)*** (2.94)*** (3.09)*** W × Average DN value within lit areas 1.168 -0.122 1.945 -0.089 (1.71)* (0.18) (2.07)** (0.10) W × Sum of lights (lit area × average DN) -0.272 -0.384 (2.93)*** (2.97)*** Spatial lag of poverty rate (delta) 0.318 0.738 0.315 0.314 0.333 0.752 0.320 0.322 (4.01)*** (18.53)*** (3.95)*** (3.94)*** (4.23)*** (19.46)*** (4.00)*** (4.04)*** R-squared (within) 0.629 0.124 0.630 0.630 0.638 0.117 0.643 0.642 Restrictions to nest spatial lag model (β2 = 0) 7.80*** 2.92* 8.01** 8.57*** 8.64*** 4.28** 9.57*** 8.84*** COMFAC to nest spatial error model (β2 = - β1) 20.50*** 4.14** 20.63*** 22.18*** 22.29*** 4.15** 23.34*** 23.30*** Total effects – direct Lit area in the NSS region -0.175 -0.178 -0.227 -0.232 (2.56)** (2.58)*** (2.41)** (2.43)** Average DN value within lit areas 0.445 -0.110 0.195 -0.668 (1.12) (0.25) (0.35) (1.09) Sum of lights (lit area × average DN) -0.180 -0.254 (2.74)*** (2.79)*** Total effects – indirect Lit area in the NSS region -0.435 -0.444 -0.642 -0.665 (5.35)*** (5.61)*** (5.67)*** (6.08)*** Average DN value within lit areas 4.512 -0.227 6.592 -0.437 (1.88)* (0.27) (1.87)* (0.37) Sum of lights (lit area × average DN) -0.441 -0.633 (5.56)*** (5.75)*** Total effects Lit area in the NSS region -0.610 -0.621 -0.869 -0.897 (13.21)*** (12.54)*** (13.26)*** (13.06)*** Average DN value within lit areas 4.958 -0.337 6.787 -1.105 (1.91)* (0.38) (1.78)* (0.89) Sum of lights (lit area × average DN) -0.621 -0.887 (13.29)*** (13.54)*** Notes: All variables in logs; all models include fixed effects for 59 NSS regions; nesting tests are chi-sq with 1 df (2 df for col (3), (7)); z-statistics in ( ), ***, **, ** for p<0.01, 0.05, 0.1. N=236. Table 3: Effects of City Growth and Town Growth on Rural Poverty Rates for 59 NSS Regions from 1993/94 to 2011/12 -------------------Illuminated Area------------------- ------------Sum of Lights (area × average DN)------------ Variable Maska Fixed Maskb Variable Maska Fixed Maskb Headcount Poverty Gap Headcount Poverty Gap Headcount Poverty Gap Headcount Poverty Gap (1) (2) (3) (4) (5) (6) (7) (8) Smaller town lit area (or sum of lights) -0.127 -0.197 -0.095 -0.128 -0.114 -0.196 -0.078 -0.114 (1.49) (1.66)* (1.13) (1.10) (1.45) (1.79)* (0.99) (1.05) City lit area (or sum of lights) -0.041 0.000 -0.038 0.002 -0.056 -0.050 -0.053 -0.047 (0.97) (0.01) (0.90) (0.04) (2.10)** (1.35) (1.96)* (1.24) W × Town lit area (or sum of lights) -0.245 -0.358 -0.241 -0.381 -0.268 -0.363 -0.269 -0.399 (2.42)** (2.52)** (2.50)** (2.82)*** (2.87)*** (2.76)*** (2.98)*** (3.15)*** W × City lit area (or sum of lights) -0.020 -0.015 -0.015 -0.003 0.013 0.026 0.017 0.035 (0.27) (0.15) (0.20) (0.03) (0.27) (0.40) (0.36) (0.55) Spatial lag of poverty rate (delta) 0.316 0.339 0.317 0.333 0.323 0.336 0.322 0.330 (3.96)*** (4.32)*** (3.93)*** (4.18)*** (4.07)*** (4.26)*** (4.01)*** (4.11)*** R-squared (within) 0.631 0.637 0.630 0.637 0.634 0.641 0.633 0.641 Total effects – direct Town lit area (or sum of lights) -0.151 -0.235 -0.118 -0.165 -0.140 -0.234 -0.103 -0.153 (2.18)** (2.45)** (1.73)* (1.75)* (2.20)** (2.65)*** (1.62) (1.72)* City lit area (or sum of lights) -0.040 0.003 -0.037 0.007 -0.055 -0.047 -0.051 -0.042 (0.86) (0.05) (0.78) (0.10) (1.76)* (1.07) (1.61) (0.95) Total effects – indirect Town lit area (or sum of lights) -0.398 -0.612 -0.379 -0.606 -0.430 -0.616 -0.415 -0.623 (3.88)*** (4.22)*** (3.99)*** (4.54)*** (4.82)*** (4.91)*** (4.97)*** (5.33)*** City lit area (or sum of lights) -0.050 -0.026 -0.042 -0.007 -0.009 0.012 -0.002 0.027 (0.48) (0.18) (0.40) (0.05) (0.14) (0.12) (0.02) (0.28) Total effects Town lit area (or sum of lights) -0.548 -0.848 -0.496 -0.771 -0.570 -0.851 -0.518 -0.775 (5.85)*** (6.30)*** (5.74)*** (6.28)*** (7.32)*** (7.70)*** (7.21)*** (7.68)*** City lit area (or sum of lights) -0.091 -0.023 -0.079 -0.001 -0.064 -0.035 -0.053 -0.015 (0.76) (0.13) (0.65) (0.00) (0.77) (0.30) (0.62) (0.13) Test of equal total effect on poverty of big city growth versus smaller town growth Direct effects 1.73 4.21** 0.93 2.21 1.44 3.62* 0.52 1.23 Indirect effects 5.69** 7.98*** 5.64** 8.92*** 14.26*** 15.95*** 14.73*** 18.50*** Total effects 9.18*** 14.40*** 7.77*** 13.08*** 19.64*** 25.28*** 17.22*** 23.19*** Notes: Poverty rates are in logs and other variables are inverse hyperbolic sine transformed so coefficients can be treated as elasticities. Variables interacted with W are spatial lags; all models include fixed effects for 59 NSS regions; tests for the effect of big city growth versus smaller town growth are chi-sq with 1 degree of freedom; z-statistics in ( ), ***, **, ** for p<0.01, 0.05, 0.1. N=236. a The variable mask approach changes the boundary between big city lit area (using a 50% luminosity threshold) and smaller town lit area (using a 20% threshold) satellite-year by satellite-year. b The fixed mask approach reserves area for the eventual expansion of the big cities based on their boundaries in the satellite-year (F18, 2012) that shows maximum big city area (and thus leaves less unmasked area available for the measured expansion in lit area of the smaller towns). Table 4: Effects of City Growth or Town Growth on Rural Poverty Rate for 59 NSS Regions from 1993/94 to 2011/12 -------------------Illuminated Area------------------- ------------Sum of Lights (area × average DN)------------ Variable Maska Fixed Maskb Variable Maska Fixed Maskb (1) (2) (3) (4) (5) (6) (7) (8) Town lit area (or sum of lights) -0.148 -0.117 -0.151 -0.111 (1.79)* (1.45) (1.94)* (1.44) City lit area (or sum of lights) -0.097 -0.097 -0.106 -0.106 (2.36)** (2.36)** (4.01)*** (4.01)*** W × Town lit area (or sum of lights) -0.258 -0.248 -0.260 -0.260 (2.77)*** (2.75)*** (2.93)*** (2.98)*** W × City lit area (or sum of lights) -0.239 -0.239 -0.120 -0.120 (3.76)*** (3.76)*** (2.90)*** (2.90)*** Spatial lag of inequality rate (delta) 0.530 0.323 0.530 0.320 0.604 0.321 0.604 0.316 (8.49)*** (4.08)*** (8.49)*** (4.00)*** (11.02)*** (4.05)*** (11.02)*** (3.92)*** R-squared (within) 0.546 0.628 0.546 0.627 0.504 0.628 0.504 0.628 Total effects – direct Town lit area (or sum of lights) -0.174 -0.140 -0.176 -0.135 (2.62)*** (2.17)** (2.84)*** (2.18)** City lit area (or sum of lights) -0.143 -0.143 -0.142 -0.142 (4.28)*** (4.28)*** (6.15)*** (6.15)*** Total effects – indirect Town lit area (or sum of lights) -0.427 -0.396 -0.429 -0.408 (5.36)*** (5.23)*** (5.62)*** (5.52)*** City lit area (or sum of lights) -0.572 -0.572 -0.427 -0.427 (6.50)*** (6.50)*** (5.33)*** (5.33)*** Total effects Town lit area (or sum of lights) -0.600 -0.537 -0.605 -0.543 (13.08)*** (13.05)*** (13.11)*** (13.12)*** City lit area (or sum of lights) -0.715 -0.715 -0.569 -0.569 (7.82)*** (7.82)*** (6.37)*** (6.37)*** Test of nesting restrictions to derive this modelc 28.26*** 1.18 26.63*** 0.94 35.52*** 4.51 33.71*** 4.03 Notes: Dep variable is log of rural headcount index, other variables are inverse hyperbolic sine transformed so coefficients can be treated as elasticities; variables interacted with W are spatial lags; all models include fixed effects for 59 NSS regions; z-statistics in ( ), ***, **, ** for p<0.01, 0.05, 0.1. N=236. a The variable mask approach changes the boundary between big city lit area (using a 50% luminosity threshold) and smaller town lit area (using a 20% threshold) satellite-year by satellite-year. b The fixed mask approach reserves area for the eventual expansion of the big cities based on their boundaries in the satellite-year (F18, 2012) that shows maximum big city area (and thus leaves less unmasked area available for the measured expansion in lit area of the smaller towns). c The test of nesting restrictions is a test of zero coefficients on the relevant encompassing models in Table 3 that are needed to derive models that have either big city growth or smaller town growth as the explanatory variables. Table 5: Effects of Big City Growth and Smaller Town Growth on Rural Gini Index for 59 NSS Regions from 1993/94 to 2011/12 -------------------Illuminated Area------------------- ------------Sum of Lights (area × average DN)------------ Variable Maska Fixed Maskb Variable Maska Fixed Maskb (1) (2) (3) (4) Town lit area (or sum of lights) -0.036 -0.027 -0.029 -0.021 (1.06) (0.81) (0.93) (0.67) City lit area (or sum of lights) 0.014 0.014 0.004 0.004 (0.81) (0.84) (0.36) (0.40) W × Town lit area (or sum of lights) 0.045 0.028 0.030 0.017 (1.18) (0.77) (0.87) (0.51) W × City lit area (or sum of lights) -0.008 0.001 0.015 0.019 (0.26) (0.04) (0.81) (1.06) Spatial lag of inequality rate (delta) 0.145 0.143 0.145 0.143 (1.46) (1.45) (1.46) (1.45) R-squared (within) 0.011 0.009 0.013 0.011 Total effects – direct Town lit area (or sum of lights) -0.035 -0.027 -0.029 -0.021 (1.25) (0.97) (1.12) (0.81) City lit area (or sum of lights) 0.015 0.016 0.005 0.006 (0.80) (0.84) (0.44) (0.49) Total effects – indirect Town lit area (or sum of lights) 0.046 0.028 0.030 0.016 (1.24) (0.80) (0.91) (0.51) City lit area (or sum of lights) -0.007 0.003 0.018 0.024 (0.20) (0.09) (0.77) (1.00) Total effects Town lit area (or sum of lights) 0.011 0.001 0.001 -0.005 (0.37) (0.04) (0.03) (0.22) City lit area (or sum of lights) 0.008 0.019 0.023 0.030 (0.19) (0.46) (0.82) (1.01) Test of equal total effect on inequality of big city growth versus smaller town growth Direct effects 2.20 1.62 1.43 0.90 Indirect effects 1.05 0.23 0.08 0.04 Total effects 0.01 0.13 0.36 0.88 Notes: Dep variable is log of rural Gini, other variables inverse hyperbolic sine transformed so coefficients can be treated as elasticities; variables interacted with W are spatial lags; all models include fixed effects for 59 NSS regions; tests for the effect of big city growth versus smaller town growth are chi-sq with 1 degree of freedom; z-statistics in ( ), ***, **, ** for p<0.01, 0.05, 0.1. N=236. a The variable mask approach changes the boundary between big city lit area (using a 50% luminosity threshold) and smaller town lit area (using a 20% threshold) satellite-year by satellite-year. b The fixed mask approach reserves area for the eventual expansion of the big cities based on their boundaries in the satellite-year (F18, 2012) that shows maximum big city area (and thus leaves less unmasked area available for the measured expansion in lit area of the smaller towns). Figure 1: Spatial Patterns in Rural Poverty Rates at Baseline Figure 2: Spatial Patterns in Rural Poverty Rates at the End of the Sample Period Figure 3: Spatial Patterns in Rates of Rural Poverty Elimination Figure 4: Location of Big Cities and NSS Regions Figure 5: Illustration of Splitting Urban Growth into Big City and Smaller Town Growth Appendix Table 1: Sensitivity Analyses for Effects of Lights on Rural Headcount Poverty Rates for 59 NSS Regions from 1993/94 to 2011/12: Spatial Durbin Fixed Effects Model ---------------Inverse Hyperbolic Sine--------------- --------------Using 30% lights threshold-------------- (1) (2) (3) (4) (5) (6) (7) (8) Lit area in the NSS region -0.149 -0.152 -0.145 -0.102 (1.75)* (1.76)* (1.85)* (1.22) Average DN value within lit areas 0.070 -0.129 -0.665 -0.648 (0.16) (0.32) (1.24) (1.19) Sum of lights (lit area × average DN) -0.154 -0.154 (1.88)* (2.09)** W × Lit area in the NSS region -0.267 -0.271 -0.283 -0.325 (2.79)*** (2.82)*** (3.13)*** (3.40)*** W × Average DN value within lit areas 1.169 -0.124 0.893 0.931 (1.71)* (0.18) (1.03) (1.09) W × Sum of lights (lit area × average DN) -0.272 -0.271 (2.93)*** (3.15)*** Spatial lag of poverty rate (delta) 0.318 0.738 0.314 0.314 0.325 0.757 0.326 0.328 (4.01)*** (18.53)*** (3.94)*** (3.94)*** (4.10)*** (19.93)*** (4.13)*** (4.15)*** R-squared (within) 0.629 0.124 0.630 0.630 0.626 0.002 0.628 0.625 Restrictions to nest spatial lag model (β2 = 0) 7.78*** 2.92* 7.99** 8.57*** 9.80*** 1.06 11.66*** 9.93*** COMFAC to nest spatial error model (β2 = - β1) 20.46*** 4.14** 20.59*** 22.18*** 24.59*** 0.23 25.34*** 25.76*** Long-run effects – direct Lit area in the NSS region -0.175 -0.178 -0.173 -0.133 (2.56)** (2.58)*** (2.75)*** (1.98)** Average DN value within lit areas 0.446 -0.110 -0.522 -0.547 (1.12) (0.25) (0.97) (0.92) Sum of lights (lit area × average DN) -0.180 -0.181 (2.74)*** (3.07)*** Long-run effects – indirect Lit area in the NSS region -0.435 -0.443 -0.461 -0.507 (5.35)*** (5.61)*** (5.91)*** (6.48)*** Average DN value within lit areas 4.517 -0.229 1.814 1.012 (1.88)* (0.27) (0.50) (0.94) Sum of lights (lit area × average DN) -0.441 -0.451 (5.56)*** (6.04)*** Long-run effects – total Lit area in the NSS region -0.610 -0.622 -0.634 -0.640 (13.21)*** (12.55)*** (13.00)*** (13.21)*** Average DN value within lit areas 4.963 -0.339 1.291 0.465 (1.91)* (0.38) (0.33) (0.39) Sum of lights (lit area × average DN) -0.621 -0.632 (13.29)*** (12.92)*** Notes: All models include fixed effects for 59 NSS regions; models (5) - (8) use logs; nesting tests are chi-sq with 1 df (2 df for col (3), (7)); z-statistics in ( ), ***, **, ** for p<0.01, 0.05, 0.1. N=236. Appendix Table 2: Sensitivity Analyses for Effects of Big City Growth and Smaller Town Growth on Rural Poverty Rates: Using 30% Luminosity Threshold to Measure Smaller Towns -------------------Illuminated Area------------------- ------------Sum of Lights (area × average DN)------------ a b Variable Mask Fixed Mask Variable Maska Fixed Maskb Headcount Poverty Gap Headcount Poverty Gap Headcount Poverty Gap Headcount Poverty Gap (1) (2) (3) (4) (5) (6) (7) (8) Smaller town lit area (or sum of lights) -0.124 -0.230 -0.068 -0.114 -0.101 -0.210 -0.057 -0.112 (1.61) (2.15)** (0.91) (1.10) (1.42) (2.13)** (0.82) (1.17) Big city lit area (or sum of lights) -0.054 -0.017 -0.051 -0.013 -0.058 -0.049 -0.055 -0.046 (1.32) (0.29) (1.22) (0.23) (2.14)** (1.29) (1.97)** (1.19) W × Smaller town lit area (or sum of lights) -0.249 -0.329 -0.256 -0.379 -0.280 -0.354 -0.282 -0.397 (2.53)** (2.38)** (2.83)*** (2.97)*** (3.17)*** (2.84)*** (3.42)*** (3.41)*** W × Big city lit area (or sum of lights) -0.026 -0.029 -0.021 -0.014 0.016 0.027 0.021 0.041 (0.35) (0.28) (0.29) (0.14) (0.33) (0.42) (0.44) (0.63) Spatial lag of poverty rate (delta) 0.316 0.336 0.321 0.335 0.336 0.343 0.337 0.340 (3.95)*** (4.24)*** (3.98)*** (4.18)*** (4.27)*** (4.37)*** (4.24)*** (4.25)*** R-squared (within) 0.630 0.638 0.628 0.635 0.630 0.639 0.628 0.638 Long-run effects – direct Smaller town lit area (or sum of lights) -0.148 -0.266 -0.091 -0.151 -0.129 -0.249 -0.083 -0.151 (2.37)** (3.07)*** (1.50) (1.79)* (2.25)** (3.13)*** (1.49) (1.94)* Big city lit area (or sum of lights) -0.055 -0.016 -0.051 -0.011 -0.056 -0.045 -0.052 -0.041 (1.20) (0.25) (1.09) (0.17) (1.79)* (1.02) (1.61) (0.90) Long-run effects – indirect Smaller town lit area (or sum of lights) -0.402 -0.582 -0.389 -0.598 -0.451 -0.619 -0.434 -0.628 (3.90)*** (4.01)*** (4.21)*** (4.60)*** (5.06)*** (4.98)*** (5.35)*** (5.57)*** Big city lit area (or sum of lights) -0.065 -0.055 -0.058 -0.032 -0.007 0.014 0.002 0.036 (0.63) (0.38) (0.55) (0.21) (0.10) (0.14) (0.03) (0.36) Long-run effects – total Smaller town lit area (or sum of lights) -0.550 -0.849 -0.481 -0.749 -0.580 -0.868 -0.517 -0.780 (5.76)*** (6.24)*** (5.60)*** (6.16)*** (7.03)*** (7.51)*** (6.87)*** (7.43)*** Big city lit area (or sum of lights) -0.120 -0.071 -0.108 -0.043 -0.063 -0.031 -0.050 -0.005 (1.04) (0.43) (0.90) (0.25) (0.73) (0.26) (0.56) (0.04) Test of equal long-run effect on poverty of big city growth versus smaller town growth Direct effects 1.45 5.42** 0.28 1.73 1.22 5.02* 0.23 1.51 Indirect effects 5.31** 6.53** 5.62** 8.21*** 15.45*** 16.05*** 16.33*** 19.48*** Total effects 8.21*** 13.22*** 6.35** 11.34*** 18.70*** 24.90*** 15.85*** 22.34*** Notes: See Table 3 for all notes, except growth of smaller towns in these sensitivity analyses is based on 30% luminosity thresholds instead of the 20% thresholds in the main analyses.