Policy Research Working Paper 10969 Afghanistan’s New Economic Landscape Using Nighttime Lights to Understand the Civilian Economy after 2021 Oscar Barriga Cabanillas Walker Kosmidou-Bradley Silvia Redaelli Eigo Tateishi Ivo Teruggi Poverty and Equity Global Practice November 2024 Policy Research Working Paper 10969 Abstract This study uses nighttime lights to examine the evolution the new economic reality in the country indicates a sig- of economic activity in Afghanistan after the August 2021 nificant economic recovery concentrated in previously regime change. A year later, nighttime luminosity had conflict-affected regions. By 2023/24, civilian luminosity dropped by 20 percent, with two-thirds of this decline had surpassed pre-2020/21 levels by 10.5 percent while, tied to the pre-planned international military withdrawal. in contrast, official gross domestic product indicates an To focus on local economic activity, the study filters out economy that is one-quarter smaller. The findings high- light emissions from foreign military installations, which light changes in economic dynamics, including increased accounted for up to 30 percent of lights over the past informality, shifts in the geographic distribution of activity, decade. Using civilian nighttime lights to understand and improved security post-Taliban takeover. This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at obarriga@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 Afghanistan’s New Economic Landscape: Using Nighttime Lights to Understand the Civilian Economy after 20211 Oscar Barriga Cabanillas Walker Kosmidou-Bradley Silvia Redaelli Eigo Tateishi Ivo Teruggi Keywords: Nighttime Lights, Synthetic Control, Afghanistan, Gross Domestic Product, Data Scarcity JEL codes: C82, O1, O4, P4, R1 1 This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. Oscar Barriga-Cabanillas (corresponding author) may be contacted at obarriga@worldbank.org. The authors want to thank Robert Beyer, Eric Le Borgne, Mamadou Ndione, Utz Johann Pape, Tobias Akhtar Haque, and Muhammad Waheed who provided substantial comments on this version of the paper. The authors would also like to thank Parth Chawla for his valuable research assistance, and Aldo Morri for his carful editing. We declare that we have no relevant or material financial interests that relate to the research described in this paper. The findings, interpretations, and conclusions expressed in this work do not necessarily reflect the views of The World Bank Group or any affiliated organizations, its Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. Introduction and motivation Following the political events of August 2021, the Afghan economy endured severe shocks. According to official estimates, Afghanistan’s GDP contracted sharply in 2021 (20.7 percent) and continued to decline in 2022 despite the resumption of (off-budget) international aid. Despite a modest GDP growth last year (2.7 percent), official GDP data indicates that by 2023, the Afghan economy was almost one-fourth smaller than in 2020. In contrast to the dynamics outlined in the GDP, other indicators show signs of economic recovery. Despite the freeze of foreign reserves and constraints to correspondent banking relationships with the rest of the world, trade data show that imports and exports have overshot the levels observed in 2020 with total exports doubling and imports growing by about one-quarter. Signs of recovery also emerge from household survey analysis. Before the Taliban takeover, half of Afghanistan's population was already living in poverty, with welfare deteriorating rapidly in the months that followed the regime change. 2 By the fall of 2021 most of the population (70 percent) did not have enough income to satisfy basic needs, and about 80 percent experienced an income decline compared to the previous year. Since then, estimates from subsequent rounds of the Afghanistan Welfare Monitoring Survey (AWMS) indicate a progressive improvement in welfare, with poverty returning to its pre-crisis level, hovering around 50 percent. 3 With the Taliban taking power and the withdrawal of international military forces, active conflict and fighting decreased significantly in Afghanistan. According to ACLED data, the number of security incidents in 2022 and 2023 marked a 91 and 98 percent reduction compared to 2021, while the decline in conflict-related deaths was 74 and 90 percent, respectively (Figure A1). The change in the security situation and the withdrawal of international military forces went hand in hand with the downsizing of Afghanistan’s war economy, with overall public spending in the security sector declining by almost 80 percent. 4 Meanwhile, thanks to an increase in fiscal revenues and the resumption of humanitarian aid, the level of civilian spending (including ITA expenditures and off-budget international aid) is comparable in magnitude to 2019. 5 2 Estimates based on IE-LFS 2021 (spring) indicate a poverty rate of 52 percent. 3 See World Bank, 2023 (AWMS). 4 See World Bank 2024 (ADU). 5 In 2019, total civilian spending amounted to Afn 435 billion, comprising of Afn 283 billion on budget and 152 billion off-budget. In 2022-23, civilian spending amounted to Afn 438 billion, comprising of Afn 98 billion on-budget spending by the Interim Taliban Administration (ITA) and Afn 340 billion off-budget spending by UN, ICRC, and bilateral donors. See World Bank 2024 (ADU). 2 Considering the increasing concerns about the sustainability of international aid moving forward, it is important to understand how Afghanistan’s economy is adjusting to the new context. However, to what extent is post-2021 economic activity in Afghanistan properly captured by official GDP estimates, especially when they reflect the economic structure prevalent in Afghanistan in the early years of the last decade? To shed light on this question, we leverage the well-established relationship between nighttime light data and economic activity (Henderson et al., 2012; Donaldson & Storeygard, 2016; Gibson et al., 2020, among others). Nighttime lights serve as a proxy for economic activity under the assumption that human activity requires light and, consequently, these light emissions reflect the level of economic output. A primary advantage of nighttime lights is that they provide an independent and consistent benchmark of economic activity (Pinkovskiya & Sala-i-Martin 2016), and their relationship with economic activity is independent of potential concerns about measuring GDP (Martinez, 2022). However, light emitted into space does not always reflect local economic dynamism on the ground. In Afghanistan, international military bases were an important source of radiance. In 2012, at the height of international troops' presence, military bases accounted for about 30 percent of all the night light captured by satellites in Afghanistan. While the presence of military bases might have undoubtedly contributed to stimulating the local economy, the radiance directly emitted by military bases was related to conflict dynamics, driven by security spending, and – ultimately – bound to disappear with the withdrawal of international forces in August 2021. We leverage information from high-definition satellite imagery and luminosity data to identify military installations in the country. By filtering the lights emitted by military installations, we reduce potential biases introduced by their luminosity when analyzing the trends and distribution of economic activity in the country. The importance of filtering out radiance from military bases is well represented by the luminosity trends of the city of Kandahar and the military airfield next to it (Figure 1). In 2013, the brightness of the military airfield rivaled that of the city, which possibly also benefited from its economic spillovers, but by 2022 the airfield had completely shut off following the military withdrawal. Consequently, not accounting for the luminosity from the airfield leads to a significant overestimation of the level of local economic activity in the broader Kandahar areas before 2021. Similarly, the size of the local economy contraction following the regime change will also be overestimated since most of the reduction in light emissions was caused by a pre-planned shut-down of the military airfield. 3 Figure 1. Comparing lights from the Kandahar city with the Kandahar Airfield. 2013 2019 2022 Source: Author’s calculation using VIIRS data Note: The figure compares the light intensity between the Kandahar Airfield and the civilian activities around the city proper. The red polygon indicates the airfield, while the black boundaries are the Kandahar city limits. Light emissions extend beyond the perimeters due to sensor oversaturation. Building on these considerations, this paper revisits the use of nighttime lights in Afghanistan to assess the evolution of economic activity following the regime change in 2021. Filtering lights from military installations to track economic activity does not mean bases did not contribute to economic dynamism, either directly through the acquisition of goods and services or indirectly through the provision of security. 6 However, our main assumption is that military bases’ contribution to the local economy was captured by the luminosity of surrounding civilian areas. From this assumption, it follows that the dimming (increasing) in luminosity post-2021 tracks the evolution of economic activity in the country. The main innovation consists of using high-resolution satellite imagery to identify military installations and use this information to create a series of “civilian” night lights which, we argue, better reflect the evolution of local economic activity in Afghanistan over the past two decades. Results of the analysis indicate that (i) compared to total radiance, the “civilian” night lights series better matches the evolution of GDP in Afghanistan, especially after the international military withdrawal began in 2014; (ii) after an economic contraction of about 20 percent in 2021, the local economy has recovered to its 2020 level, a result validated using a synthetic control approach that benchmarks civilian luminosity in Afghanistan with that of neighboring countries; and (iii) the nature and geography of economic activity 6It is not possible to provide a definite answer on the extent of these channels. For instance, foreign military installations depended almost completely on foreign supply chains, limiting their expenditure in the local economy. Similarly, while bases provided security, their presence made these areas a military objective for the insurgency, reducing the net security benefit. 4 in Afghanistan have changed post-Taliban takeover and new baseline surveys are needed to update the country’s National Accounts estimates. The paper is structured as follows. Section 1 presents a literature review on using nighttime lights to measure economic activity. Section 2 provides context on Afghanistan's economic evolution and the role of foreign aid in past growth trends. Section 3 describes the sources of the nighttime light data, the process to identify military installations, and the netting out of their luminosity to construct a civilian nighttime light series. Section 4 presents the paper’s results, and Section 5 concludes. 1. Literature review Our research relates to different branches of economic literature, including best practices to process and clean luminosity data, the understanding of the evolution of local economic activity, and its use to benchmark economic growth when there are concerns about the quality of national account information. In terms of best practices for processing luminosity data, the literature has for a long time recognized the need to address “background noise” from temporary light emissions sources like forest fires, reflection from snowpacks, and gas flares (Elvidge et al., 2009; Abrahams et al., 2018; Gibson et al., 2021; Bluhm & Krause, 2022). In the case of Afghanistan, we are not the first to point out the outsized role of military installations in the total light emitted and their sudden dimming after the withdrawal of international troops from the country (Elvidge et al., 2022). Nevertheless, our contribution consists of a methodology to systematically identify foreign military installations and filter out their radiance to produce a civilian luminosity series. Besides the technical aspects related to processing raw nighttime light data, a frequent concern is that night lights are better at capturing economic activity in urban areas and those in the service and manufacturing sectors (Chen & Nordhaus, 2019). The weaker relation between nighttime lights and agricultural GDP is easily understood in a context where a large percentage of the rural population lacks access to electricity and, even when electricity is available, it is more intensively used for activities that do not generate light at night, such as providing energy for water pumps (Doll & Pachauri, 2010). However, a new generation of satellite sensors has largely addressed these concerns by being able to 5 capture finer traces of luminosity, allowing them to track economic activity outside cities (Gibson et al., 2021). 7 Similarly, there are concerns that the strength of the relationship between economic activity and night light intensity declines steadily with economic development since there are physical limits to the amount of light emitted from any given location. Data validates this concern as the elasticity of GDP to nighttime light is, on average, higher for countries with lower levels of GDP per capita (World Bank, 2017). 8 In terms of understanding economic trends, nighttime lights have been used as an independent benchmark when there are concerns about the quality of GDP data, estimating the size of the informal economy, and assessing the spatial distribution of economic activity. Accurately capturing GDP is particularly challenging in developing countries, especially when there are concerns about the quality of the national account information due to capacity constraints or large segments of the informal economy concealing a significant share of the final goods and services from the statistical agencies (Martinez, 2022). 9 Likewise, informal sectors are only partially captured in GDP data but still produce light. Therefore, as shown by Ghosh et al. 2010 for India, the size of the informal economy can be largely explained by the gap between the official state GDP and the level of economic activity predicted by nighttime light. Additionally, nighttime lights have also served as a metric of economic activity in low-information environments, particularly in conflict-affected areas. This literature has a significant history in Afghanistan. For example, local surges of violence reduced local growth for up to one-quarter but had no long-term effect on economic activity (Galdo et al., 2021). This result mirrors the impacts of lower conflict levels found by de Roux et al. 2024, where nighttime lights indicate an increased level of rural economic activity following the 2016 Colombia peace accords. 7 The capacity to measure economic activity in rural areas has been improved as a new generation of satellites is better able to detect light in rural and low population density areas (Chen and Nordhaus, 2015). Launched in 2011, the VIIRS Day- Night Band (DNB) provides accurate measurements across a broader range of lighting conditions, with higher spatial accuracy and temporal comparability. Moreover, this argument has lost relevance with the large increases in rural electrification in South Asia (from around 60 percent to almost universal in the past 20 years), and the declining share of agriculture over aggregate GDP. Considering a large cross-section of countries around the world, the inverse Henderson elasticities are 0.13 for agriculture, 0.37 for manufacturing, and 0.43 for services, with the relationship being statistically significant for all three sectors (World Bank, 2017). 8 For reference, in our case, the elasticity of GDP to total luminosity is 0.19 for the period 2003-2020, increasing to 0.26 when considering only civilian lights. This result is aligned with the values expected for a country with the level of economic development of Afghanistan, which on average is 0.31 (World Bank, 2017). 9 His results show authoritarian regimes are more likely to overstate (understate) positive (negative) changes in GDP. 6 Closest to our objectives, Sänger et al. (2023) use nighttime lights to estimate the extent of Afghanistan’s economic activity reduction following the regime change in 2021. Their estimates predict an economic contraction of 28 percent between 2021 and mid-2022, close to the reduction reported in the official GDP series. However, they do not account for the role of the pre-committed closure of foreign bases, which would have occurred even in the absence of a traumatic dislocation of economic activity due to sanctions, disruptions to the banking system, and the sudden stop of international aid. 2. Afghanistan context Afghanistan is a landlocked country with one of the lowest GDP per capita levels in the world. The Afghan population is largely rural (75 percent of the total), and the agriculture sector provides income for close to half of its population. Over the last two decades, the performance of the Afghan economy has been severely affected by international aid, conflict, and insecurity, as well as by natural and political shocks. Growth dynamics over the past two decades can be characterized into three distinct phases (Figure 2a). Between 2003 and 2012, Afghanistan experienced a period of high-paced growth, averaging 9.4 percent. Part of this exceptional growth performance can be explained by the high level of aid Afghanistan received in the past decade, which has produced higher aggregate demand for goods, services, and construction. Official development aid and military assistance grew continuously, from US$404 million in 2002 to more than US$15.7 billion in 2010, the equivalent of 98 percent of GDP (Figure 2b). About one‐third of these aid flows went into developing civilian infrastructure and services such as education, health, electricity, and roads (World Bank 2014). Military operations also surged during this period, with large increases in the presence of troops from around 30,000 in 2008 to more than 100,000 by 2012 and the geographical scope of military operations. With the start of international troops withdrawal in 2014, the decline in aid, and the increase in conflict and insecurity, the economy entered a new phase of sluggish growth that lasted till the Taliban takeover in August 2021. Over this period, growth averaged about 2 percent, and per capita income declined. Besides the deterioration in security and heightened political uncertainty, the poor economic performance was spurred by the substantial decline in civilian aid flows, from US$4.2 billion in 2012 to around US$2.8 billion toward the end of the period. Negative growth rates translated into significant deterioration in welfare, with poverty increasing in both the urban and rural areas. In 2019-20, 7 approximately 50 percent of the Afghan population had a consumption level below the national poverty line. Following the political events of August 2021, the Afghan economy entered a phase of sizeable economic contraction. According to official estimates, Afghanistan’s GDP contracted sharply in 2021 (20.7 percent) and continued to decline in 2022 despite the resumption of (off-budget) international aid. Despite a modest GDP growth last year (2.7 percent), official GDP data indicates that in 2023 the Afghan economy was almost one-quarter smaller than in 2020. Figure 2. Main macroeconomic indicators. Panel a. Evolution of GDP (year-on-year growth) Panel b. International grants (billions USD) 12 Civilian 10 Security 8 BILLION USD 6 61% 60% 34% 0% 4 17% 2 - 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 Source: Author’s calculation using NSIA data. Note: GDP data is calculated over a fiscal year (January to December) until 2019, and information in the following years is expressed in a solar calendar (June to March). After that year, the aid flow data includes both in- and off-budget, with data in billions of USD. Of particular concern in the case of Afghanistan is that national account data reflects the economic structure of 2016. The limited infrastructure and statistical capacity have always posed a challenge to obtaining reliable GDP data in the country. The economy was highly informal before the regime 8 change, with about 72 percent of total economic activity in the informal sector, and therefore not properly registered in official statistics (World Bank, 2005; SIGAR, 2018), while the sampling frames used to capture firm activity and relative prices would in any case by 2024 be considered outdated. Nevertheless, all these challenges have been exacerbated by the shifts in the economic structure, with economic activity moving toward informal sectors, a reconfiguration of economic sectors due to changes in international aid flows, and a new spatial distribution of activity driven by a potential peace dividend. 3. Data and methods We rely on two datasets to reconstruct the evolution of economic activity in the country between 2003 and 2023. First, for the long-term trends between 2000 and 2014, we use the Global NPP-Visible Infrared Imaging Radiometer Suite-like –(VIIRS-like) nighttime light dataset developed by Chen et al. (2021). It has a high spatial resolution (500m x 500m) but is only available yearly. Starting in 2014, we used the VIIRS Stray Light Corrected Nighttime Day/Night Band Composites dataset. The dataset provides the same spatial resolution as the VIIRS-like dataset (500m x 500m) but a monthly temporal frequency. Both datasets include basic background noise reduction layers. 10 Figure A2 compares the evolution of total luminosity from both data sources, showing that despite differing levels, they display the same trends. In addition, Annex B discusses potential measurement errors and tests their presence in our data, finding that they play little role in the results presented in the paper. We implement a two-step process to filter out the luminosity produced by military installations. In the first step, we rely on the OpenStreetMap (OSM) land-use layer data and extract polygons labeled as military facilities. This includes the larger and more permanent military installations in the country, but it is not exhaustive as it depends on users’ voluntary inputs. To create a more comprehensive list of military installations, we use the light signatures from the 558 military-labeled polygons identified in the OSM data to detect additional military facilities. For this, 10The Visible Infrared Imaging Radiometer Suite Day-Night Band (VIIRS DNB) data, despite having a limited temporal range, provides a substantial improvement in measurement accuracy, range, and spatial resolution over the earlier Defense Meteorological Satellite Program (DMSP) (Gibson et al. 2020, 2021). The VIIRS-like dataset covers 2000 to 2022 and was obtained from Google Earth Engine (projects/sat-io/open-datasets/npp-viirs-ntl). The background noise filters were developed by the Earth Observation Group, Payne Institute for Public Policy, Colorado School of Mines, generating the VIIRS Stray Light Corrected Nighttime Day/Night Band Composites Version 1. While it is stray-light-corrected, other temporal lights, such as background noises, aurora, fires, and boats, are not filtered out. Both the processing of the VIIRS and VIIRS-like data is available on a GitHub repository: https://github.com/worldbank/afg_viirs_analysis. 9 we first computed the over-time standard deviation of nighttime light values within the OSM-military polygons to discover additional VIIRS pixels with similar standard deviation patterns. We can assume that these additional VIIRS pixels with similar standard deviation patterns to those in the OSM- military polygons are potential military-related lights not captured in the first filtering step. The rationale for using standard deviation as a key statistic to explore potential additional military-related lights is that we can assume that the opening and closing of military bases— with strong light intensity during their operation—yield a higher standard deviation (that is, variation in light intensity) compared to relatively stable and weak artificial lights such from cities, towns, and roads. Figure A3 displays military polygons identified using OSM data, and Figure A4 includes examples of the distinct layout of military bases, which allows them to be visually recognized on satellite imagery. We targeted OSM-military polygons whose area size is larger than, or equal to, 4 km2 to compute the threshold standard deviation. The result shows that the mean standard deviation of these relatively large bases is 8.477. By exploring the VIIRS pixels outside the OSM-military polygons using the threshold (Standard deviation above 8.477), we detected 214 VIIRS pixels as candidates for additional undetected military bases. Using the characteristic layout of military installations, we manually checked the polygons against high-resolution commercial images, identifying 52 additional bases not registered in the OSM database. It is important to note that eliminating exclusion and inclusion errors is difficult. For inclusion errors (i.e., pixels incorrectly labeled as military bases), our visual verification of the base layout makes the misidentification of military pixels. Meanwhile, in the case of exclusion errors (i.e., pixels labeled as civilian but containing a military base), we are confident the unlabeled bases represent those smaller in size and of a more temporary nature. 11 While acknowledging these limitations, the proposed dual filtering process (OSM-military polygons + standard deviation thresholding with manual verification) can capture and filter out the VIIRS pixels with military facilities in Afghanistan as much as possible. This dual filtering process reduces potential bias introduced by military installations in the total light intensity and trends, which previous studies in the country have ignored. 11To capture the light emitted from military installations, it is necessary to delineate a polygon around each base’s perimeter. In bases in isolated areas, polygons can be drawn as to capture as much of the lights emitted, including the overglow into neighboring pixels. However, in the case of areas highly saturated by light, such as Kabul’s city center, it is not possible to completely filter out the radiance from military bases that glows into adjacent civilian areas. This constitutes an additional form of measurement error on our civilian lights that cannot be completely filtered. 10 After applying the basic background noise filters, the total luminosity mask contains 76,754 pixels. To construct the mask of civilian nighttime lights, we subtract 1,324 pixels identified as military bases, whereas 1,110 pixels come from the OSM-military dataset. The paper results use the sum of all the light captured in each pixel of the civilian and military masks, which, in the case of the VIIRS, data, is available at monthly intervals and yearly for the VIIRS-like series. This pixel-level processing allows us to produce civilian and military luminosity series at any level of geographical aggregation. 4. Results Our analysis is divided into three main components. As a first step, we show that civilian night lights correlation with GDP is higher than for total luminosity (including light from military bases). Moreover, we show that the spatial distribution of civilian and total nighttime lights differs due to the concentration of military bases in strategically important areas. Using these two results, we argue civilian lights provide a better benchmark for local economic activity, especially after the withdrawal of the international military presence began in 2014. Our second set of results compares the contraction of luminosity and GDP in 2021 and its evolution in the following years. We show that most of the reduction in total luminosity would have occurred in the absence of any economic dislocation since two-thirds of the contraction was driven by the shutdown of foreign military installations. Tracking the evolution in civilian nighttime light post-2021, we find that civilian luminosity by 2023 is above its 2020 levels, a result we interpret as local economic activity also displaying a significant recovery. We further validate this result using a synthetic control approach that benchmarks civilian luminosity in Afghanistan with that of neighboring countries. Finally, we explore the spatial changes in the distribution of economic activity. Results highlight that all regions experienced the 2021 shock (and its latter recovery) equally. We conclude by arguing for the need to update the country’s National Accounts system to capture the new economic realities after the Taliban takeover. 4.1 Compared with total luminosity, civilian nighttime lights are better at tracking the evolution of economic activity Night light trends follow GDP growth dynamics (Figure 3). Between 2004 and 2013, thanks to significant investments in improving electricity access and initial low levels of electrification, total and civilian luminosity presented high growth rates which averaged 22 and 17 percent, respectively, during 11 the period. 12 This substantial increase in radiance coincides with the period of highest economic growth, foreign aid, and the expansion of military installations. On their part, military installations went from representing a marginal share of total luminosity to about 30 percent of all the lights in the country at its peak in 2013. Reflecting the sluggish economic growth from 2014 until the 2021 Taliban takeover, nighttime lights presented a lower and more inconsistent growth. Over the period, civilian luminosity increased only by 13 percent and presented negative growth rates in four of the seven years covered. Total luminosity experienced a reduction of 4 percent due to the gradual reduction in the international military presence that started in 2014. Yet, military lights still represented close to 18 percent of total light emissions in Afghanistan by 2020. Figure 3: Evolution of total, civilian, and military luminosity. 100,000 90,000 Total (incl. military) 80,000 Civilian 70,000 TOTAL RADIANCE Military 60,000 50,000 40,000 30,000 20,000 10,000 - 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 Source: Author estimations based on VIIRS-like (pre-2014) and VIIRS (post-2014). Notes: Nighttime lights are expressed as total radiance captured during the year in nW/sr/cm^2. The series contains information from the VIIRS-like (2003-2014) and VIIRS (2014-onwards) series. Differences in the trends between economic activity captured by civilian and total night light across different periods are reflected in the correlation with GDP (Table 1). Our results show that between 2003 and 2014, the correlation of GDP with total luminosity (including military bases) and our civilian series was close to 0.9, a value in line with estimates for countries with similar levels of development (Table 1, panel a). 12Between 2003 and 2014, electricity access went from 20 to 90 percent of the population (WDI indicators). This large increase in electricity access helps explain total and civilian luminosity increasing by about four and three-fold between 2004 and 2013. 12 As military lights began a process of continual dimming in 2014, the correlation between total luminosity and GDP fell drastically to only 0.07 for the 2014-2020 period. Nevertheless, the civilian lights series maintains a high correlation with GDP at around 0.75 (panel b). Importantly, this drop in the correlation is robust to including different begin and end periods. 13 These results highlight that civilian lights present a consistent series to track the evolution of economic activity in Afghanistan. More importantly, they represent a better baseline for assessing the evolution of economic activity post-2021. Table 1. Correlation of GDP and nighttime lights (multiple periods) GDP Panel a. 2003-2013 Total radiance (incl. military lights) 0.95 Nighttime lights Civilian lights 0.94 Panel b. 2014-2020 Total radiance (incl. military lights) 0.07 Civilian lights 0.75 Panel c. 2003-2020 Total radiance (incl. military lights) 0.95 Civilian lights 0.97 Source: Author estimations using VIIRS-like and NSIA GDP data. Note: Total radiance includes military lights. The correlations include the official GDP series and VIIRS-like luminosity (2003-2020). A similar analysis using the VIIRS (2014-2020) series was performed, and all results hold. Correlation results are robust to changes in the start and end years and are consistently lower for the 2014-2020 period. The analysis of local economic activity using civilian night light reveals distinct patterns from the ones implied by total luminosity (Figure 4). For instance, in 2012, the Kabul region was the main economic activity hub; however, while it represented 43.4 percent of total luminosity, it accounted for as much as 58 percent of civilian light emissions. On the other hand, in the Southwest, the analysis of total nighttime lights would attribute as much as 24.4 percent of the national economic activity to this region against only 13.9 percent when using civilian light emissions. With the gradual withdrawal of foreign troops and the progressive winddown of military operations, the difference between civilian and total nighttime lights closed over time, with only significant differences remaining in the Central region by 2020. Still, the persistent existence of this gap for almost two decades highlights the importance of filtering out light emissions from military installations when 13 Correlation results are robust to changes in the start and end years and are consistently lower for the 2014-2020 period. Figure A5 shows the evolution of GDP, as well as total and civilian luminosity. 13 understanding the distribution of economic activity around the country, as they artificially increased the perceived level of activity in military-relevant areas (see World Bank, 2018). Figure 4. Total and civilian luminosity (Share of national). 2012 2016 2020 Source: Author estimations based on the VIIRS-like and VIIRS. Note: Regions are defined as Central (Kapisa, Logar, Maydanwardag, Panjsher, Parwan); East (Kunarha, Laghman, Nangarhar, Noristan); Kabul (Kabul), North (Balkh, Faryab, Jawzjan, Samangan, Sar-e-Pul); Northeast (Badakhshan, Baghlan, Kunduz, Takhar); South (Ghazni, Khost, Pakteka, Paktya); Southwest (Hilmand, Kandahar, Nemroz, Uruzgan, Zabul); West (Badghes, Farah, Hirat); West-central (Bamyan, Daykundi, Ghor). 4.2 Civilian lights show economic activity is growing post 2022 To the extent that the series of civilian night light data provides a better proxy for local economic activity, the analysis of trends in civilian lights also indicates a different magnitude of the shock that accompanied the Taliban takeover and the extent of the economic recovery, or lack thereof. Total nighttime lights in 2021 fell by almost 23 percent following the regime change. However, almost two-thirds of the reduction was caused by the pre-committed military withdrawal, with our estimates showing that the decrease in civilian luminosity over the same period was only 7.1 percent. 14 Following the closedown of the military bases, the civilian nighttime soon encompasses all the luminosity emitted from the country by late 2021 (Figure 5a). Civilian nighttime lights indicate the economy further contracted in 2022/23 but enjoyed a significant recovery in the following year (Figure 5b). Importantly, official GDP data does not fully reflect the recovery in 2023/24. While both series showed an uptake in economic activity, the recovery in civilian luminosity (21.1 percent) was orders of magnitude higher than that of GDP (2.7 percent). Annex B provides robustness tests that show that the significant increase in civilian lights is not the result of 14Result from comparing the level of civilian nighttime lights in the solar year 2021 with the total level of lights present in the solar year 2020. 14 measurement error on the VIIRS data, as well as tests that confirm the direction and magnitude of our estimates under additional filters on the luminosity data. The high year-on-year growth rates of civilian nighttime lights imply that the total civilian luminosity had more than recovered by 2023/24 and is 10.5 percent above its 2020/21 level (Figure 5c). In contrast, total luminosity is still 8 percent below its 2020/21 recorded level. The reason for this discrepancy is that total luminosity includes in the 2020/21 benchmark the significant contribution of foreign military installations which, as explained, were due to be shut down. We argue that civilian lights provide a better benchmark than total luminosity when assessing Afghanistan's current economic dynamics. While foreign military bases undoubtedly contributed to stimulating the economy, their radiance was related to conflict dynamics and security spending, and the bases’ contribution to the local economy was captured by the luminosity of surrounding civilian areas. Ideally, having a series of GNP would help to complement the argument about foreign military bases' contribution to the country’s economy. Unfortunately, these data are not available for Afghanistan. Moreover, the 21.1 percent increase in civilian luminosity during the last year should not be interpreted as the economy expanding at the same rate. Calculating the actual change in GDP given the growth in civilian lights requires properly calibrating the GDP to nighttime light elasticity. However, properly estimating this elasticity is technically challenging, especially when a new economic reality with lower conflict and flows of international aid likely influenced the sectoral and spatial distribution of the economy, thus altering the value added per unit of light. We will further develop these ideas in the next section. 15 Figure 5. Evolution of nighttime and real GDP. Panel a. Relative reduction in nighttime lights: Civilian and total luminosity Total (incl. military) 12.8 Civilian 12.7 RADIANCE (IN LOGS) 12.6 12.5 12.4 12.3 12.2 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2018 2019 2020 2021 2022 2023 2024 Panel b. Year-on-year growth Panel c. Levels relative to 2020/21 30% 30% Civilian NTL Total NTL GDP Civilian NTL Total NTL GDP 21.1% 21.1% 20% 20% 10.5% 10% 10% 2.7% 0% 0% -1.8% -1.8% -6.2% -8.0% -10% -7.1% -22.6% -10% -7.1% -22.6% -8.8% -24.0% -20% -20.7% -20% -30% -30% -20.7% -23.7% -25.7% -40% -40% 2021/22 2022/23 2023/24 2021/22 2022/23 2023/24 Source: Author estimations using VIIR and NSIA GDP data. Note: Panel a vertical red line indicates the change in the political regime. GDP calculated on a solar calendar. To capture the shock due to the change in the political regime, luminosity growth rates use only last three quarters of the solar year; conclusions remain if full solar year is used. To confirm the recovery patterns implied by the analysis of civilian night light data, we implement a counterfactual analysis using a synthetic control approach (Cattaneo et al., 2021). Using the information on night light data emission from neighboring countries, we estimate the counterfactual civilian night light series, i.e., we recreate estimates of night light data emission that could have been observed in the absence of the shock that accompanied the Taliban takeover. This analysis directly answers concerns in the nighttime light literature about luminosity data working best at identifying cross-sectional differences in economic activity than its evolution over time. By using the synthetic 16 control approach, we account for the evolution of luminosity in neighborhood countries to better understand if the civilian luminosity gap with respect to 2020/21 has, in fact, closed. In a nutshell, we use quarterly nighttime light information from neighboring countries' provinces and calculate weights that recreate a new series that closely matches the observed quarterly evolution of the civilian lights in Afghanistan. Specifically, we use information from every quarter between 2014 and 2020 and then predict a counterfactual evolution for Afghanistan's civilian lights from 2021 until 2023. Annex C provides further details on the estimation methodology and data used and several placebo tests that confirm our capacity to predict the evolution of Afghanistan's civilian nighttime lights. The synthetic control needs two main assumptions in order to provide a consistent series to track the evolution of civilian luminosity in the absence of the 2021 regime change. First, there must be no spillovers from treated units to those used to estimate the predicted series. This is equivalent to the events in August 2021 and their economic ramifications not driving changes in luminosity patterns in provinces in the neighboring countries. While it is not possible to completely discard that events in Afghanistan have regional ramifications, we argue that any spillover affecting nighttime lights in neighboring countries was negligible. Second, synthetic control estimates must closely match the civilian nighttime light series in the training period. Figure 6 allows for visual validation, with the synthetic control series closely tracking the observed evolution of civilian luminosity in the shared area, which includes the period used to train the synthetic control model (2014 to 2020). The unshaded part of Figure 6 displays the relative gap between observed civilian lights and the synthetic control series estimated through the synthetic control approach. The figure also includes post-treatment prediction intervals for the counterfactual with at least 90% coverage probability. Interestingly, our series predicted the evolution of civilian lights in the first two quarters of 2021 before the events that led to the regime change, providing further confidence in its capacity to track civilian nighttime lights in the following quarters. We interpret this relative gap between civilian lights and the synthetic control series as the distance to the potential GDP that would be expected given the evolution of nighttime lights in neighboring countries. A significant gap developed between the observed and predicted civilian series precisely in the third quarter of 2021. The timing of the gap leaves little doubt that Afghanistan suffered a major economic shock in the aftermath of the regime change and the following large economic dislocations 17 caused by the withdrawal of international organizations, the lack of access to hard currency, and disruptions in international trade and banking systems. The significant gap that developed in 2021 and 2022 had closed by 2023. In the second half of 2021, civilian light emissions were 84 percent of those expected under the synthetic control counterfactual, with this gap remaining relatively constant until mid-2022. Starting in late 2022, this relative gap began to close. By 2023, while civilian lights remained below their predicted trend, the difference fell well within the prediction’s confidence intervals. 15 Even if the exact value of our synthetic control estimates should be interpreted cautiously, the sharp increase in civilian lights and the closing of the gap with the synthetic control series suggest that Afghanistan’s economy is stabilizing in such a way that matches information sources such as household welfare measures and trade data. Figure 6. Counterfactual evolution of civilian nighttime lights: Synthetic Control approach. Civilian Synthetic control 12.9 12.8 12.7 RADIANCE (LN) 12.6 12.5 12.4 12.3 Training period Predicted period 12.2 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 2018 2019 2020 2021 2022 2023 Source: Author estimations based on VIIRS data. Notes: Radiance reports the logarithm of the civilian lights captured and its synthetic control prediction. The synthetic control weights are calibrated using neighboring country provinces between 2014 and 2020. The post-treatment prediction intervals are estimated as in Cattaneo et al. (2021). Results displayed from 2018 onwards. The shaded area indicates the period used to calibrate the synthetic control weights, and the vertical red line indicates the change in the political regime. 15These intervals are not equivalent to traditional confidence intervals but provide an estimate of the statistical uncertainty in the predicted trends that comes from the sample’s variance (Cattaneo et al. 2021). Even if these conditional intervals are large, they clearly show that civilian lights were significantly below expected values between the third quarter of 2021 and the end of 2022. 18 4.3 Nighttime lights indicate that the spatial distribution of economic activity has changed since 2020 Having assessed the reliability of the trends in local economic activity and the extent of recovery implied by the civilian night light data analysis, we explore the spatial dimension of the shock that accompanied the Taliban takeover and the subsequent economic recovery. The shock that accompanied the regime change was not equally distributed across Afghanistan’s regions. It was the highest in areas most dependent on conflict-related economic activities and aid. In 2021, the reduction in civilian lights concentrated around Kabul (and its surrounding area) and in the South (Kandahar, Helmand and Zabul). In these areas, civilian luminosity fell by more than 15 percent compared with their 2020 level, which is aligned with the significant military presence in those provinces. Meanwhile, large areas of Afghanistan experienced no significant change in civilian lights, with some even displaying marginal increases (Figure 7a). Similarly, the no-conflict dividend and the extent of economic recovery have not been equally distributed (Figure 7b). The level of local economic activity captured by night-time lights in 2023 was higher compared to the civilian lights baseline of 2020 in most provinces, but the increment was particularly pronounced in Khost (63.4 percent), and in Hirat, Nemroz, Badakhshan and Noristan (about 38 percent). The marked exceptions are Parwan, where civilian luminosity in 2023 is 43.3 percent lower than two years prior; the provinces of Kapisa and Zabul, with a luminosity about 11 percent lower; and Kabul, where civilian lights emitted 8.1 percent less radiance in 2023 compared to 2020. 16 In the case of Parwan, it is important to consider that 75 percent of all the nighttime lights in the province during 2020 16 were produced by military installations. Thus, the economic impact of the military withdrawal was highest in the province. 19 Figure 7: Change in civilian nighttime lights relative to 2020. Panel a. 2021 Panel b. 2023 Source: Author estimations based on VIIRS data. Notes: Data for 2021 compares last two quarters with the same period in 2020. Regions are defined as Central (Kapisa, Logar, Maydanwardag, Panjsher, Parwan); East (Kunarha, Laghman, Nangarhar, Noristan); Kabul (Kabul), North (Balkh, Faryab, Jawzjan, Samangan, Sar-e-Pul); Northeast (Badakhshan, Baghlan, Kunduz, Takhar); South (Ghazni, Khost, Pakteka, Paktya); Southwest (Hilmand, Kandahar, Nemroz, Uruzgan, Zabul); West (Badghes, Farah, Hirat); West- central (Bamyan, Daykundi, Ghor). 20 The combined effect of differences in the impact of the regime change shock and in the extent of economic recovery have determined a change in the spatial distribution of economic activity in Afghanistan (Figure 8). This result, we argue, is linked to the no-conflict dividend and the emergence of new economic activities in areas previously more affected by the conflict. In particular, while Kabul remains the country's main contributor to the country’s overall level of economic activity (26 percent of total civilian lights in 2023), its share fell by 14 percent compared to 2020. The opposite situation occurred in the north and west regions of the country, which increased their contribution to total economic activity during the period. Figure 8. Regional civilian lights (% of National) Source: Author estimations based on VIIRS data. Note: Regions are defined as Central (Kapisa, Logar, Maydanwardag, Panjsher, Parwan); East (Kunarha, Laghman, Nangarhar, Noristan); Kabul (Kabul), North (Balkh, Faryab, Jawzjan, Samangan, Sar-e-Pul); Northeast (Badakhshan, Baghlan, Kunduz, Takhar); South (Ghazni, Khost, Pakteka, Paktya); Southwest (Hilmand, Kandahar, Nemroz, Uruzgan, Zabul); West (Badghes, Farah, Hirat); West-central (Bamyan, Daykundi, Ghor). 5. Conclusions After filtering out the light from military installations, we show that the resulting series of civilian nighttime lights correlates more with past economic activity. Moreover, our results showcase that civilian lights represent a better baseline for understanding the evolution of economic activity after the 2021 events. Civilian lights indicate the local economy has recovered to its 2020 level. However, a new spatial distribution of luminosity hints at changes in economic dynamics arising from the end of active conflict, which represented a significant deterrent for sustained economic growth. The changes in luminosity are likely to reflect higher economic activity in the informal sector, outside the capital, and in rural areas. Unfortunately, we lack the information to translate changes in the spatial distribution of civilian lights to monetary values. While higher luminosity indicates more local economic dynamism, it is difficult 21 to assign a concrete monetary value to the goods produced without additional on-the-ground information. Since Afghanistan has never produced provincial or regional GDP information, strong assumptions would be required to translate the increments in provincial luminosity to GDP. Official GDP data shows a slight recovery in economic activity in 2023; however, this recovery falls well below the dynamics detected by civilian lights. The divergence between civilian luminosity and official GDP serves as an appeal for improving the national account system in Afghanistan. The current system of national accounts uses 2016 as the base year and input surveys dating back to 2011- 2014, a period when aid-driven economic activity favored service sector expansion in urban areas and conflict limited investments and economic development, especially in rural areas. Therefore, improving GDP estimation would require a new sampling frame that properly accounts for sectoral shifts, the spatial distribution of economic activity, and the significant increases in the informal sector. Moreover, considering potential changes in the households' consumption baskets, a new expenditure survey will guarantee accurate inflation tracking. Our results open a broader research agenda that uses geospatial data to track the evolution of economic activity and the distributional implications of the new economic reality. First, using more granular luminosity data at the level of settlements can provide additional high-resolution information on how economic activity has evolved and shifts in the economy's sectoral composition. This settlement-level analysis can shed light on the reasons for the weaker economic activity in Kabul and better understand the economic spillovers from military bases. Second, given its role in providing employment, it is crucial to understand the evolution of the agriculture sector, including changes in productivity and crop composition, as well as improvements in the functioning of markets in rural areas. Finally, geospatial indicators can help to track trade patterns both within the country and across borders. 22 References Abrahams, A., Oram, C. and Lozano-Gracia, N., 2018. Deblurring DMSP nighttime lights: A new method using Gaussian filters and frequencies of illumination. Remote Sensing of Environment, 210, pp.242-258. Andersson, M., Hall, O., & Archila, M.F. (2019). How data-poor countries remain data poor: Underestimation of human settlements in Burkina Faso as observed from nighttime light data. ISPRS International Journal of Geo-Information, 8(11), 498. Barriga-Cabanillas, O.E., Chawla, P., Redaelli, S. and Yoshida, N., 2023. Updating Poverty in Afghanistan Using the SWIFT-Plus Methodology (No. 10616). The World Bank. Bluhm, R. and McCord, G.C., (2022). What can we learn from nighttime lights for small geographies? Measurement errors and heterogeneous elasticities. Remote Sensing, 14(5), 1190. Cattaneo, M.D., Feng, Y. and Titiunik, R., 2021. Prediction intervals for synthetic control methods. Journal of the American Statistical Association, 116(536), pp.1865-1880. Ciarli, T., Kofol, C. and Menon, C., 2015. Business as unusual. An explanation of the increase of private economic activity in high-conflict areas in Afghanistan. SERC discussion papers (SERCDP0182). Chen, X., and Nordhaus, W.D., 2015. A test of the new VIIRS lights data set: Population and economic output in Africa. Remote Sensing, 7(4), pp.4937-4947. Chen, X., and Nordhaus, W.D., 2011. Using luminosity data as a proxy for economic statistics. Proceedings of the National Academy of Sciences, 108(21), pp.8589-8594. Chen, X., and Nordhaus, W.D., 2019. VIIRS nighttime lights in the estimation of cross-sectional and time-series GDP. Remote Sensing, 11(9), p.1057. Chen, H., Sun, C., Xiong, X., Sarid, G. and Sun, J., 2021. SNPP VIIRS day night band: ten years of on-orbit calibration and performance. Remote Sensing, 13(20), p.4179. Dahmani Scuitti, A., Knippenberg, E., Kosmidou-Bradley, W., Belanger, J.L., 2023. Geospatial Analysis of Displacement in Afghanistan. Policy Research Working Papers; 10583. World Bank, Washington, DC. Donaldson, D. and Storeygard, A., 2016. The view from above: Applications of satellite data in economics. Journal of Economic Perspectives, 30(4), pp.171-198. Elvidge, C.D., Zhizhin, M., Ghosh, T., Hsu, F.C. and Taneja, J., 2021. The annual time series of global VIIRS nighttime lights was derived from monthly averages from 2012 to 2019. Remote Sensing, 13(5), p.922. Elvidge, C. D., Ziskin, D., Baugh, K. E., Tuttle, B. T., Ghosh, T., Pack, D. W., & Zhizhin, M. (2009). A fifteen-year record of global natural gas flaring derived from satellite data. Energies, 2(3), 595–622. Galdo, V., Kitzmueller, 23 Elvidge, Christopher, Tilottama Ghosh, and Mikhail Zhizhin. "The dimming of lights in Afghanistan 2022 versus 2021." Payne Institute Commentary Series: Research (2022). Galdo, V., Acevedo, G.L. and Rama, M., 2021. Conflict and the composition of economic activity in Afghanistan. IZA Journal of Development and Migration, 12(10). Gibson, J., Olivia, S. & Boe‐Gibson, G. (2020). Night lights in economics: Sources and uses. Journal of Economic Surveys, 34(5), 955–980. Gibson, J., Olivia, S., Boe-Gibson, G., & Li, C. (2021). Which night light data should we use in economics, and where? Journal of Development Economics, 149, p.102602. Ghosh, T., Elvidge, C., Sutton, P. C., Baugh, K. E., Powell, R., & Anderson, S. (2010). Shedding light on the global distribution of economic activity. The Open Geography Journal. 3, 147-160 Henderson, J.V., Storeygard, A. and Weil, D.N., (2012). Measuring economic growth from outer space. American Economic Review, 102(2), 994–1028. Martinez, L.R., (2022). How much should we trust the dictator’s GDP growth estimates? Journal of Political Economy, 130(10), pp.2731-2769. NSIA, (2020). Population Projection and Its Impact on the Development. Mimeo, Kabul, Afghanistan. National Statistic and Information Authority. Pinkovskiy, M. and Sala-i-Martin, X., (2016). Lights, camera… income! Illuminating the national accounts-household surveys debate. The Quarterly Journal of Economics, 131(2), pp.579-631. Sänger, T., Kapstein, E. and Sircar, R., 2022. Estimating the Collapse of Afghanistan's Economy Using Nightlights Data. Available at SSRN 4287663 World Bank. 2014. Afghanistan - Pathways to inclusive growth. Washington DC: The World Bank. World Bank. 2017. Growth out of the Blue. South Asia Economic Focus. Washington DC: The World Bank. World Bank. 2022. Afghanistan Development Update: Adjusting to the new realities. October 2022. Washington DC: The World Bank. World Bank. 2023. Afghanistan Welfare Monitoring Survey: Round 3. October 2023. Washington DC: The World Bank. World Bank. 2024. Afghanistan Development Update: Navigating economic recession and deflation. April 2024. Washington DC: The World Bank. 24 Annex Annex A: Additional figures and tables Figure A1. Conflict intensity (Total fatalities) 50,000 All conflict-related events 40,000 Total fatalities from political violence 30,000 20,000 10,000 - 2017 2018 2019 2020 2021 2022 2023 Source: Author’s calculation using ACLED data. Figure A2. Comparison of total luminosity VIIRS-like and VIIRS data 140,000 VIIRS-like VIIRS 120,000 100,000 80,000 60,000 40,000 20,000 - 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 Source: Author’s calculations using VIIRS-Like and VIIRS data. 25 Figure A3. Examples OSM polygons with military installations. Source: Author’s calculations from OSM data Figure A4. Examples of military installations identified through our two step process. Source: Author’s calculations Note: The figure displays the characteristic layout of foreign military installations in Afghanistan. 26 Figure A5. Evolution of luminosty and GDP 1,300 1,200 85,000 1,100 TOTAL LUMINOSITY 1,000 65,000 900 GDP 45,000 800 700 25,000 600 Total (incl. military) Civilian GDP 500 5,000 400 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 Source: Author’s calculations using VIIRS-Like data and NSIA GDP data. Note: GDP in constant billions of Afg. 27 Annex B: Further data validation There are several technical challenges in working with nighttime light data, requiring significant preprocessing steps before economists can use it. These steps include accounting for seasonal factors, like differences in summer glare and annual light usage patterns, as well as filtering out "ephemeral" light sources such as auroras, fires, and gas flares (Gibson et al., 2021). Physical factors also play a role in how the data is captured. These include measurement errors due to sensor oversaturation over high- lit areas (top coding), blurring of adjacent pixels when the sensors lack the resolution to capture luminosity at sufficient resolution, and calibration of measurements (Bluhm and Krause 2018). Additionally, atmospheric conditions, such as the presence of clouds, can limit the capacity to collect consistent luminosity measures. Implementing well-accepted background noise filters and using a single luminosity measure address several of the previous challenges. In our case, we rely on VIIRS satellites' improvements, which have improved optical instruments that minimize, to a large degree, the challenges present in previous generations of sensors (Abrahams et al. 2018). As a first check on the data, we verify the consistency of the reading at the pixel level. Out of 74,106 pixels in the civilian mask, all the pixels (99.04 percent) consistently registered monthly readings. The remaining 0.96 percent include pixels that appear on and off due to complete cloud coverage. We also assess if changes in the distribution of cloud coverage can explain the higher luminosity in 2023. Table B1 shows that pixels are, on average, only half of the time free of cloud cover. The share of cloud-free days is relatively equally distributed across the pixels with low (quintile 1) and high (quintile 5) radiance. Interestingly, the average share of cloud-free coverage pixels is marginally lower in 2023 than in 2020. A similar result occurs when calculating cloud-free coverage at the regional level (Figure B1). These results indicate that, on the aggregate, changes in atmospheric conditions did not drive the increase in luminosity detected in 2023. 28 Table B1: Could-free coverage pixel-level measurements (Civilian night time lights) Pixel radiance quintiles in 2020 Quintile 1 2 3 4 5 Year 2020 2023 2020 2023 2020 2023 2020 2023 2020 2023 Average 48.0% 46.5% 47.7% 45.1% 47.2% 44.7% 47.8% 45.1% 48.7% 45.2% Min 11.8% 11.8% 22.7% 15.1% 17.5% 15.3% 23.8% 20.5% 37.8% 30.1% median 49.0% 47.4% 48.2% 45.8% 47.9% 45.2% 48.2% 45.2% 49.0% 45.5% max 57.3% 54.2% 57.3% 54.5% 57.3% 54.5% 57.3% 54.2% 57.3% 54.2% Source: Author’s estimations using VIIRS data including only civilian pixels. Figure B1: Could-free coverage pixel-level measurements at the regional level Civilian night time lights Kabul Southwest Central West North Northeast East South West-central Source: Author’s estimations using VIIRS data including only civilian pixels . 29 We also assess for the role of outliers in the main results of the paper. Despite using standard background noise filters, the data present significantly large tails for all the regions (Figure B2). To estimate the role of high-luminosity pixels, we estimate the paper's main results after dropping pixels with luminosity levels in in each region's top 1 percent. Figure B2. Pixel-level distrinution of luminosity: Civilian night time lights Source: Author’s estimations using VIIRS data including only civilian pixels. Pixels in the upper 1 percent of the distribution are a significant share of regional luminosity, contributing almost 10 percent of all the civilian lights captured. This large contribution to total radiance is caused by the pixel’s location at the center of the major urban areas. Providing confidence in the robustness of our results in the presence of potential measurement errors, results on the distribution of civilian lights hold even after removing high-luminosity pixels. 30 Table B2: Pixel-level distrinution of luminosity Civilian night time lights Percentage of total Change in total luminosity 2020-2023 luminosity (upper 1%) 2020 2023 Total luminosity Bottom 99 percent Kabul 13% 11% -9% -7% Central 10% 9% -23% -22% East 11% 10% 7% 8% North 11% 10% 14% 15% Northeast 11% 9% 19% 22% South 11% 11% 37% 37% Southwest 11% 13% 9% 7% West 13% 18% 38% 29% West-central 8% 6% 22% 25% Author’s estimations using VIIRS data including only civilian pixels. Upper 1 percent of pixels by luminosity identified from the year- region specific distribution. 31 Annex C: Synthetic controls methodology We use the synthetic control methodology to construct a counterfactual for the evolution of civilian luminosity in Afghanistan. The synthetic control approach estimates the effect of an intervention by calculating a weighted average of untreated units that best resembles the civilian luminosity evolution prior to the first quarter of 2021. The counterfactual series uses these weights to project the expected evolution of the civilian luminosity in the absence of regime change. The effect of the regime change (treatment) is then estimated by comparing the value of the observed civilian nighttime lights after August 2021 (post-treatment outcomes) with the predicted values from the synthetic control approach (counterfactual series). In our case, the predicted series of civilian nighttime lights uses the weighted average of luminosity from provinces in the Islamic Republic of Iran, Pakistan, Tajikistan, Turkmenistan, and Uzbekistan. These data received the same basic background filtering as Afghanistan’s, except for the netting out of military installations. Table C1 provides descriptive stats for all the country provinces included in the synthetic control methodology. For the synthetic control approach to accurately predict the evolution of civilian lights, several assumptions must be met. First, the weights must closely match the civilian nighttime lights series prior to 2021. Second, there must be no spillovers from treated units to those used to estimate the predicted series. In our case, this means that the events in August 2021 and their economic ramifications should not drive changes in luminosity patterns in neighboring provinces. While it is not possible to completely rule out regional ramifications from events in Afghanistan, we argue that given the large spatial dispersion of the non-Afghan provinces used in the estimation of the synthetic control approach, any spillover effects on the levels or trends in nighttime lights radiance in neighboring provinces are, if present, small. Additionally, following recent developments in the synthetic control literature (Cattaneo et al. 2021), our approach includes prediction intervals that reflect statistical uncertainty. These intervals account for two sources of randomness. First, there is potential misspecification and noise in the assignment of the synthetic control weights during the pre-treatment period. This refers to the inaccuracy in selecting the optimal weights for constructing the synthetic control. For example, provinces with higher variation will lead to less accurate counterfactual post-treatment series and larger prediction intervals. Second, random errors and unobservable factors affect outcomes in the post-treatment 32 period. Larger variation in the control units during the prediction period also leads to larger prediction intervals, particularly when occurring in the units assigned the highest weights. We calibrate the weights for each of the neighboring countries’ provinces by minimizing the errors such that the weighted average of luminosity most closely matches Afghanistan’s luminosity in the pre-treatment period. The estimated weights are given in Table C2. We use the quarterly data between 2014 and 2020 to calibrate the weights, as it completely isolates the weight calculation from anticipating the 2021 foreign military withdrawal. Our weights closely replicate the evolution of civilian night light radiance in Afghanistan prior to 2021, providing confidence in the synthetic control capacity to track civilian nighttime lights between 2021 and 2023. Moreover, they closely track the evolution of civilian lights in the first two quarters of 2021 before the events that led to the regime change. We conducted a placebo analysis to confirm the robustness of our approach further. We estimated the weights using quarterly data between 2014 and 2018, comparing the counterfactual series with the observed evolution of civilian lights until the last quarter of 2020. Our results closely tracked the luminosity during the period. This result provides further confidence that the synthetic control approach can track the evolution of civilian lights following the regime change of 2021 (Figure C1). Figure C1: Placebo test: Predicting civilian luminosity between 2018 and 2020. Source: Author’s calculations using VIIRS data. Note: The synthetic control weights are calibrated using neighboring country provinces between 2014 and 2018. The post-treatment prediction intervals are estimated as in Cattaneo et al. (2021). 33 Table C1: Percent of national civilian luminosity for each province in the synthetic control Iran, Islamic Rep. Share national radiance (%) Share total radiance (%) Province 2014 2020 2023 2014 2020 2023 Ardebil 1.2% 1.3% 1.2% 0.8% 0.9% 0.8% Bushehr 5.2% 5.0% 4.9% 3.5% 3.4% 3.4% Chaharmahal-o bakhtiyar 1.0% 1.1% 1.0% 0.7% 0.7% 0.7% East Azarbayejan 3.9% 4.0% 3.5% 2.6% 2.7% 2.4% Esfahan 6.7% 6.6% 6.1% 4.5% 4.5% 4.2% Fars 7.1% 7.3% 7.1% 4.8% 5.0% 4.9% Ghazvin 1.7% 1.6% 1.6% 1.2% 1.1% 1.1% Ghom 1.4% 1.3% 1.2% 0.9% 0.9% 0.9% Gilan 2.1% 2.1% 2.1% 1.4% 1.5% 1.4% Golestan 1.3% 1.4% 1.4% 0.9% 0.9% 0.9% Hamedan 1.9% 1.7% 1.6% 1.3% 1.1% 1.1% Hormozgan 4.1% 4.0% 4.0% 2.8% 2.7% 2.8% Ilam 2.0% 2.1% 2.1% 1.4% 1.4% 1.4% Kerman 4.2% 4.4% 4.3% 2.8% 3.0% 2.9% Kermanshah 2.3% 2.1% 1.9% 1.6% 1.4% 1.3% Khorasan 7.2% 8.2% 8.0% 4.9% 5.6% 5.5% Khuzestan 14.4% 13.7% 15.6% 9.8% 9.3% 10.8% Kohgiluyeh va boyerahma 2.3% 1.7% 2.8% 1.5% 1.2% 2.0% Kordestan 1.2% 1.3% 1.4% 0.8% 0.9% 1.0% Lorestan 1.7% 1.7% 1.7% 1.2% 1.2% 1.1% Markazi 2.5% 2.4% 2.4% 1.7% 1.6% 1.7% Mazandaran 3.7% 3.9% 3.6% 2.5% 2.7% 2.5% Semnan 1.5% 1.6% 1.3% 1.0% 1.1% 0.9% Sistan-o baluchestan 2.4% 2.4% 2.4% 1.6% 1.7% 1.6% Tehran 10.7% 10.3% 10.2% 7.3% 7.0% 7.1% West Azarbayejan 2.7% 2.8% 2.7% 1.8% 1.9% 1.8% Yazd 2.4% 2.5% 2.6% 1.6% 1.7% 1.8% Zanjan 1.2% 1.3% 1.2% 0.8% 0.9% 0.9% Pakistan Share national radiance (%) Share total radiance (%) Province 2014 2020 2023 2014 2020 2023 Balochistan 5.1% 5.5% 4.8% 0.7% 0.9% 0.7% Islamabad 2.7% 3.0% 3.4% 0.4% 0.5% 0.5% KP 11.4% 12.0% 11.4% 1.6% 1.9% 1.8% Punjab 47.6% 53.6% 55.7% 6.8% 8.5% 8.6% Sindh 33.3% 25.9% 24.6% 4.8% 4.1% 3.8% 34 Tajikistan Share national radiance (%) Share total radiance (%) Province 2014 2020 2023 2014 2020 2023 Badakhshoni Kuni 1.3% 1.6% 1.7% 0.0% 0.0% 0.0% Khatlon 30.4% 29.7% 28.6% 0.3% 0.4% 0.4% Sogd 23.4% 26.9% 27.5% 0.2% 0.4% 0.4% Tadzhikistan Territories 44.8% 41.9% 42.1% 0.5% 0.6% 0.6% Turkmenistan Share national radiance (%) Share total radiance (%) Province 2014 2020 2023 2014 2020 2023 Chardzhou 14.3% 12.6% 13.6% 1.3% 0.9% 0.9% Mary 20.4% 17.5% 18.0% 1.8% 1.2% 1.1% Tashauz 9.7% 8.7% 8.6% 0.9% 0.6% 0.5% Turkmenistan Territories 55.5% 61.1% 59.9% 4.9% 4.3% 3.8% Uzbekistan Share national radiance (%) Share total radiance (%) Province 2014 2020 2023 2014 2020 2023 Andijan 7.9% 7.6% 7.6% 0.6% 0.6% 0.6% Bukhara 10.7% 7.9% 7.6% 0.9% 0.6% 0.6% Fergana 6.6% 7.6% 8.4% 0.5% 0.6% 0.7% Jizzakh 4.0% 3.6% 4.1% 0.3% 0.3% 0.3% Karakalpakstan 2.5% 4.0% 3.6% 0.2% 0.3% 0.3% Kashkadarya 16.5% 12.5% 11.8% 1.3% 0.9% 0.9% Khorezm 1.6% 2.5% 2.7% 0.1% 0.2% 0.2% Namangan 6.7% 6.0% 6.1% 0.5% 0.5% 0.5% Navoiy 4.2% 4.9% 4.6% 0.3% 0.4% 0.4% Samarkand 7.3% 8.9% 10.7% 0.6% 0.7% 0.9% Sirdarya 2.7% 2.9% 2.9% 0.2% 0.2% 0.2% Surkhandarya 4.0% 5.0% 4.9% 0.3% 0.4% 0.4% Tashkent 16.7% 17.3% 17.6% 1.3% 1.3% 1.4% Tashkent city 8.5% 9.2% 7.6% 0.7% 0.7% 0.6% Source: Author estimations based on VIIRS data. 35 Table C2: Provinces in synthetic control with their corresponding weights Country Province Weight Iran, Islamic Rep. Fars 17.8% Iran, Islamic Rep. Zanjan 7.8% Iran, Islamic Rep. Khorasan 5.1% Iran, Islamic Rep. Kohgiluyeh va Boyerahmad 2.6% Pakistan Balochistan 30.7% Tajikistan Tadzhikistan Territories 0.5% Turkmenistan Tashauz 12.3% Uzbekistan Tashkent 5.4% Uzbekistan Jizzakh 4.4% Uzbekistan Khorezm 4.1% Uzbekistan Sirdarya 3.3% Uzbekistan Bukhara 2.6% Uzbekistan Kashkadarya 2.5% Uzbekistan Namangan 0.2% Source: Author estimations based on VIIRS data. 36