Evidence from Special Economic Zones in India

Place-based Development: Evidence from Special Economic Zones in India Abstract Little is known about the aggregate and distributional effects of Special Economic Zones (SEZs) in developing countries. We investigate the influence of Indian SEZs by exploiting spatial variations in the timing of zonal operations. Using satellite and survey data, we establish that SEZs boosted economic activity within areas several times the size of the zones. The zones also drove a structural change in the local economy with resources shifting away from the informal sector and the formal sector growing in size and productivity. This growth, however, differently benefits workers at the higher end of the income and skill distributions. 1 Introduction Special Economic Zones (SEZs) are a popular policy tool used by both developed and developing economies to boost growth in specific geographic areas and economic sectors. Governments typically direct infrastructure development and regulatory concessions at narrow regions in order to influence the location of large and productive firms within them. The resulting agglomeration of economic activity is expected to increase local competition, produce denser input and labor markets, and produce knowledge spillovers among firms and workers that drive long-term development. Empirical evidence of such benefits is, however, limited and focused on the study of such programs in developed economies.1 The unique institutional and economic features of developing countries, such as the size of their informal sectors and tight credit markets, are certain to influence the success of their SEZ programs. There is hence a need to understand their zonal development initiatives separately. In this paper, we provide one of the first empirical evaluations of the Indian SEZ policy to understand its aggregate and distributional impact on the Indian economy. India introduced the 2005 Special Economic Zones (SEZ) Act as one of the biggest pushes to industrial development in its history. Over ten years since 2005, it has annually invested roughly 0.5% of its GDP, totalling 62 billion US dollars. The size of its investment is substantial by both Indian and international standards.2 As of 2016, 221 SEZs began operation across the nation, attracting firms through tax exemptions, infrastructural benefits and regulatory concessions, directly employing 1.4 million people and contributing to almost one-third of the annual national exports.3 However, the observed economic activity of SEZs is insufficient evidence of the policy’s success. It is necessary to study the indirect effects of such spatial policies when evaluating them. One particular concern is that these observed benefits to regions receiving SEZs could be offset by losses elsewhere in the economy through resource relocations, which would produce little, if any, aggregate gains at considerable costs.4 Using the stages of approvals for SEZs as a source of quasi-experimental variation, our study not only reveals an increase in general economic activity in the immediate neighborhoods of SEZs but also confirms that regions several times the size of the zones benefit- up to the sub-national level of a district. Additionally, 1 In the case of the US, while such policies seem to increase employment and wages (Busso, Gregory and Kline (2013); Empowerment Zones) the benefits are offset by losses elsewhere in the country (Kline and Moretti (2013); Tennessee Valley Authority program). 2 Roughly half the amount was spent over a decade by the Indian government for financing infrastructural development and providing tax exemptions to all firms in two eastern Indian states (Shenoy, 2016). According to Kline and Moretti (2013) and Busso, Gregory and Kline (2013), the investments in the Tennessee Valley Authority program and the Empowerment Zones are around 20 billion and 3 billion US dollars, respectively. 3 Statistics are sourced from the Ministry of Commerce website and Mukherjee and Bhardwaj (2016). 4 The literature on such place-based development policies stresses the necessity to study the response of the wider private sector to the policies and measures their cost-effectiveness based on the programs’ ability to form linkages with the local economy. See Duranton and Venables (2018) , Farole et al. (2011) and Aggarwal (2011) for a discussion. 1 we study how the SEZ policy interacts with the informal and formal sectors of the economy, a dual economy framework which is unique to low income countries. The most pertinent characteristics of informality in India are the absence of any form of regulation in production, omission from the tax base, and workers that are often unskilled and that do not receive social security benefits. While the Indian informal sector is characterized by low productivity, it is a major source of employment hiring about 80% of the total labor force.5 Hence it is important to analyze the effects of SEZs in the dual economy framework for a more comprehensive understanding. Our analysis reveals that SEZs instigate a structural change of the economy, increasing firm size (in terms of production, employment and investment) and productivity in the formal sector while causing a shift of resources out of the informal sector and crowding out its production. Greater formality is considered more desirable in developing countries because it brings a larger part of the economy under beneficial government regulation (in matters such as worker safety and welfare) and broadens the tax base. Moreover, a shift of resources from the less productive informal sector would increase the overall productivity of the economy.6 While our result show-cases the potential of such policies to bring about long-term development, we also find evidence suggestive of a rise in inequality in the short-term. Our results are especially striking given the unique nature of Indian SEZs. The Indian government incentivized the participation of the private sector in zonal development and allowed SEZs to be of substantially smaller physical sizes than found elsewhere in the world. Small zones were doubted to be capable of producing significant additional economic activity. In addition, private sector participation in SEZ development was viewed as tax-free profit generation at the cost of issues such as misuse of land and inequities favoring large companies over small ones.7 The challenge to our identification strategy comes from the non-random nature of program location. This is a common concern for studies analyzing place-based development policies. Due to private sector participation in program location, however, we face the opposite concern of studies generally involving zones in developed economies, which usually target under-industrialized regions. In order to credibly isolate the effect of SEZs on the regional economy, we exploit a source of variation in the government-regulated approval process for SEZs. We consider only those regions surrounding SEZs that have reached the penultimate stage of approval before beginning operation. This stage, henceforth referred to as ‘notification’, ensures that both the regions and the developers possess qualities that make the project viable- the government signals its approval of the developer’s detailed business plan, and the developer signals keenness in the region and his commitment to the plan by completing the purchase or rental of land. Forming our analysis based on 5 6 Authors’ calculations from the 2005 round of National Sample Survey (NSS) on Employment and Unemployment. Hsieh and Klenow (2009) estimate a 40% increase in overall productivity from reallocating resources to larger (formal) firms. 7 See Aggarwal (2006) for a summary of the policy debate. 2 comparisons only among regions that were actively targeted by SEZ developers takes care of the first-order concern that the targeted regions may be different, for example in terms of potential for growth, than other parts of India. While we base our analysis on the variation in the preparatory levels of SEZs, we adopt separate empirical frameworks for exploring different questions. Firstly, we trace the pattern of SEZ influence through time and space using granular satellite lights data as a proxy for economic activity. For this, we apply panel data analysis to a fourteen-year panel of all 1-square kilometer cells that are within 15 kilometers of any SEZ in our sample. This helps us establish that the beginning of operations inside an SEZ sets off an increase in economic activity not only within the SEZs, but also in the immediate neighborhoods around it. We find the effect to be moderately persistent across time and up to areas that are comparable in size to the administrative division of an Indian village. We also find that areas farther away are not hurt significantly (by a potential withdrawal of resources) thus recording net-positive effects up to areas spanning 1,200 square kilometers, one-fourth the size of a median district in our data-set.8 Our main results on the aggregate effects of SEZ spillovers are derived from a difference-in-differences framework using a rich set of variables on firms and workers drawn from formal and informal sector surveys. We compare outcomes between regions that have at least one operating SEZ (treated regions) and those with at least one SEZ that pass notification but no operating SEZ (control regions). In choosing control regions in this fashion, our strategy is in line with Busso, Gregory and Kline (2013) in not comparing areas to their geographical neighbours, limiting the concern of spillovers to the treated regions from control regions (such as movement of workers) that may mechanically bias our treatment effect. Additionally, we are able to separate the direct effects of SEZs (on firms and workers within SEZs) from the indirect effect on the local non-SEZ economy by studying subsets of firms and workers with varying possibilities of being located within SEZs. For robustness, we conduct a series of pre-trends analyses for the regions in our study to address the concern that there may be serious differences among areas of consideration in the timing of SEZ operations. Our results indicate that the average labor productivity of formal sector manufacturing firms in treated districts increases differentially by 24% between 2005 and 2010. We also find evidence for within-industry expansion in formal production by 46%, employment by 18%, and investment in plant and machinery by 37% over the same time period. Along with productivity gains and an increase in the demand for labor, the wages in the treated formal sector experienced a differential increase of 14% over wages in the control district. Our findings also suggest that the resulting agglomeration spillovers from SEZs structurally transform the economy away from informal lines of production towards greater formality. This is especially true in the 8 The district is the main level of local government below the state in India. 3 case of informal manufacturing where we observe a halving of total production within industries of treated districts with total employment declining by 24% and labor productivity by 42%. One possible explanation for this trend is a selection effect driven by an increase in registrations among the most productive firms in the informal sector or those that previously stayed “under the radar” to escape taxes and other regulations. The presence of SEZs likely increased the demand for higher quantity and quality of local goods and services, thus motivating informal sector firms to register themselves to signal quality and expand customer base.9 We also find evidence of a significant decrease in employment in firms at the lower end of the productivity spectrum, such as small household businesses. This suggests a reduction in “forced informality,” which is usually a result of insufficient formal employment opportunities. Although the impact of the SEZ policy has been positive in terms of stimulating formalization in the economy, we find evidence suggestive of increasing inequality represented by a non-uniform effect on overall worker wages. While workers at the 90th percentile of the income distribution gain as much as 38% over the years in which their district was treated with SEZs, those at the lower end of the wage and educational distribution seem not to gain significantly. This suggests that workers are left out of the wage benefits due to their inability of being absorbed by the formal sector. It corroborates the current concern in both developed and developing countries about the problem of “skill gap” where the workforce is unable to fulfill the demand for skilled labor thus holding back further prospects of development. Our work makes novel contributions to the study on place-based development policies. Firstly, we provide a general framework for evaluating place-based development policies using remote sensing data, which can be easily adapted for analyzing other such programs in India or other countries. Our framework is especially useful when policies are not implemented at an administrative level such as state, province or county10 . Our cell-level analysis of policy impact provides a method for studying changes that are within these larger regions, such as industrial corridors, technology parks and other important spatially narrow infrastructure projects. Secondly, we draw important insights from our focus on how zonal development programs affect the littleexplored dual economic structure of the Indian economy. Earlier works studying SEZs and similar programs in India and China, such as Wang (2013), Alder, Shao and Zilibotti (2016), Chaurey (2016) and Shenoy (2016), do not touch upon the effect of such programs on the formal-informal sector dynamics. The closest work to ours, in this respect, is Magruder (2013) who finds that the change in minimum wage rule in Indonesia acted as a big-push mechanism leading to greater formalization. Lastly, we contribute on a positive note to the mixed evidence on the role of firm agglomeration in boosting 9 An upcoming work evaluates if increasing registrations and the quality channel are strong explanatory channels. Such policies allow for analytical methods such as spatial discontinuity design along administrative borders, as employed by Shenoy (2016) in his analysis of the New Industrial policy in the Indian state of Uttarakhand. 10 4 productivity and development of a region. Our analysis of the formal sector is comparable to the work of others in studying the impact of place-based programs in developed countries that do not have prominent informal sectors. Greenstone, Hornbeck and Moretti (2010) find positive productivity gains to firms located in the same county as “million dollar plants” in the United States while Kline and Moretti (2013) find that agglomeration gains from the Tennessee Valley Authority program are offset by losses elsewhere in the country. The evidence is also mixed in the case of programs in developing countries when only the impact on the formal economy is considered. Wang (2013) finds that municipalities receiving early waves of Chinese SEZs experience productivity gains while Chaurey (2016) does not find state-level productivity gains from firm agglomeration in the Indian state of Himachal Pradesh as a result of the New Industrial Policy.11 Our findings of an increase in employment and wages in the formal sector of treated regions, both real and nominal, are in line with those of Kline and Moretti (2013), Busso, Gregory and Kline (2013), Wang (2013) and Chaurey (2016). Given the traditionally low level of labor mobility in India12 , real wage increases are also consistent with the prediction of Moretti (2010) that low labor mobility implies that any benefits from a shock to labor demand accrues to workers residing within a region.13 The paper proceeds in the following steps: Section 2 provides the reader with an overview of Indian SEZs. In Section 3, we study the pattern of spillovers caused by SEZs on the surrounding areas. Section 4 explores the aggregate effects of these spillovers and their implications for firms and workers. Section 5 summarizes our insights and highlights further areas of research. 2 The Indian SEZ Experience The Government of India’s SEZ policy was influenced by the success story of the Chinese SEZs. Impressed by the SEZs in Guangdong province in 2000, the Commerce Minister of India initiated changes in India’s Export-Import policy, which converted existing Export Processing Zones (EPZs), which were industrial estates that produced export-oriented goods, to Special Economic Zones (SEZs).14 SEZs were envisioned as comprehensive industrial townships with social facilities like housing blocks, schools and hospitals. The real growth in SEZ activity was kick-started by the SEZ Act of 2005, which officially proclaimed its objectives to be: (a) generation of additional economic activity, (b) promotion of exports of goods and services, (c) 11 The movement of resources to treated regions from other regions is potentially responsible for muted gains at a spatially aggregated level. Glaeser and Gottlieb (2008) point out that this shift may still have an overall welfare impact on the aggregate economy if the elasticity of productivity to agglomeration is greater in places receiving the programs. Empirically, non-linearity of agglomeration effects is a challenge to establish and many papers, including this one, focus on attempting to document if there are positive net effects on the surrounding economy, up to a sub-national level of aggregation. 12 Topalova (2010) finds that the landmark trade reform of 1991 that officially opened India to international trade had a surprisingly little impact on the already low inter-district migratory patterns. 13 High mobility, on the other hand, would predict an in-migration of workers who would apply an upward pressure on land prices and cancel out the effect of any increase in nominal wages. 14 EPZs do not form part of our sample. 5 promotion of investment from domestic and foreign sources, (d) creation of employment opportunities, and (e) development of infrastructure facilities. The Indian SEZs differed in two key ways from SEZs and other place-based programs in the world, including the Chinese model. First, the minimum size requirement was much lower. This resulted in Indian SEZs being physically much smaller than municipality-sized SEZs in China and census tracts designated as Empowerment Zones in the United States. The size requirements were sector-specific; while Information Technology (IT) SEZs were allowed to be as small as 0.1 square kilometers, multi-product SEZs needed to be at least 10 square kilometers of area. The second distinguishing feature of Indian SEZs is that they were open to development by both the public and private sectors, resulting in 70% of the SEZs being either private or joint sector initiatives.15 These features resulted in two main trends in zone location16 : both public and private sector SEZs tended to locate in urbanized areas with already existing industrial clusters, or they clustered in belts to promote the development of a new industry within the state. Hence, despite the small size of an individual SEZ, the tendency to cluster increased the potential of agglomeration spillovers to impact regional productivity and economic growth. Studying this unique pattern of SEZ development could thus provide useful lessons to countries that find it economically and politically infeasible to develop large-sized SEZs. Similar to other place-based development programs through the world, India provided largely fiscal incentive packages to the SEZ developers as well as to the firms locating within SEZs (henceforth referred to as units). Table A.1 in the Appendix provides an overview of these incentives. The most notable of these is the 100% tax exemption on profits for the first five years of operation, which converted to a 50% exemption in the next five after which the same rate was applied to any profit that was reinvested into SEZ activity.17 Additionally, both developers and SEZ firms were exempted from paying the Minimum Alternate Tax (MAT), which is currently set at 18.5% of book profits in India.18 The MAT is a compulsory tax levied on companies that make substantial profits but have low, or even zero, tax liability due to the host of deductions and exemptions available under the income tax law. According to a representative survey of SEZ developers and companies by Mukherjee and Bhardwaj (2016), 84% of the interviewed units declared that the tax exemptions formed the biggest motivating factor for them to begin production within SEZs. Apart from tax benefits, both developers and companies wishing to locate within SEZs enjoyed an ease in administrative procedures through the “single window mechanism”. Applications were reviewed jointly by both the Central and State governments through a single regulatory body, the Board of Approval (BoA), 15 16 17 18 Authors’ estimates. See Aggarwal (2011) for a detailed survey of SEZ developers on issues including zone location and development. The corporate tax rate is 35% in India. This exemption was reversed in 2011. 6 which was set up to facilitate a fast pace of clearances and resolution of bureaucratic red-tape typically surrounding the starting of a business venture. All of these incentives helped to create a relatively hasslefree environment for firms, wished to operate in a country not known for its ease of doing business19 . Our empirical analysis hinges on the approval process for establishing SEZs. Applications to develop SEZs were submitted to the BoA, which met quarter-yearly and reviewed them based on the following criteria: the quality of the business plan, the plan for financing, land type targeted20 and prior approvals of the state government.21 While these factors were repeatedly highlighted during the decision making process, the relative importance of each, and whether this list is exhaustive, is unclear. If the application meets the requirements, the developing company is issued a formal approval. After this, it needs to revert to the BoA with documentation on land rental or purchase agreements as well as with any revisions to the development plan suggested by the BoA. At this stage, the body issues a notification for the SEZ. This is usually brought to the attention of the general public through news articles as well as notice boards erected at the site of the planned SEZ. Construction then commences and an SEZ is considered operational once the first unit starts production within it. Figure A.1 in the Appendix illustrates the approval process of the SEZs. In our analysis, we only consider those SEZs that pass the penultimate stage, i.e. notification. The Ministry of Commerce and Industry is the primary source of information on Indian SEZs in the form of lists of notified and operational SEZs in India. We merge the lists to obtain information on the developer, date of notification, whether the zone has started operation, zone size, the industrial sector as well as the location of each SEZ (down to the village level, and occasionally to the street level). We then add the actual starting dates of operational SEZs, defined as the year in which the first unit within the SEZ becomes operational, which we source from newspaper articles, BoA meeting minutes and developers’ websites. Our data-set, given the rich location details, can easily be analyzed at multiple levels of aggregation- at the neighborhood level with geo-coded location data, village, and district level. This is helpful in merging it with secondary data of different aggregation possibilities (as we do in this paper). Our sample includes all notified SEZs in the states of Andhra Pradesh, Gujarat, Maharashtra, Karnataka, Kerala, Uttar Pradesh and Tamil Nadu, which host more than 80% of total operational SEZs. At any point in time, we compare regions that host (or are about to host) operating or notified SEZs. 19 In 2017, India ranked 130 out of the 190 countries considered in the World Bank’s Ease of Doing Business Index, and has been consistently ranked below countries such as Iran, Nicaragua and Uganda. 20 The land should not only meet the minimum size requirements, but it should also be a contiguous area that is preferably waste land or unsuitable for double-crop cultivation. 21 Information is derived from the BoA meeting minutes. 7 3 Patterns of SEZ Influence Across Space and Time 3.1 Data Overview Our main data-set for analyzing the pattern of spillovers from SEZs is the Nightime Lights (henceforth referred to as NTL) time series. The raw data (National Geophysical Data Center, 2013) is obtained from the National Centers for Environmental Information. The resolution is 30 arc-seconds, which is equivalent to approximately 854 meters when measured around the center of India, and the data values range from 0 (background noise) to 63 (brightest). We clean the series from 2000 to 2013 by removing ephemeral events and gas flares to obtain persistent and stable lighting. Nighttime luminosity has been extensively utilized as a proxy for economic growth and development (Henderson, Storeygard and Weil, 2012), degree of urbanization (Ma et al., 2012), degree of electrification (Min et al., 2013), population density (Sutton et al., 1997), etc.22 The NTL has three main advantages- it is available annually at a high level of disaggregation and provides a neutral measure of a region’s economic activity. In the Indian context, given that economic data is spatially coarse, available at infrequent intervals and may fail to capture the extent of economic activity in its vast unorganized sector, the usage of a neutral measure of economic activity enriches the overall examination and adds robustness to the findings derived from other data-sets that are available at higher levels of spatial aggregation. The annual availability further helps in confirming the effects of an SEZ that appear precisely in the year it becomes notified or operational, over neighbourhoods of varying proximity. Our reliance on nighttime lights data to capture general economic activity at the sub-national level is further bolstered by Bhandari and Roychowdhury (2011) who find that district-level GDP in India can be significantly explained by lights data and by Dugoua, Kennedy and Urpelainen (2018) who find that lights data has strong predictive power for village-level electrification in India. Furthermore, by being agnostic to the nature of economic activity captured by nighttime lights, we handle the criticism that sub-national comparisons of economic development using nighttime lights is problematic. Works such as Mellander et al. (2015) argue that the correlation between the nighttime lights and economic activity is weak for wages and strong for population and establishment density in the case of Sweden. In our case, even if nighttime lights are only correlated with population density, not economic growth, it still provides us with meaningful interpretation since we are also interested in spatial reallocation of human resources. Using the NTL data and geo-coding the locations of all the SEZs in our data-set, we overlay a fine grid 22 The frequency and disaggregated nature of the lights data allows for creative applications in studying a variety of interesting and previously unexplorable issues, such as regional political favoritism as in Hodler and Raschky (2014), and the effects of spatial distribution of ethnicities before colonization (Michalopoulos and Papaioannou, 2013). 8 layer over the map of India and use each cell as the unit of analysis. Each cell is defined as a square with the length of 0.01 decimal degree, which is approximately 1.025 kilometers at around the center of India. In order to see the indirect or spillover effects of SEZs on the cells, we restrict our attention to cells that are believed to be strictly outside of SEZs. For identifying cells outside the zones, we assume that SEZs are circularly shaped since the exact shape of SEZs are unknown. Using the area of the SEZ reported by its developer, we calculate the radius of the SEZ and draw a circle around the point. The circular shape assumption is a strong one and creates a concern that we might label area that is actually inside the SEZ as non-SEZ area. To avoid defining inside-SEZ cells as outside-SEZ ones, we take a conservative approach and increase the radii of SEZs by 10%.23 We then record for each cell, the NTL reading over the years24 and the distances between the centroid of the cell and the projected boundary of every SEZ in our database, thus linking SEZ-level information (such as notification and operation years) with cell-level information. We restrict our attention to cells that are at most 15 kilometers away from their closest SEZ, and the resulting number of observations (cells) is 62,386 per year. We additionally use the Gridded Population of the World (GPW) data (Center for International Earth Science Information Network, 2016) series in order to measure the extent to which the effects recorded by our analysis of the NTL data is driven by population movements. This data-set, downloadable from the NASA Socioeconomic Data and Applications Center (SEDAC), is available at 5-year intervals from 2000. The lowest level of data resolution is the same as the NTL (roughly 1 kilometer-wide cells). The data for each cell is derived using the population listed by national and sub-national administrative units. In the case of India, this is the level of the sub-district. Each grid cell is assigned with values of population density per square kilometer according to a proportional allocation gridding algorithm, which allocates the same value to all cells within a sub-district.25 We hence acknowledge that the data-set may have limitations for study at a granular level. However, since the cell-level population is at least not derived from its NTL reading, it is useful to analyze cell-level NTL per population. This provides us with an estimate for how important population movements are as an explanation for the effects produced by the NTL analysis. Table 1 provides us with an overview of the SEZs in our sample. We have 251 notified SEZs in our sample, with 133 of them operational by 2014. The median size of an SEZ is about a third of a square kilometer, the size of around 32 soccer fields. The mean is much larger at 1.5 square kilometers due to the presence of a few exceptionally large SEZs such as that the Mundra SEZ in Gujarat, which spans 64 square kilometers. Since the IT and electronic sector SEZs make up 69% of the total and their minimum size requirements are small, 23 See Figure 1 that depicts the fine grid of cells superimposed on circles that indicate SEZ locations as well as size. In cases there were two satellites collecting data for a cell, we take the average of the two readings. 25 The allocation is based on an assumption that the population of a grid cell is the exclusive function of the land area within that pixel. Water area such as lakes, rivers, and ice-covered areas are excluded. 24 9 such size distribution is to be expected. It is also apparent from Panel A that SEZs are largely a private sector venture with 70% of the zones being developed by purely private or joint sector entities. The average year of notification of an SEZ is 2008 (Panel A) with operating SEZs being notified slightly earlier (Panel B). The difference between average year of notification and operation for SEZs in Panel B shows us that the developers take an average of two and half years to secure the necessary permits, complete substantial construction and attract their first tenants. Panel A of Table 1 also displays the average NTL and NTL per population in 3 kilometer-neighborhoods around SEZ boundaries. These neighborhoods experienced an increase in economic activity, which goes beyond the increase in population judging by the values of both variables after the initiation of the SEZ Act. It is important to note that due to the NTL data being top-coded at 63, our analysis may be capturing the lower bound of the actual growth that took place. This is because SEZs tend to locate in relatively urban areas with already high values of NTL and may result in the right-censoring of our empirical results. We, therefore, anticipate spillover effects to be underestimated in our study. Figure A.3 indeed confirms that by the end of sample period, 2013, a non-negligible fraction of the data is top-coded implying that our estimates are conservative.26 3.2 Empirical Strategy We utilize the fourteen-year panel data-set of cells, which is constructed as detailed in the earlier section. Our general approach is to compare cells based on their proximity to SEZs and measure the differential effects on them through time due to an SEZ’s presence. For clarity of interpretation, we focus on the time and distance dimensions in separate analyses. The time-dimension analysis uses an event-study framework and examines how a given area reacts to the event of a nearby SEZ beginning operation. The cells considered in the event-study are those that have exactly one SEZ within 3 kilometers that begin operation between 2006, the first year of SEZ operations, and 2013, the last year of observation of the NTL. In making this selection, we aim to (a) focus on the reaction of the immediate neighborhood and (b) reduce the number of SEZs that affect a cell.27 For such a cell i that is situated outside of SEZs in year t, we run the following specification: k6X =−1 log(lightit ) = αi + βt + γk ∗ Dikt + it , (1) −13≤k≤7 where the outcome variable is logged luminosity of cell i at time t, which is the logged transformation of 26 See Section A.2 in the Appendix for a brief discussion. Since the treatment is at the SEZ level and the analysis at the cell level, it is possible that a cell is influenced by multiple SEZs in its vicinity, with overlapping notification and operation timelines, especially given the tendency of SEZs to cluster. 27 10 the cell’s NTL value incremented by 1. A binary variable Dikt takes the value of one if the SEZ within 3 kilometers away from cell i has been operating for k years in year t. Year 0 is the initial year of operation of an SEZ. We use cell and year fixed effects to add robustness to our findings. Additionally, we correct the standard errors for spatial autocorrelation following the specification of Conley (1999) up to a cutoff of 30 kilometers28 . The estimates of interest are γk ’s, for k ∈ [−13, 6]. Each γk can be interpreted as the change in brightness (in log deviations) of a cell, k years since the operation of the nearby SEZ, relative to the year before its operation (γ−1 is normalized to 0). We expect γk to be positive and its magnitude to be increasing in k after the initial year of operation, indicating persistence in the effect of operating SEZs. Prior to operation, we should not expect any significant trend in γk since we do not expect regions to be affected by an SEZ even before it establishes its presence. We also use the same framework and have similar expectations from the coefficients when we study the event of notification of an SEZ. For the distance-dimension analysis, we classify the SEZs at any point in time into three main age groups: period0 denotes years before the SEZ is notified, period1 covers the post-notification and pre-operation years, and period2, the years after operation. For a cell i that is not located in any SEZ, an SEZ in the x-th distance ring (x − 1 to x kilometers away from the cell) exerts an effect that is dependent on whether the SEZ is in period 0, 1 or 2 of its lifetime. In order to study the varying effect of an SEZ across distances, we conduct the following analysis for a particular value of x: log(lightit ) = αi + βt + γx ∗ period1ixt + δx ∗ period2ixt + 15 X 2 X λθd periodθidt + it , (2) d=x+1 θ=0 The main outcome variable is logged luminosity of cell i in year t and we run the above specification for each integer value of x from 0 to 15. For the xth distance ring analysis, we consider all cells that have their closest SEZ in this ring. periodθixt is the number of periodθ SEZs that are in the xth ring away from cell i in year t, for θ values 0, 1 and 2.29 The average difference in the effects of period 2 and period 0 SEZs (the base group), denoted by δx , captures the additional effect felt on a cell due to the operation of an SEZ in the xth ring. The average change to the NTL of a cell due to the notification of an SEZ in the xth ring is similarly denoted by γx . δx can be thought of as a long-run effect of an SEZ on a cell, and γx , the short-run effect. Both measures are useful in developing an understanding of the changes that the local economy experiences, although the latter effect is of greater economic interest. The double summation term indicates that we control for all farther away SEZs affecting the cell (but not within the xth ring), in whatever period of life 28 The results are robust when the cutoff is alternatively set to 10, 20, 50, and 100 kilometers. The empirical results are qualitatively unchanged when the terms become dummy variables of whether or not there is at least one SEZ in each period. 29 11 they may be.30 Just as in Specification 1, the standard errors are corrected for spatial autocorrelation. We expect δx to be decreasing in x, i.e. cells closer to SEZs experience the most positive effects while cells farther away experience less positive or even negative effects. The latter would be the case if there was a movement of resources away from farther areas to areas closer to SEZs. This is a reasonable expectation given that the zones bring in new firms that attract workers as well as other firms to the region due to an increase in opportunities for work. We also expect that the effects of operating SEZs are greater in magnitude than the effects of notified SEZs at all distances from the SEZs (δx ≥ γx ). Using the specification above, we are also able to test if changes to the neighborhoods of SEZs in terms of NTL is driven by population movements, a channel we can test at the granular level with the main outcome variable being logged lights per population. Since the population data is available at 5 year intervals, the period of life of an SEZ is updated every 5 years. We expect that a significant portion of the expected increase in NTL at neighborhoods close to SEZs will be driven by an increase in population in that neighborhood. The magnitudes should, however, be taken as less reliable than in the analysis using only NTL data due to the way in which population data is constructed for India, as discussed in Section 3.1. 3.3 Results Inside SEZs: Event Study of Activity within SEZs An event-study of areas inside SEZs confirms the relevance of Specification 1. Figure 2a reports the percent change in the cells’ NTL (derived from the γk ’s), which shows a clear and persistent increase in activity after the beginning of SEZ operation. The figure, however, also indicates an upward trend in economic growth prior to the SEZs’ operation. One possible explanation for this trend is the preparatory activity undertaken in and around SEZs after notification (in terms of building, road construction, setting up water distribution networks, etc.). As discussed in Section 3.1, the data shows that it took 2 to 3 years on average after their notification for the operating SEZs in our sample to begin operation. Figure 2b, which studies the event of notification, confirms that the pre-operation growth in Figure 2a may indeed be driven by post-notification activities. SEZ Spillovers over Time We then restrict our attention to areas strictly outside, but within 3 kilometers from any SEZ to study the spillover effect of SEZs across time. Figure 2c displays a similar trend as the within-SEZ analysis. When an SEZ starts operating, the immediate neighborhood experiences a significant increase in NTL, which is lower 30 Figure A.2 contains a pictorial representation of Specification 2 for greater clarity. 12 than the increase within SEZs as expected. It is persistent in magnitude but loses significance over time. We suspect that this is due to the small number of SEZs older than 4 years old by 2013, considering the mean year of initial operation is 2009 according to Table 1. We still observe the upward trend in NTL prior to operation. Since the area of analysis is physically outside zones, the increase in post-notification activity can be a result of both construction activities that may extend outside zones such as building external connecting roads as well as the surrounding economy preparing for the impending shock to local demand for goods and services. Examples of the latter would be the construction of hostels and residential properties to host potential out-of-area SEZ workers. This may still not explain the upward-sloping trend prior to even the notification of SEZs as Figure 2d shows.31 However, we can still argue that regions around notified SEZs experience a noticeable increase in the slope of NTL, i.e. growth rates, upon notification. Figures A.4a and A.4b display results from an alternate specification in which we modify the event to be the year of the earliest operating/notified SEZ within 3 kilometers. This specification is more flexible in the selection of cells than the current one and allows for the presence of multiple SEZs within the 3 kilometer radius. For example, these cells could be representative of more urban areas, which attract a greater number of SEZs. This does not, however, produce significantly different patterns, adding to the robustness of our finding.32 SEZ Spillovers across Space Figures 3a and 3b, plot γx ’s and δx ’s in Specification 2, illustrating the spatial extent of SEZ spillovers in the long- and short-run, respectively. The resulting trends are in line with our expectations. In the longand short-run scenarios, we observe a sharp increase in NTL in neighborhoods of operational SEZs and notified yet not operating SEZs (relative to the average level in the pre-notification period), respectively. The positive impact especially continues to be significant for areas with the closest SEZ within 4 kilometers in the long run. Interestingly, from Figure 3a, we do not find evidence of a zero-sum relocation of resources due to the SEZ policy, at least at the level of aggregation considered in this paper- at areas up to 15 kilometers away from SEZs.33 While areas at a distance greater than 5 kilometers from SEZs seem to return to their normal growth pattern, there is no strong evidence of farther regions being negatively affected by a withdrawal of resources. This area is equivalent to quarter the size of a median district in our data-set. We cannot, 31 The pattern is preserved when we control for time trends in addition to year fixed effects. We also experiment with increasing the radius of the neighborhood to 5 kilometers to get similar results, which are not included in the data appendix and are available on request. 33 This holds true when we check patterns over a wider area of 20 kilometers’ radius (results of which are presented in Figure A.5) that spans 1,200 square kilometers. 32 13 however, rule out zero-sum effects over wider regions with this framework. One concern about Specification 2 is that it considers cells that have their closest SEZ in the xth distance ring and thus looks at different sets of cells for the analysis across distances, i.e. cells that do have SEZs in the xth distance ring are not considered for the xth distance ring analysis if they have at least one SEZ that is nearer than x kilometers. We carry out an alternate, less restrictive specification by considering all cells that have at least one SEZ in the xth distance ring, and controlling for both nearer and farther away SEZs when studying the effects of SEZs at a particular distance ring x around a cell.34 Results from this alternate specification is displayed in Figures A.6a and A.6b. While the pattern of influence is similar to the ones produced with the original specification, there is stronger evidence that regions farther away than 5 kilometers do not get hurt and in fact experience positive and significant effects due to operational SEZs. In Figures 3c and 3d, we adopt Specification 2 to show the effect of having an SEZ nearby by the distance to the closest SEZ on NTL per population while acknowledging the drawback of the population data detailed in Section 3.1. The population movements do seem to drive some, but not all, of the increase in NTL especially in the long-run scenario. This suggests that there are other channels at play, especially in the case of effects produced by operating SEZs.35 The chief takeaway from the analysis in this section is that we find evidence of Indian SEZs producing persistent and positive spillovers on the local economy as measured by nighttime lights. While they boost economic activity in areas up to 4 kilometers from zones (comparable to the size of a typical Indian village), farther away areas do not seem to be significantly hurt. SEZs, therefore, can be viewed as generators of net positive effects on areas several times the actual size of the zones and even comparable to the size of an Indian district. An important caveat to our findings is that we are unable to determine if there are net positive effects at the national level. It is possible that activity shifts from elsewhere in the country to areas around SEZs, producing little if any net benefits at the aggregate level. Our results speak only to regional positive effects. This finding, however, motivates a deeper look into the forces behind the spillovers generated in the local economy. In addition, it also lend credibility to the results derived in the following section which use more traditionally available data-sets on firms and workers in the Indian economy that are available at the spatially aggregated level of the village or district. 34 See Section A.5 in Appendix. Acknowledging the limitation of the GPW data, we run a separate analysis examining the influence on SEZs on sub-district level population movements by aggregating the NTL and GPW data at the sub-district level. A detailed explanation and the corresponding result is presented in Section A.7 in Appendix. In short, it is unlikely that there are any population movement across sub-districts that can be attributed to the introduction fo SEZs. 35 14 4 Aggregate Effects of SEZs on Firms and Workers We have shown in the previous section that the introduction of SEZs into a region promotes general economic activity. The effect appears positive up to a level of geographical aggregation, a district, which is of interest to political and administrative authorities. A district is the main level of local governance below the state and is divided further into sub-districts that consist of villages and towns. In our sample, there are 68 districts, each of which is, on average, divided into 19 sub-districts. Each sub-district is further divided into an average of 9 villages. In order to understand the general equilibrium effects of SEZ activity, we now examine individual and firm behavior in response to the introduction of SEZs.36 We employ multiple data sets at different levels of administrative units- the village, sub-district and district. 4.1 Key Variables At the village level, we analyze firm and worker numbers using the Economic Census (EC). The strength of the EC data is that it is a complete enumeration of all enterprises in India (except those engaged in crop plantation and cultivation) and is identifiable at the village level. Through firm-level information on employee size, industry and ownership type, it provides us with an overview of the distribution of activity across industrial sectors, both manufacturing and services, in every village or town. The EC data covers information on 28 million firms located in the districts of our interest in each of the two rounds available, 2005 and 2012.37 At the district level, we make use of more detailed firm characteristics such as production and wages to shed light on the mechanism behind SEZ effects. For studying movements within the formal manufacturing sector38 , we use the Annual Survey of Industries (ASI) dataset. The ASI dataset is an annual survey of firms in manufacturing that are considered formal, i.e. those registered under the Factories Act39 . It is a complete enumeration of firms above a 100 in worker strength with an annual survey of a repeated random cross-section of smaller firms. This data contains more information than the EC, which allows for a deeper analysis of firm-level variables such as size (employment, asset base and production), new firm formation and costs of production including wages and rents. The data set covers around 30,000 firms in the districts of our interest annually from 2000 to 2009. We also make use of a similarly rich set of firm-level information provided by the informal sector counterpart to the ASI, the National Sample Survey (NSS) Unorganized Manufacturing and Services quinquennial survey data, which covers firms in the unregistered sector of the 36 By ‘firm’, we imply establishment or factory-level information. Results pertaining to village- and sub-district level analyses can be found in the Appendix. 38 Nationally representative surveys on formal service firms have not been conducted so far in India. 39 Registration under the Factories Act is required for firms above 10 workers if the unit uses power, and above 20, if not. This is also the standard definition of formality adopted by researchers on the Indian economy. 37 15 Indian economy. Each survey round in our study (2000, 2005 and 2010) contains information on around 35,000 firms in the districts of our interest. We complement the firm analysis with worker-level information which allows us to analyze worker wage effects within districts taking into consideration individual characteristics such as education level attained and household demographics. This information is derived from the NSS Employment and Unemployment Surveys. The data is a repeated cross-section of a nationally representative sample of workers from across all industrial activities. Information on firm type and industry, wages, household characteristics, education and consumption is provided. The data set covers around 104,000 workers in every round considered (2000, 2005 and 2010) among the districts in our study. While the worker-level survey data does not have explicit indicators for whether the worker is employed in the formal or the informal sector, we make the distinction using the 10- worker rule of the Factory Act in order to analyze the effects of SEZs separately on the informal and the formal work-force. Table A.2 in the Appendix provides an overview of the regions and population studied in 2005. The regions in our study were on average much denser than the all-India average of 382 people per square kilometer. More than 90% of the working population in a district received no education above secondary level. Average firm size was small with 96% of the firms employing below 10 workers. This indicates the highly skewed firm size distribution and the vast size of the informal sector.40 The extent of informality in the economy is also apparent from indicators such as the proportions of firms that hire no workers at all (40%), operate without power (34%), do not have external financing options (96%) and rely on informal sources of finance (around 40%). We also find evidence of the low level of productivity in the informal sector with wages in the formal sector manufacturing being on average almost 10 times that of those in the informal sector. 4.2 Empirical Strategy In this section we adopt a difference-in-differences framework which chiefly requires that both treated and control regions follow common trends prior to the initiation of the SEZ policy. This is a challenge since the locational choices of such place-based policies are not random. In the case of developed economies, zones are usually located in under-industrialized regions. In the Indian context, however, because SEZ development was mainly a private sector initiative, the program targeted regions with a greater degree of urbanization, human capital quality and profitability. A simple comparison of areas with SEZs and those without would therefore be unsatisfactory as it will likely violate the common pre-trends assumption and bias our difference-in-differences estimate of effects upwards. Our solution is to adopt a similar strategy to the one commonly used in the literature on place-based 40 The figure is comparable to estimates in Amirapu and Gechter (2014). 16 development policies. We use the the approval process of SEZs as a source of quasi-experimental variation to compare regions that have at least one SEZ that has passed notification but none that is operating in them (the control group) with those that have at least one operating in them (the treated group). The exact set of regions that fall into treatment and control categories differ according to the frequency and level of aggregation allowed by the multiple data-sets used.41 In the case of Busso, Gregory and Kline (2013) and Kline and Moretti (2013), control regions were formed out of rejected areas. There is an added strength in our case such that the control areas were never disqualified by administrative authorities and were indeed expected to have operating SEZs in the near future. A point immediately in favor of our identification strategy is that it takes care of the first-order concern that areas attracting SEZs, both notified and operational, may be fundamentally different from other areas in terms of worker and industrial composition and growth potential. Table 3 provides some evidence for this when we compare across treated, control and other districts in the states in our sample. Both treated and control districts are almost ten times as dense as other districts in the states, consistent with the trend of SEZs locating in relatively urban areas. Treated and control districts also seem to have substantially different worker compositions than the rest of the state, especially with respect to the proportion of workers employed in agriculture and manufacturing. The rest of the state seems to predominantly depend on agriculture for its livelihood with 64% in related professions compared to only 7% in treated or control districts. The average monthly income of workers and their education level in the districts compared against each other in our analysis are also clearly higher with the proportion employed in informal household businesses being lower by around 10 percentage points. This table tells that results from our proposed comparison of treatment and control districts will be more credible than a simple comparison of SEZ and non-SEZ districts. The second-order concern comes from the fact that treated regions seem to be disproportionately among those with earlier notified SEZs, as shown in the bottom panel of Table A.3. This could imply that the order of timing in application and notification is correlated with unobservables relevant for the outcomes studied. However, anecdotal evidence points to developers being spatially restricted in their choice of location. Locations were not chosen solely based on profitability but also giving weight to the own-state bias of SEZ developers, both public and private. State governments always started SEZs within their states, and private sector developers usually chose locations within the state or district in which they are headquartered.42 It is then quite likely that the control areas did not form part of the choice set of developers in treated regions. This could be more due to their out-of-state location rather than growth potential. Moreover, we find that 60 percent of the control districts were targeted earlier by SEZ developers that did not manage to reach the 41 42 Refer to Table A.3 in Appendix for an overview. Evidence is gathered from interviews with SEZ developers in Tamil Nadu. 17 stage of notification. We derive this information from the BoA meeting minutes which discuss the decisions made on all the SEZs that ever applied for approvals. These areas attracted initial interest around the same time as treated areas according to the bottom rows of Table A.3. We find that SEZs in both control and treated districts received early formal approvals within the first year of the SEZ Act. The average difference between the two groups receiving a formal approval was only about 8 months. Given that treated and control regions received developer interest within a comparable time frame, one is also naturally concerned about the delays to notification and eventual operation of SEZs in control regions, and whether this was correlated with the potential of the regions and the selective success of the SEZ policy. If the set of selection criteria for regions were completely known, as was in the case of Busso, Gregory and Kline (2013) in their evaluation of Empowerment Zones, we could use techniques such as propensity score weighting to increase the comparability between treated and control regions. In our case, however, the exact set of conditions used by the BoA to deem an SEZ notification-worthy, is unknown. What we know from the BoA’s meeting minutes is that these included other criteria than economic indicators for the region- such as the ability of the developer to possess the land, obtain environmental clearances and propose a viable development plan. The difficulties in land acquisition and clearing administrative steps including environmental clearances, which were prerequisites for notification, are also well documented by Mukherjee and Bhardwaj (2016) and Aggarwal (2006). Arguably these measures are related to state-level factors such as the lack of land banks, or delays in administrative clearing that can be believed to be time-constant at least within the short-frame of our analysis of the SEZ policy. There is also the possibility that treated regions differ in the number of ‘capable’ developers than control regions given that they have earlier operating SEZs; we address these issues with the addition of region fixed effects to our framework which would take into account non-time varying differences in business and administrative potential among regions. We derive additional evidence of common trends among treated and control regions by utilizing NTL data to compare time trends among cells in 5 kilometer-neighborhoods of SEZs that eventually became operational versus those that never became operational (by 2013, the last observation year for NTL). Figures 4a and 4b show that both kinds of regions did not experience significantly different pre-trends (as captured by trends in NTL) before notification, regardless of whether the SEZs in them ever began operation. Since the areas covered by a 5 kilometer radius is roughly twice the size of an average village in our sample, the analysis confirms common trends among relative large portions of land. We also find evidence for ‘developer seriousness’ in control regions from the trend in Figure 4b where areas around notified SEZs seemed to experience an increase in economic activity in the initial couple of years after notification judging from the spike in economic activity before the region returned to its normal growth path. This supports our belief that the developers of notified SEZs were committed to the region and to the project. 18 We also take advantage of the ASI data to analyze pre-trends in district totals of production, assets used and employment in formal manufacturing sectors prior to the SEZ Act. Figure 5 shows that there are no discernible differences in trends among treated and control districts, except for a slightly higher trend in employment in treated districts. We also carry out the following falsification test using the same information at the firm-level, to see if operating and notified SEZs produced effects on formal firm activity even before their introduction in districts: log(yf idt ) = α0 + αi + βt + γ ∗ N o.EventuallyOperatingSEZsd (1) + δ ∗ N o.EventuallyN otif iedSEZsd + εf idt The outcome variable yf idt takes the logged values of variables related to an average firm f in 2-digit industry i in district d at time t: such as production, investment, employment, wages and average productivity. The main regressors are the number of eventually operating and notified SEZs that the district receives after 2005. Since the analysis is over the time period between 2000 and 2005 (before the announcement of the SEZ Act.), the corresponding coefficients,γ and δ , should not show any significance if there had not been any differential growth pattern before the introduction of the SEZ Act. The results can be found in Table 2 and confirm that there were not any significant differences in growth patterns in the pre-periods. Given the supporting evidence for parallel trends among districts with notified SEZs and those with operational SEZs before the initiation of the SEZ policy, we first utilize the annual data on the formal manufacturing sector in a generalized difference-in-differences framework to study the year-on-year effects of additional SEZs as they become notified or operational within a district. We only consider districts that have been notified with at least one SEZ before 2010. The treatment variables are now the stock of operating SEZs in district d at time t and the stock of notified but not yet operating SEZs in the same district at time t: yf idt = α0 + α1 ∗ N o.OperatingSEZsdt + α2 ∗ N o.N otif iedSEZsdt + βi + γd + δt + εf idt (2) yf idt is the outcome variable of an average formal manufacturing firm f in industry i and district d at time t. These include logged values of production, average labour productivity (defined as total production per worker), wages, employment and value of plant and machinery. N o.OperatingSEZsdt refers to the total number of operational SEZs in district d at time t and N o.N otif iedSEZsdt refers to the number of notified but not yet operational SEZs in district d at time t. District, 2-digit industry and year fixed effects are included with the standard errors being clustered at the district level. In the above specification, we assume that every SEZ, operational and notified, has a uniformly additive effect on the outcomes of a firm in a 19 district-industry cell. We expect positive effects on firm-level measures such as average labor productivity, employment and production due to the presence of an additional operating SEZ, i.e. α1 to be positive. This would provide proof of the push that SEZs give to local demand and of the productivity spillovers that they are capable of generating. The specification also allows us to evaluate the changes that are brought about by notified SEZs in the region (α2 ) which could reveal the mechanism behind the increase in light activity following SEZ notification shown in the previous section. In order to analyze the formal sector data along with the less frequently available informal sector data, we use the following framework: yirt = α0 + α1 Tr + α2 Tr AF T ERt + βi + γr + δt + εirt (3) yirt is the average firm-or worker-level outcome variable (such as logged worker wages or firm size) in industry i in region r in year t. Tr is the treatment indicator, which takes the value of 1 for regions that were treated with at least one operational SEZ before the post-treatment period as stated in Table A.3 in Appendix. The value is 0 for regions that have at least one SEZ notified before this time but none operational yet. AF T ERt is the time indicator, which takes the value 1 for the post-treatment period, and 0 otherwise. Region, industry43 and year fixed effects are included with standard errors cluster-robust at the level of the region. α2 is the coefficient of interest, which describes the change in an outcome such as average employment of a firm located in the treated region with respect to the control region due to the presence of at least one operational SEZ. Depending on our analysis of the formal or the informal sector, our expectations differ about the effect that SEZs are bound to have on firms. For formal sector firms, we expect positive productivity spillovers that encourage production, and boost investment and employment. We also expect an increase in wages paid by the formal firm due to the increase in productivity as well as greater demand for labor. For firms in the informal sector, we expect a priori that the increase in labor demand and wages in the formal sector may lead to a reduction in sustenance-level self-employment with workers moving to formal firms that are expanding, paying more and offering greater job security. The increase in demand for local goods by SEZs could also motivate more productive firms in the informal sector to pay the cost of being regulated and gain from the increased profitability of being formal. This would result in a reduction in the overall size of the informal sector in terms of employment, assets and production. 43 In an alternate specification, we consider region-industry fixed effects. 20