WPS8385 Policy Research Working Paper 8385 Obstacles on the Road to Palestinian Economic Growth Roy van der Weide Bob Rijkers Brian Blankespoor Alexei Abrahams Development Research Group & Development Data Group March 2018 Policy Research Working Paper 8385 Abstract This paper quantifies the impact of market access on traversal times. These are combined with population data local GDP in the West Bank, proxied by nighttime lights, to construct a time-varying market access measure for each using the deployment of road closure obstacles by the locality. Market access has a significant and substantial Israeli army between 2005 and 2012 as a quasi-natural effect on local light emissions. This association is robust experiment generating exogenous temporal and spatial to controlling for conflict, and strengthens when market variation in accessibility. Minimum travel times between access is instrumented by the number of obstacles located locality pairs are computed using road network and obsta- in a radius between 10 and 25km away from the locality. cles data supplemented with information on checkpoint This paper is a product of the Development Research Group and the Development Data Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at rvanderweide@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 Obstacles on the Road to Palestinian Economic Growth∗ Roy van der Weide Bob Rijkers Brian Blankespoor Alexei Abrahams† March 27, 2018 ∗ 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 of Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the countries they represent. All errors are our responsibility. † Correspondence to: Roy van der Weide (email: rvanderweide@worldbank.org), Bob Rijkers (email: brijkers@worldbank.org), Brian Blankespoor (email: bblankespoor@worldbank.org), Alexei Abrahams (email: alexei abrahams@alumni.brown.edu). We are grateful to Arie Arnon, Shanta Devarajan, Toan Do, Francisco Ferreira, Caroline Freund, Aaditya Mattoo, David McKenzie, Aart Kraay, staff at UN- OCHA and the PCBS, as well as seminar and conference participants at the World Bank, the IMF, and the Tinbergen Institute for providing useful comments. The authors thank the World Banks Mul- tidonor Trust Fund for Trade and Development and the Strategic Research Partnership on Economic Development for funding. 1 Introduction How productive is market access? Answering this question is difficult because the in- frastructure that connects markets typically evolves slowly and non-randomly over time, which makes it challenging to isolate its impact on economic performance. Exploiting variation in the provision of transportation infrastructure such as roads (Faber (2014), Allen and Atkin (2017), Alder (2017), Banerjee et al. (2012)), railroads (Donaldson and Hornbeck (2016), Donaldson (2010)), and waterways (Feyrer (2009)), existing studies typically find sizeable returns to improving connectivity. Research addressing this ques- tion by examining the impact of changes in borders (Redding and Sturm (2008)) has reached similar conclusions. These studies either examine large accumulated changes in transport infrastructure over extended periods of time or changes in accessibility brought about by large isolated shocks. A notable exception is the study by Storeygard (2016), which uses annual variation in transport costs generated by interacting distances between cities with oil prices, to assess short-run impacts of changes in transport costs on economic performance. This paper exploits unique variation in travel times within the West Bank resulting from the deployment of road closure obstacles by Israel to identify the impact of market access on local GDP, proxied by nighttime lights (NTL). The obstacles take the form of manned and unmanned physical barriers, including roadblocks, checkpoints, earth mounds, trenches, and a separation barrier wall. According to the Government of Israel, these restrictions are for the purpose of enhancing the security of Israel and Israeli citizens. As the placement of obstacles is not (directly) driven by local economic performance, but rather by security considerations, it is largely exogenous to local economic conditions. Moreover, the number, intensity, and configuration of obstacles are subject to frequent and unanticipated changes, yielding sizeable short-run variation in local market access. This context thus provides a quasi-natural experiment to assess the impact of short-term fluctuations in accessibility on economic performance. We adopt a market access approach (following Deichmann (1997) and Donaldson and Hornbeck (2016)) to quantify how the intermittent re-configuration of road closure obstacles determines each location’s connectivity to markets. The resulting measure takes into account both the type and intensity of the obstacles as well as their position- ing. This is important since one strategically placed checkpoint can reduce accessibility more severely than a multitude of roadblocks when alternative connections are avail- able. In addition, restrictions may be mutually reinforcing and market access in a given locality is in part determined by obstacles in relatively distant areas, the impact of which simple counts of the number of obstacles would fail to capture. 2 The market access measure is a comprehensive metric of proximity to destination localities’ population, with more importance attached to localities that take less time to reach. Estimating travel times between localities requires data on the road network (including road type and travel speeds for all road segments), detailed information on the precise geolocations and timing of the obstacles (including the separation barrier wall), and on checkpoint traversal times. To construct these data we processed obstacle and road maps provided by UNOCHA and conducted repeated interviews with UN- OCHA officials to obtain estimates of the time cost of traversing each checkpoint at any given time during our sample period. We calculate the minimum travel time be- tween each locality and the governorate capitals in the West Bank via the road network, and recalculate optimal routes and attendant travel times for all of the configurations of obstacles observed each year throughout the 2005-2012 period. We show that year-to-year changes in market access positively predict changes in nighttime light emissions (NTL), which serves as our preferred measure of local eco- nomic performance given the absence of spatially disaggregated GDP measures in the West Bank.1 According to our preferred estimates, a 10% improvement in market access increases local output by 0.6%, assuming a lights-to-GDP elasticity of 0.3 (cf. Hender- son et al. (2012)).2 A back-of-the-envelope calculation suggests that in the absence of road blocks, GDP per capita in the West Bank would have been 4.1% to 6.1% higher each year over our sample period. These ceteris paribus estimates must be interpreted with caution as they depend on the assumed GDP-to-lights elasticity as well as the choice of distance decay parameters, which govern how much weight is attached to destinations that take more time to reach. Moreover, they abstract away from other distortions and uncertainty that may have attenuated the economy’s output response to changes in market access. The main threat to identification is the potential endogeneity of market access, which could result from omitted variables bias. To address this concern we use geo- referenced data on fatalities to control for local conflict intensity, arguably the most likely confounder of a relationship between local economic performance and mobility restrictions. In addition, we construct an instrument for market access, notably the number of obstacles located in a radius between 10 and 25 km away from the locality. These obstacles are orthogonal to local conditions yet an important determinant of local market access. Our IV estimates are substantially larger than our simple OLS 1 NTL are by now widely accepted as a credible proxy for local GDP/economic performance (Chen and Nordhaus (2011), Henderson et al. (2012), Pinkovskiy (2013), Alesina et al. (2016), and Storeygard (2016)). Henderson et al. (2011) furthermore confirms that log differences in NTL is a suitable proxy for growth in GDP. 2 A 10% improvement in market access yields a 1.9% increase in lights emissions. 3 estimates, suggesting the latter are downwards-biased, possibly reflecting measurement error in our market access measure. The positive relationship between market access and lights is also obtained when we estimate a difference-in-differences specification in which we relate 4-year growth rates in lights to changes in market accessibility over the same period. It is furthermore robust to using alternative market access measures that hold the population fixed, such that variation in market access is purely driven by obstacle deployment, as well as to using market access measures that account for access to foreign markets. Results also remain when we use different distance decay parameters. Finally, our results are robust to using bottom-coded instead of top-coded lights data, which capture light emissions from economically more vibrant localities better. Our paper builds on and contributes to different strands of literature. To start with, by focusing on annual (and comparatively moderate) variation in market access, it complements earlier studies on the returns to investing in transportation infrastructure that have exploited large network expansions often materializing over many years, if not decades. Donaldson and Hornbeck (2016), for example, estimate the effect of changes in market access brought about by a major expansion of US railroads on the value of agricultural land. Faber (2014) and Allen and Atkin (2017) estimate the gains from trade that are made possible by unprecedented expansions of the highway network in respectively China and India.3 Our study is particularly closely related to Storeygard (2016), which is the only other paper we are aware of that uses annual variation. He cleverly exploits variation in oil prices (interacted with distances between cities) to identify the impact of transportation costs on economic development in Africa, taking routes as given. We endogenize the choice of optimal routes by adopting a market access approach that builds on the seminal work of Donaldson and Hornbeck (2016). In a robustness analysis we consider access to both domestic and foreign markets (cf. Jedwab and Storeygard, 2017). More generally, our findings resonate with a sizeable body of literature that demonstrates that reducing trade frictions can enhance welfare (Arkolakis et al. (2012), Anderson and Van Wincoop (2004), Head and Mayer (2011)). 3 Changes in transportation infrastructure have been shown to impact the structure of economic ac- tivity more broadly. Baum-Snow (2007), Duranton and Turner (2012), and Gonzalez-Navarro and Quintana-Domeque (2016) study the impact of transportation infrastructure on sub-urbanization, growth of cities, and property values respectively. Allen and Arkolakis (2014) assess how it shapes the spatial organization of the economy, while Adukia et al. (2017) and Asher and Novosad (2017) investigate the impact of rural road expansion on investments in education and the structural trans- formation of the labor market. Blankespoor et al. (2017) estimate the impact of road improvements on local employment and specialization. Jedwab and Moradi (2016) and Jedwab et al. (2017) examine how the collapse of railroad network impacted on road network investments as well as on the spatial distribution of economic activity. 4 Several other studies exploit obstacles to mobility as a source of identifying varia- tion in the West Bank and Gaza. Abrahams (2018a), in a closely related study to ours, focuses on how Israeli obstacles disrupted Palestinian commuting, causing employment rates to decline in labor-supplying locations but to increase in labor-demanding loca- tions. Cal`ı and Miaari (2018) document a negative association between local employ- ment rates and mobility restrictions in the short-run. Amodio and Di Maio (2017) study the impact of trade restrictions between the West Bank and the outside world on the efficiency of input allocation among Palestinian firms. Finally, Etkes and Zimring (2015) use the blockade of Gaza as a natural experiment to quantify the gains from trade by using the West Bank as the counterfactual. The remainder of this paper is organized as follows. The next section elaborates on the context. Section 3 discusses the construction of the accessibility index, which is our key independent variable. Lights and fatalities data are presented in section 4. Section 5 presents our empirical strategy. The results are presented in section 6. A final section concludes. 2 Context The West Bank is an elevated plateau bordered on its north, west, and south by coastal Israel, and to its east by Jordan. It is small: about 56km (34.8 miles) at its widest and about 133km (82.6 miles) at its lengthiest, roughly 1/4 the area of New Jersey. Approximately 2.7 million Palestinians live there, while approximately 330,000 Israeli civilians live in over 160 separate, segregated settlements and outposts throughout the West Bank.4 Since 1967, a major part of the West Bank has been controlled by the Israeli army. The army maintains a constant troop presence of bases, garrisons, patrols, and watch- towers throughout the West Bank’s interior, while all borders of the West Bank (in- cluding its border with Jordan) are controlled by Israeli security fencing, walls, and staffed by Israeli civil administrators and border police at every port of entry and exit. The Israeli army responds to ongoing Israeli-Palestinian conflict by both offensive retaliations (Jaeger and Paserman, 2008) and defensive efforts aimed at intercepting militants before they reach Israel or its settlements. This latter policy intensified under the auspices of ‘Operation Defensive Shield’, and involved the deployment of hundreds of physical obstacles (Cal` ı and Miaari, 2018) inside the West Bank, along the internal road network, and the construction of a 500km wall that separates Israel from the West 4 See Niksic et al. (2014). Since our study does not concern East Jerusalem, we count Palestinians and Israelis inside the West Bank only. 5 Bank.5 These mobility restrictions serve to safeguard Israeli civilians who have settled beyond the wall inside the West Bank: Israel’s primary justification for the movement restrictions is that they are necessary to protect Israelis within its jurisdiction and Israelis living in the West Bank or traveling on West Bank roads. ...the settlement enterprise, including the roads built for it, was one of the primary fac- tors in shaping the restrictions regime that Israel has forced on the Palestinians since the beginning of the Second Intifada. B’Tselem (2007) The ethos behind obstacle employment reduces concerns about endogeneity: mobility restrictions served security objectives and did not target market access nor economic performance. We worry that Israeli security objectives accidentally and indirectly re- lated to Palestinian economic conditions, but we do not worry that they are deliberately or directly related.6 Due to a lack of local GDP data, we use nighttime lights as a proxy for local economic output. In theory, Israel could control lights in the West Bank by regulating the supply of electricity to Palestinian localities. If cutting of electricity were used as a punishment device, then this “unobserved variable” could correlate both with NTL and the deployment of road closure obstacles. There is no documented (or anecdotal) evidence however that Israel is manipulating electricity supply as a penalty device in the West Bank. Israel has on occasion halted the supply of electricity but in all cases this has been for commercial reasons, i.e. in response to overdue electricity bills. Thus, there is no reason to believe that this will bias our results. Another potential cause for concern is that Palestinian firms may have relocated in response to loss of market access caused by obstacle deployment. We expect firm 5 According to UNOCHA, decisions regarding the placement of obstacles in the West Bank are made by the brigade commander of the Israeli Defense Forces assigned to the location or governorate of interest, often in consultation with the Israeli security forces. The decision to introduce or replace an obstacle is motivated by safety concerns, not by economic considerations. 6 One conceivable endogeneity concern, for example, is that in addition to obstacle deployment, the Israeli army performed other security measures such as raids, arrests or curfews, which could be correlated with both obstacle placement and local economic prospects. Though we attempt to assuage these and other endogeneity concerns later by using an instrumental variable approach, an important argument against this concern is that obstacles were essentially a defense security measure, and affected Palestinian market access only as an afterthought. By contrast, raids, arrests, and curfews were all offensive (proactive) security measures taken deliberately against specific Palestinian locations. We do not expect defensive and proactive measures to correlate spatially. For a more in-depth contextual overview we refer the interested reader to Abrahams (2018b). 6 relocation to be limited for several reasons, however. To start with, property rights have been very insecure during this period. Palestinians have tended to rely on their family and patriarchal clan (hammouleh ) to protect their property. Such protection however only applies when one lives near relatives, which helps explain low voluntary migration rates. Indeed, using census data Abrahams (2018a) finds negligible evidence of firm relocation. With regard to market access specifically, lack of relocation has an even more obvious explanation: obstacles were plainly a temporary phenomenon, so it made sense only for the firms fleetest of foot to pay the fixed costs of relocating during a time of such uncertainty and change. In interpreting our results it is important to keep in mind that road obstacles rep- resent only one class of mobility restrictions limiting the economic success of the West Bank. Other restrictions include international trade restrictions, Area C restrictions, and limited control of public utilities and communications infrastructure (for further discussion, see Niksic et al. (2014) or AIX Group (2013)). While mobility restrictions have gradually been eased over time, these other restrictions have remained, so our estimates of the benefits of market access may be unduly conservative. 3 Measuring market access in the West Bank 3.1 Defining market access Market access as we will define it evaluates the ability of roads to connect people and firms. Being better connected to populated localities means being better connected to potential consumers as well as workers. The productivity of a locality is enhanced by cheaper access to markets and labor. Donaldson and Hornbeck (2016) derive a measure of market access that emerges as the solution to a set of implicit equations where market access for a given locality equals a weighted sum over other localities’ access to markets, with the weights being a function of the cost of interacting with these other localities and their population counts (see eq. (9) in Donaldson and Hornbeck (2016)). Donaldson and Hornbeck (2016) also derive a first-order approximation to this solution which yields a measure of market access for locality i that is of the following form:7 M Ait = h(Tijt )Pjt , (1) j 7 Donaldson and Hornbeck (2016) confirm that the approximation is highly correlated with the exact measure of market access, and that their empirical results are not sensitive to switching between the approximate and exact measure. 7 where Pjt denotes destination locality j ’s population count in year t, Tijt is the minimum travel time between localities i and j in year t, and h(T ) denotes the distance decay function (i.e. a positive and monotonically declining function of T ).8 For the distance 1 2 decay function h(T ) we will use: h(T ) = exp (− 2 T /θ2 ), where the parameter θ governs the rate at which markets located further away are being discounted.9 From the perspective of a producer at some origin location, not all possible desti- nations are equally important: their size and distance help determine their relevance. Ceteris paribus, destinations that take more time to reach are less relevant, since the price of the exported good will be higher there (assuming transport costs increase with distance), and therefore less competitive with local substitutes. Destinations with smaller populations are less relevant, since they contain fewer consumers (and workers). Moreover, market access at a given origin i may respond to changes in road connec- tivity well beyond the direct neighborhood of locality i; if connectivity between i and j changes, it will obviously affect market access for localities i and j , but it will also have an effect on market access of localities connected to i and j , and the localities connected to these localities etc. We compute market access for a balanced panel of grouped (aggregated) localities in the West Bank at an annual frequency covering the period 2005-2012 (the source of time-variation will be discussed in the next subsection). For computational convenience, the set of destinations is restricted to the governorate capitals, excluding Jerusalem (which is located at the other side of the barrier) and using nearby Al Ram instead. In a robustness check in which we attempt to account for access to foreign markets we add Tel Aviv as an additional destination (see section 5.2). The governorate capitals are excluded as origin localities for the same reason Donaldson and Hornbeck (2016) exclude origin locations from their set of destinations, notably concerns of endogeneity bias. A household’s decision to reside in a given locality is plausibly in part determined by the economic prospects for that location, such that the population count will be co- determined with economic success. Excluding a locality’s own population count should alleviate this concern to an important extent.10 We will present additional robustness 8 This measure resonates with the accessibility measures put forward in the regional sciences and transportation literature that dates back to the 1950s, see e.g. Harris (1954), Hansen (1959), Ingram (1971), Wachs and Kumagai (1973), Dalvi and Martin (1976), Black and Conroy (1977), Koenig (1980), Guy (1983), Heikkila and Peiser (1992), Allen et al. (1993), Geertman and Van Eck (1995), Song (1996), Deichmann (1997), Kwan (1998), and Geurs and Van Wee (2004). 9 There are a handful of commonly used choices for the distance decay function, of which Gaussian decay is one, see e.g. Ingram (1971), Guy (1983), Song (1996), Deichmann (1997), Kwan (1998). h(T ) = T −θ is another popular choice, which is adopted by Donaldson and Hornbeck (2016). Results not reported here to conserve space but available upon request confirm that using the latter choice of distance decay does not alter our findings. Our results are also qualitatively robust to the choice of distance decay parameter, as is shown in sections 5.2 and 6.2. 10 Our results are found to be robust to the decision to include or exclude the governorate capitals 8 Figure 1: The road network and governorate capitals in the West Bank checks that address this concern in section 6.2. Figure 1 shows the geography of the West Bank along with key data inputs that will feature in the computation of market access: the locations of disaggregated localities in the West Bank (green dots), including the 11 governorate capitals (red squares with name labels), and the road network that connects them.11 Locality locations and population counts The Palestinian Central Bureau of Statistics (PCBS) conducted a population census in 1997 and 2007. In addition to collecting population statistics, the PCBS also records the geographic coordinates of the localities (centroids in 1997 and “urban footprint” stored as polygons in 2007), which they kindly shared with us. We count 687 localities in 1997 and 545 localities in 2007; a number of localities were merged or split between the two census periods. In order to build a balanced panel of localities, we created a themselves from our regression analysis. 11 The 11 destination cities include: Jenin, Tubas, Tulkarm, Nablus, Qalqiliya, Salfit, Ramallah, Jericho, Ar Ram, Bethlehem, and Hebron. 9 new locality identifier that equals the 1997 identifier in case of splits and equals the 2007 identifier in case of a merge (i.e. the most dis-aggregated identifier that can be tracked over time). This new identifier counts 533 unique “stable” localities that can be tracked over time. A number of localities are dropped from the analysis, includ- ing Jersualem localities that lie outside of the West Bank and localities with a 2007 population count below 1000. This reduces the number of stable localities that will be used in the regression analysis to 241. Annual population counts Pjt are obtained by interpolating (and extrapolating) the log of the 1997 and 2007 counts for each locality separately. For the geographical coordinates of the localities we work with centroids, which were “snapped” to our road network. Road network data The geo-referenced road network data were kindly provided to us by UNOCHA (who further developed the road network data produced by the European Commission’s Joint Research Center (JRC)). UNOCHA’s road network data counts approximately 26,000 geographic features for the West Bank (compared to approximately 20,000 geographic features for JRC’s data). Each road segment is classified into one of 6 road types with corresponding estimates of the average amount of time required for a typical civil- ian/commercial vehicle to traverse the segment: Regional (60kph), Primary (50kph), Secondary (40kph), Tertiary (30kph), Residential (20kph) and Track (10kph). The data also record different categories of restrictiveness for each road segment concerning use by Palestinians: No Restriction, Restricted Use, Partially Prohibited, and Totally Prohibited. We drop road segments classified as “Totally Prohibited”, and assume that the necessary permits have been acquired to permit use of all remaining roads. The analysis requires road segments to be perfectly connected (and origin and destination localities to be perfectly connected to the road network) for the the network analyst algorithm to traverse the road network. Road segments that did not exactly connect were modified to connect based on our best judgment. Travel times The computation of market access as defined in eq. (1) requires data on travel times Tijt . We used ArcGIS 10.3’s Network Analyst software package to solve for the optimal route that minimizes travel time between all origin and destination pairs of interest. Selected examples of these optimal routes (depicted in bold purple lines) are plotted in Figure 2. 10 Figure 2: Optimal routes between Ramallah and selected cities in the absence of ob- stacles 11 3.2 Road closure obstacles as a determinant of market access Road closure obstacles The Israeli army deployed physical obstacles along roads and borders to defend Israelis against violence originating in the West Bank. A 500 km wall or separation barrier was built along (and sometimes beyond) the Armistice Line of the West Bank to prevent the entry of threats from the Palestinian side into Israel. Henceforth, all cross-border traffic was forced to pass through any one of nine green-line checkpoint border crossings, guarded by the Israeli army (see B’Tselem (2007)). Throughout the uprising, however, nearly a quarter million Israeli civilian settlers were dwelling deep inside the West Bank, beyond the protection of the wall. To defend those settlements, the army deployed hun- dreds of manned checkpoints, roadblocks, boulders, earthmounds, and other physical obstacles along the West Bank’s internal road network with the intention of monitor- ing Palestinian traffic passing along roads in the vicinity of Israeli settlements. Most of these obstacles constitute a “full stop”, meaning that vehicles cannot pass though them, prompting a detour for traffic making their way from origin to destination. Checkpoints denote an exception. While these do not constitute a “full stop”, passing through a checkpoint consumes a certain amount of time due to the screening of passengers and vehicles by the Israeli army and potential traffic congestion. The UNOCHA Map Center did an excellent job of recording the progress of the wall’s construction, and keeping track of road obstacles’ geolocations, frequently updating their maps throughout and after the uprising as obstacles were moved and removed. Over this period, UNOCHA published Closure Update reports, which include detailed maps (in PDF format) showing the locations of obstacles that were deployed at the given points in time. We obtained a copy of their database in ArcGIS shapefile format which contained: (a) the geolocation of the separation barrier in its completed form, and for each road closure obstacle, (b) start-date, (c) end-date (if no longer active), and (d) type of obstacle. The left panel of Figure 3 displays a map constructed from our data depicting a snapshot of the different types of obstacles along the major arteries of the West Bank’s internal road network as of January 2006. For reference, the map also includes the Palestinian Authority’s 11 West Bank governorate capitals. The right panel zooms in on a northwestern section of the West Bank, depicting locations of Israeli army ob- stacles lying along roads connecting the governorate capitals of Tulkarm, Nablus, and Ramallah. Various types of obstacles are visible, including checkpoints, less vigorously enforced checkpoints dubbed partial checkpoints, roadgates, and unmanned obstacles such as earthmounds. The fact that obstacles were deployed on some roads and not 12 Figure 3: West Bank roads with road closure obstacles (January, 2006) 13 others provides cross-sectional (spatial) variation in the degree to which each Pales- tinian location was obstructed by obstacles. Completing the road closure obstacles database The original database we received did not include: (a) construction dates for each seg- ment of the separation barrier (needed to account for changes in optimal routes as segments are added to the wall over time), (b) obstacles prior to 2008, and (c) esti- mates of the “crossing times” associated with crossing any given checkpoint (which vary over time). We completed (a) and (b) by inspecting the obstacle maps UNOCHA has published over the years, and by manually: (1) dating each segment of the separation barrier, and (2) adding the geolocations and attributes of any obstacles appearing on published UNOCHA maps prior to 2008 into our shapefile. Moreover, obstacles whose locations did not perfectly align with the road network were snapped to the most plausi- ble road segment based on available documentation on these obstacles and consultation with UN-OCHA officials. Since checkpoints play a critical role in controlling Palestinian traffic in the West Bank, accounting for the time it takes to traverse active checkpoints improves our es- timates of optimal travel times over the road network. Working in coordination with UNOCHA, we conducted repeated retrospective interviews between 2009 and 2013 of their field workers in an effort to estimate average “crossing times” for each checkpoint and for each year between 2005 and 2012. For selected checkpoints the reported cross- ing times are found to range between 15 minutes and 240 minutes depending on the time period, highlighting the importance of accounting for fluctuations in checkpoint traversal times. Updating optimal routes as obstacles are relocated or removed Between 2005 and 2012 obstacles are frequently relocated, replaced by other types of obstacles, or removed altogether. As the spatial configuration and intensity of obstacles changes, we recalculate the optimal routes between origin-destination pairs (by mini- mizing travel times). Figure 4 provides an example, plotting the optimal route between Tulkarm and Nablus as it evolves over time. Nablus is Palestine’s historical economic capital and the West Bank’s second largest city after Hebron. It was surrounded by a large number of obstacles in the years following the uprising that made it difficult for commercial (and civilian) traffic to travel to and from Nablus; all traffic was forced to travel through the Huwwara and Awarta checkpoints, located to the south of the city, significantly prolonging travel times. As some of these obstacles were being removed during the 2009-2010 period, more direct routes to/from Nablus were being permitted 14 as can be observed in the bottom panel of Figure 4.12 Optimal routes also evolved with the construction of the separation barrier wall. Construction of the barrier wall started during the Second Intifada and it continued throughout the period of analysis.13 Figure 5 provides two examples of optimal route changes between 2006 and 2012: Qalqiliya to Hebron (left panel map) and Qalqiliya to Ramallah (right panel maps). Temporal variation in travel times Changes in optimal routes generally yield changes in travel times. But travel times may also change when optimal routes have not, namely when the time it takes to traverse the checkpoints located along the optimal routes have changed (without in fact altering the optimal routes). In Figure 6 we plot the kernel density of optimal travel times between all origin-destination pairs for two points in time (2006 and 2010). The third density, depicted in bars, corresponds to a hypothetical world where all road closure obstacles have been removed. Our calculations demonstrate that obstacles have greatly increased the travel times between origins and destinations in the West Bank, and that these travel times are subject to sizeable fluctuations over time as obstacles are intermittently reconfigured. For example, traveling from Hebron in the southern West Bank, we calculate that an obstacle-free journey to Nablus would have taken 2 hours and 17 minutes. In 2005-2008, however, the same journey would have taken around 7 hours. In 2009 travel times declined to “just” 4 hours and 40 minutes, and to less than 4 hours in 2010. 3.3 Market access over time and across space Estimates of travel times Tijt are combined with population count data Pjt to obtain estimates of market access M Ait , following eq. (1), for all localities i and times t. The distance decay parameter for our main specification is set at θ = 40. Other values for θ will be considered as part of the sensitivity analysis (see Sections 5.2 and 6.2). We compute market access on an annual basis. This matches the frequency with 12 The figures show that in 2006 and 2007 the obstacles in place forced traffic to take a longer route compared to later years when these obstacles were removed. In 2006-2007, commercial traffic traveled via the Huwwara or Awarta commercial checkpoints south of Nablus. In 2009, the Beit Iba checkpoint was dismantled which opened up the North-West route out of Nablus towards Tulkarm. This eliminated the need to travel via the Jit checkpoint. 13 As of July 2011, approximately 437 kilometers of the barrier wall was completed, 58 kilometers was under construction, and 213 kilometers was planned and not yet under construction (UNOCHA, 2011). As a result, 71 of 150 settlements are on the Israeli side of the separation barrier. Upon completion of the planned separation barrier, part of the total barrier (15%) is near the 1949 Armistice Line (Green Line) or within Israel, while the remainder of the total barrier (85%) is within the West Bank. 15 Figure 4: Optimal routes between Tulkarm and Nablus over time 16 Figure 5: Impact of barrier wall on optimal routes Optimal routes between Hebron and Qalqiliya (left panel) and between Ramallah and Qalqiliya (right panels) before (2006) and after (2012) construction of the barrier wall north of Salfit 17 Figure 6: Kernel density plots of travel times (in minutes) which the lights data (our outcome variable which will be introduced in the next section) are observed. Note that our obstacles data are not updated on an annual schedule. For the years after 2008, our database includes exact dates on which obstacles were being deployed or removed. Prior to 2008, the precision with which we can track obstacles over time is determined by the frequency with which UNOCHA updated its Closure Update reports and maps, which varies with the intensity with which obstacles were being reconfigured (the time between updates ranges roughly between 3 and 12 months). First, we compute market access for the irregularly spaced time periods dictated by data availability (every month prior to 2008, and the times of the UNOCHA updates after that). Second, we collapse these data to obtain annual estimates of market access by simply averaging the intra-year data. Figure 7 shows the landscape of market access in the West Bank, and how it has evolved from 2006 to 2010. Darker colors are associated with higher values of market access. It can be seen that localities surrounding Nablus and Ramallah experienced a marked increase in market access over this period. Figure 8 presents the same data but in the form of kernel density plots. The recon- figuration of obstacles has reshaped the distribution of market access over the years. The evolution from a bi-modal to a uni-model distribution reflects a number of signif- icant changes between 2006 and 2010 that have improved market access (transferring density at lower levels of market access to higher levels of market access). Lifting key 18 Figure 7: Market access (θ = 40): 2006 (left) versus 2010 (right) 19 Figure 8: Kernel density of market access (with θ = 40) 1 .8 .6 density .4 .2 0 9 10 11 12 ln market access 2007 2008 2009 2010 restrictions around Nablus in late 2009 arguably denotes the most prominent change. Not only because Nablus was surrounded by some of the most restrictive obstacles, effectively amounting to a blockade, but also because the city is central to the West Bank, both geographically and economically.14 Another notable change is the gradual improvement in connectivity between the north and the south of the West Bank (from late 2007 to 2010) resulting from a reduction in Wadinar checkpoint traversal time (which controls North-South traffic). As some of the restrictions to north-south traf- fic were eased, more centrally located localities were able to re-connect to the south’s large market/population. The improvements in market access brought about by the above mentioned changes are also apparent in Figure 9 which shows a time-series of log market access by means of a box plot. The width of the boxes show the cross-sectional variation in (log) market access (across locations). 14 It is the second largest city, home to the Palestinian Stock Exchange (which is small by international standards), and enjoys close geographic proximity (in the absence of obstacles) to all but the two most southern governorate capitals (which are Hebron and Bethlehem). 20 Figure 9: Box plot time-series of market access (with θ = 40) 4 Nighttime Lights and local conflict data 4.1 Nighttime Lights as a proxy for economic output This section introduces the nighttime lights (NTL) data. Since the 1970s the Defense Meteorological Satellite Program (DMSP) has been recording the intensities of night- time light emissions from the Earth’s surface. For all satellites and all years since 1992, the National Oceanic and Atmospheric Administration (NOAA) has generated freely accessible annual composite images that average the intensities of light recorded on cloud-free nights at each pixel. Apart from their obvious usefulness for tracking electri- fication rates (Min, 2015), year-to-year changes in lights have also been shown to track year-to-year GDP changes (Henderson et al., 2012). This has popularized lights as a proxy for GDP in under-developed or conflict-prone contexts where spatially disaggre- gated GDP data are often unavailable or unreliable (see, for example, Alesina et al. (2016) or Pinkovskiy (2013)). The left-hand-side map of Figure 10 shows an annual composite image of the West Bank recorded by the DMSP’s F16 satellite in the year 2006. We can see brighter and larger concentrations of light around the major West Bank cities of Nablus, Ramallah, and Hebron, among others. To decide whether or not to assign a pixel’s light intensity (digital number or DN) to a Palestinian location, we obtain urban footprint polygon data for all Palestinian census locations through the UNOCHA Map Center. Poly- 21 Figure 10: Top-coded vs. bottom-coded lights F16-2006 (top-coded) Lights image (left) and RADCAL (bottom-coded) Lights image (right) 22 gons were finely drawn to circumscribe built-up areas and are highly accurate for the time period of interest. Throughout our analysis we assign a pixel’s value to a Pales- tinian location if the pixel’s centroid lies within 500 meters of the location’s polygon. This “generous” assignment rule allows for potential geolocation errors in both ArcGIS’ under-the-hood projection of NTL data from WGS-1984 coordinates to UTM-36N, and for shifting error in the NTL data themselves (Tuttle et al., 2013).15 Summing up the DNs of all assigned pixels per location, we find that the brightest three towns in 2006 were Hebron (3577 counts), Nablus (2309), and Ramallah (1830). Potential truncation of DMSP-NTL imagery can arise because of topcoding. Based on 1970s technology, DMSP satellites were only equipped to record up to 6 bits of information per pixel, rendering DNs of 0 to 63. Moreover, the gain setting was typically quite high in order to detect faint light sources. When the satellite viewed bright urban sources, it often topcoded (maxed out at 63). The overwhelming majority of West Bank Palestinian locations were too dim to suffer from this problem, but few neighborhoods in the greater Jerusalem area (Bethlehem, Ramallah) are topcoded for most years. As a result, we cannot tell for most of these locations if they grew brighter or dimmer; they report 63 for most or all of their pixels. As described below, we perform robustness checks to guard against the influence of these topcoded locations on our results. Most importantly, parallel to the topcoded DMSP-NTL series, NOAA also generated a ‘radiance-calibrated’ series for select years, including 2006 and 2010. The radiance-calibrated (RADCAL) series was constructed from a lower average gain setting, avoiding topcoding at the cost of introducing bottom-coding. We show that our main results survive when we use these non-topcoded data. The right-hand-side map of Figure 10 shows the RADCAL image captured by the F16 satellite for 2006, which illustrates that the bottom-coded setting does not pick up light emissions from dimly lit areas. Another potential concern is that OLS sensors on DMSP satellites deteriorate over years in orbit, and that their average gain settings vary from year to year and satellite to satellite. As a result, the images of the West Bank in any two given years are not directly comparable. To facilitate interpretation of our graphs, however, we applied intercalibration parameters to make the data as comparable over time as possible. For topcoded imagery, we applied the parameters recommended in Wu et al. (2013). For non-topcoded imagery, we applied the parameters recommended in Hsu et al. (2015). In our regressions, we rely on year effects and locality fixed effects, as well as the use 15 In a robustness check that we do not present to conserve space but is available upon request we experimented with using a more conservative assignment rule, where a pixel’s value is assigned to a location only if its centroid lies inside the polygon. The results obtained using this alternative assignment rule are very similar to the ones presented in the main text. 23 Figure 11: GDP of the West Bank & Gaza and Nighttime Lights (NTL), 2000-2010 2.4 2.2 ln(output) 2 1.8 2000 2002 2004 2006 2008 2010 Year ln(GDP) ln(NTL) of natural-logs, to minimize this source of error. DMSP-NTL appear to track West Bank Palestinian economic growth accurately for the period under study. To construct a NTL series for the West Bank as a whole, we sum the intercalibrated pixel values (i.e. the quantities of light observed) within 500m of West Bank Palestinian polygons. After taking natural logs, we plot the resulting time series for 2000-2010 in Figure 11, along with series of logged Palestinian GDP. The GDP data were obtained from the World Bank, which unfortunately does not disaggregate series separately for the West Bank and Gaza. The NTL series tracks the time-evolution of national GDP well. Note for example the marked decline in light before and during the early years of the Second Intifada, which commenced in late 2000. Note also the slump in both series from 2005 to 2007 resulting from the introduction of sanctions in response to the election of a majority-Hamas legislature. The subsequent recovery coincided with improvements in market access. Joint evolution of Nighttime Lights and Market Access Figures 12 and 13 provide a preview of our regression results. Figure 12 plots the residual of the natural log of lights per capita against the residual of the natural log of market access, controlling for locality and year fixed effects, as well as proxies for local conflict, which will be discussed below. The positive slope suggests that higher market access is associated with more output (proxied by lights). Similarly, Figure 13 plots 24 the difference in log lights per capita over 2006-2010 against the difference in log lights, controlling for local conflict. The slope is again positive, indicating that improvements in market access are associated with output growth. In Sections 5 and 6 we make an effort to establish causality. Figure 12: Nighttime Lights (NTL) vs. market access .4 .2 ln NTL pc (Residual) -.2 0 -.4 -1 -.5 0 .5 1 ln Market Access Θ=40 (Residual) coef = .061, se = .014 t = 4.23 Nighttime Lights (NTL) vs. market access (controlling for locality, year and fatalities) 4.2 Fatalities data We use geo-referenced data on Israeli and Palestinian fatalities in the West Bank from B’Tselem16 as our indicators of local conflict. For each violent incident, B’Tselem records the exact date, geolocation, and number of fatalities for Israelis and Palestinians separately. We combine these data with our geo-referenced locality data for the West Bank, to obtain the number of Israeli and Palestinian fatalities that have been recorded within a 5km radius of each locality during the course of the year, which we use as a control variable. 16 B’Tselem is an Israeli independent non-profit organization that records a variety of statistics related to the conflict between Israel and Palestine. It refers to itself as an Information Center for Human Rights in the Occupied Territories. The statistics published by B’Tselem we use can be downloaded from: http://www.btselem.org/statistics. 25 Figure 13: Change in Nighttime Lights (NTL) vs. change in market access 1 ∆ ln NTL pc (Residual) 0 -.5 .5 -.5 0 .5 1 ∆ ln Market Access (Θ=40) (Residual) coef = .124, se = .044 t = 2.82 Change in Nighttime Lights (NTL) vs. change in market access (controlling for fatalities), 2006-2010 5 Empirical strategy Taking Palestinian locations as our units of analysis, we attempt to identify the causal effect of market accessibility on annual local nighttime light output per-capita. Our main specification is: ln yit = β ln M Ait + γxit + λi + δt + εit , (2) where the outcome variable yit is the total lights per-capita observed at location i in year t, M Ait is our measure of market access from eq. (1) with θ = 40 (different values for θ are considered in the sensitivity analysis), xit is a vector of time-varying control variables for location i, λi denote location fixed effects, and δt denote year fixed effects. Our controls xit are the number of Palestinian and Israeli fatalities that occurred in year t within a 5km radius. β is the (short-run) elasticity of lights with respect to market access, the key parameter of interest. We drop locations with fewer than 1000 residents (in 2007), so as to ensure that results are not driven by small villages or temporary bedouin encampments. Finally, we only keep localities for which we have data for all 8 years. The resulting regression sample is a balanced panel of 241 locations for the years 2005 to 2012. Standard errors are clustered by locality and are robust to heteroskedasticity. 26 The year effects control for West Bank level time-varying conditions, including shifts in public policy, terms of trade, international aid flows etc. They also help us account for the deterioration and changes of the satellites over the years (see Section 4.1). Location fixed effects capture the time-invariant features of location i that affect annual light output, such as initial size of the location, local geographic conditions etc. Our motivation for controlling for local violence, by including xit , is that it could conceivably be a driver of both accessibility and changes in lights. 5.1 Instrumental variables estimation OLS estimation of β from eq. (2) may be vulnerable to endogeneity bias if other factors not accounted for are correlated with changes in both local obstacle deployment and local lights. The sign of such bias is theoretically ambiguous. We can think of at least three different scenarios that would predict biases of different signs. A positive bias (i.e. OLS over-estimating β ) might emerge when the omitted variable denotes complementary penalties imposed on Palestinian localities. These would be negatively correlated with market access and negatively correlated with local economic success, hence predicting a positive bias. For a candidate explanation for a negative bias consider the possibility that our measure of market access is subject to measurement error which would introduce clas- sical attenuation bias. Alternatively, consider a scenario where fixed checkpoints are being replaced by “flying” checkpoints, which are temporary and mobile checkpoints that can pop-up anywhere and anytime depending on current circumstances. They can be in place for a couple of hours or a couple of days, but rarely much longer. Because of their fleeting nature it is harder to collect reliable data on them, such that they are currently not captured in our database. This “omitted variable” would be positively correlated with our market access variable (as fixed checkpoints are replaced by flying checkpoints, our measure of market access and the flying checkpoint count go up to- gether) but negatively correlated with local economic success, which would predict a negative bias. We address this concern by building an instrument for market access that omits obstacles located within 10 kilometers from a given location. Note that given the very compact size of the West Bank discussed in section 2, 10km is a considerable distance (recall that the West Bank is 56km at its widest and 133 km at its lengthiest). The idea is that this instrument is orthogonal to local “incidents”, thereby uncorrelated with the above mentioned omitted variables, but nevertheless correlated with market access since changes in the deployment of obstacles located further away from a given 27 locality will have an impact on how well the locality is connected to the governorate capitals. One option would be to recompute our measure of market access for each location, where any obstacles within 10 kilometers are ignored, and then use this as an instrument. However, this is highly computationally intensive. As a computationally more convenient alternative we compute the number of checkpoints located farther away than 10km but no farther than 25km, and will refer to this as the checkpoint “donut count”. This instrument is also orthogonal to possible measurement error to the extent that this error is driven by guesstimation errors in “crossing times” (associated with crossing the checkpoints). This may serve as an additional argument for using the checkpoint donut count rather than the alternative market access variable as an instrument. 5.2 Addressing additional concerns This section elaborates on a variety of robustness checks addressing additional concerns. Holding the population constant Changes in our measure of market access over time are not only due to changes in the configuration of road closure obstacles but also due to changes in the population size of the destinations (i.e. changes in size of the respective markets). Population growth however is arguably co-determined with economic prospects (Baum-Snow et al. (2017)). Hence, one concern is that growth in lights may in part be driven by unobserved local shocks that also impact on population growth in nearby governorate capitals. For this reason, population at origin is not counted in our market access measure (by dropping the governorate capitals from the sample). To further address this concern, we will re- compute each location’s market access by holding the destination population constant. Varying distance decay Market access is also a function of the distance decay parameter θ (see eq. (1)). For our main specification we set this parameter equal to 40 (in practice this implies that destinations that take about 90 minutes or more to travel to will get relatively little weight). While one could in theory estimate θ, it is a difficult parameter to estimate precisely (see e.g. the discussion in Donaldson and Hornbeck, 2016). As a robustness check we will consider different values of θ ranging between 30 and 80. It should be noted that the level (variance) of market access will mechanically increase (decrease) as we increase θ. Therefore, one way of evaluating the impact of the choice of θ is to compare estimates of light emission in a West Bank with and without closure obstacles 28 (the latter denoting a hypothetical West Bank), where in both cases market access is computed using the same choice of θ. The results of this back-of-the-envelope exercise are presented in Section 6.3. Accounting for access to international markets Economic performance depends both on domestic and foreign market access. To account for potential impacts of changes in road connectivity on external market access we compute an augmented measure of market access in which we proxy external market access by adding Tel Aviv as a 12th destination (in addition to the 11 governorate capitals), considering a range of different population sizes for the external market.17 We compute the shortest travel time to Tel Aviv allowing travel via three different border crossings, notably Al Jalama, At Tayba, and Tarqumiya.18 Top-coded versus bottom-coded NTL (+ dif-in-dif estimation) We use two different variations of original NTL data: top-coded and bottom-coded lights data (see Section 4.1). The latter will also be referred to as radiance-calibrated lights data, or RADCAL in short. These two variatons are obtained by employing different settings on the same satellite, akin to adopting different settings concerning light-sensitivity on a photo camera. When the top-coded setting is adopted, the satellite is sensitive to dim sources of light, thus providing accurate readings of the brightness of these relatively dim sources. However, it is unable to distinguish between lights if their brightness exceeds a certain level of intensity. The opposite holds true for bottom- coded (RADCAL) data; it accurately measures bright light but is unable to distinguish between relatively dim sources. This means that the bottom-coded series is better equiped to track lights for large and more developed cities, while the top-coded series may be preffered when tracking smaller less developed locations. Top-coded data are recorded annually while bottom-coded data are only available 17 In the default measure we assign a population of 10 million to Tel Aviv as it serves to represent all external markets. This number is admittedly arbitrary but serves as a crude approximation to the relative size of the Palestinian external market. Israel, which had approximately 8 million inhabitants in 2012, is the main trading partner of the West Bank and Gaza; Palestinian sales to Israel account for roughly four-fifths of all Palestinian exports while purchases from Israel account for about two thirds of all imports. In robustness checks presented in the Appendix we assess the robustness to increasing the size of the external market to respectively 25 and 50 million people. 18 Our approach to accounting for access to external market is inevitably somewhat crude. We do not take into consideration potential time penalties associated with crossing the border nor do we quantify other transactions costs associated with crossing borders, including tariffs, obtaining travel permits etc. We also do not account for border closure days. Accounting for all of these additional factors would presumably only reduce our external market access measures. Put differently, our external market access proxies might be conservative in the sense that by not accounting for these additional frictions we allow for a maximal impact of external market access. 29 for 2006 and 2010 (for our period of study). For this reason we use the top-coded data for our main specification. To verify how sensitive our findings are with respect to this choice, we will present regression results obtained using bottom-coded data as a robust- ness check. As we only have two years of bottom-coded data, which are 5 years apart, we will estimate the parameters from eq. (2) using a difference-in-differences approach. For completeness, the same difference-in-differences regressions are also applied to the top-coded data. Examining how the relationship between changes in market access and changes in lights holds up over this longer term constitutes a valuable robsutness check in and of itself. 6 Results To set the stage, columns 1 and 2 of Table 2 present first stage regressions using the natural log of market access as dependent variable and the checkpoint donut count, the number of checkpoints located in a radius between 10 and 25km from the localitys cen- ter, as the main explanatory variable. Although obstacles captured by this donut count measure are not deployed within the locality but rather in its vicinity, they clearly sig- nificantly adversely impact local market access; each additional obstacle reduces market access by approximately 3.81 percentage points according to the estimates presented in column 1. This estimate is not sensitive to controlling for local conflict as is demon- strated in column 2, suggesting that checkpoint deployment outside ones own locality is largely orthogonal to local conflict. Column 3 provides corroborative evidence for this conclusion by demonstrating that our proxies for local economic conflict are not significant predictors of the donut count measure, which now serves as the dependent variable. In summary, the number of obstacles deployed within a 10 to 25 km radius of the locality is a suitable instrument for market access. 6.1 Main findings Turning to the main specifications, Table 2 presents estimates of equation 2. Lights emissions are clearly strongly and significantly correlated with market access. Simple OLS estimates presented in column 1 imply that a 10 percentage points increase in market access increases lights by 0.6 percentage points. If we assume that the GDP- to-lights elasticity is 0.3 (cf. Henderson et al., 2012), this corresponds to an increase in local output of approximately 0.2 percentage points. This relationship is robust to including proxies for violence, as is done in column 2. When we instrument market access using our donut count measure of obstacles 30 Table 1: First Stage Regressions (1) (2) (3) ln Market Access ln Market Access Checkpoints 10-25km OLS OLS OLS Checkpoints 10-25km -0.0381*** -0.0380*** (-8.59) (-8.60) ln(1+Pal.Fat.<5km) -0.0175 0.0639 (-1.11) (0.56) ln(1+Isr.Fat.< 5km) 0.0006 -0.0132 (0.02) (-0.04) Year FE Yes Yes Yes Locality FE Yes Yes Yes N 1928 1928 1928 adj. R2 0.671 0.671 0.478 Notes: * p < 0.10, ** p < 0.05, *** p < 0.01. Heteroscedasticity robust t -statistics clustered by locality in parentheses, Pal.Fat.<5km (Isr.Fat.<5km) denotes the number of Palestinian (Israeli) fatalities within a 5km radius in a given cal- endar year (Source: B’Tselem). Checkpoints 10-25km is the number of checkpoints deployed in a radius between 10 and 25km away from the locality’s center (Source: UNOCHA). Market Access is calculated using: M Ait = j h(Tijt )Pjt , where P denotes destination population, T denotes travel time, and h(T ) = exp (− 1 2 T 2 /θ2 ) is a distance decay function with θ = 40 (see Section 3 for details). located in a radius between 10 and 25km of the locality (columns 3 and 4), the asso- ciation between lights and market access strengthens substantially. According to our preferred IV estimates that control for local violence, presented in column 4, a 10 per- centage points increase in market access leads to a 1.9 percentage point increase in lights emissions (or a 0.6 percentage point increase in local GDP). OLS estimates of the relationship between market access and lights are thus downward biased (a finding which is consistent with existing studies, see e.g. Jedwab and Storeygard (2017), Red- ding and Turner (2015)). We speculate that this may reflect attenuation bias due to measurement error in our market access measure. Whatever the reason may be, the IV regressions lend credence to a causal interpretation of the positive association between market access and economic performance. 31 Table 2: Main Regressions (1) (2) (3) (4) ln NTL pc ln NTL pc ln NTL pc ln NTL pc OLS OLS IV IV ln Market Access 0.0610** 0.0624** 0.188*** 0.190*** (2.41) (2.50) (2.76) (2.81) ln(1+Pal.Fat.<5km) 0.0137 0.0162* (1.59) (1.84) ln(1+Isr.Fat.<5km) 0.0182 0.0181 (1.05) (1.06) Year FE Yes Yes Yes Yes Locality FE Yes Yes Yes Yes N 1928 1928 1928 1928 adj. R2 0.452 0.453 0.344 0.346 Notes: * p < 0.10, ** p < 0.05, *** p < 0.01. Heteroscedasticity robust t -statistics clustered by locality in parentheses. NTL pc measures local (top-coded) Lights emissions per capita (Source: NOAA) as described in section 4. Market Access is calculated using: M Ait = j h(Tijt )Pjt , where P denotes 1 2 2 destination population, T denotes travel time, and h(T ) = exp (− 2 T /θ ) is a distance decay function with θ = 40 (see Section 3 for details). Pal.Fat.<5km (Isr.Fat.<5km) denotes the number of Palestinian (Israeli) fatalities within a 5km radius in a given calendar year (Source: B’Tselem). In columns 3 and 4 Market Access is instrumented using Checkpoints 10-25km, the number of checkpoints deployed in a radius between 10 and 25km away from the locality’s center (Source: UNOCHA). 6.2 Robustness checks Robustness checks for our preferred IV estimates that control for local violence (column 4 of Table 2) are presented in Table 3. Columns 1-3 demonstrate how our estimates change when alternative distance decay parameters are used, notably θ = 30, θ = 60, and θ = 80, respectively.19 While the choice of the distance decay parameter impacts the magnitude of the coefficients, the qualitative pattern of results remains the same; market access is significantly positively correlated with lights, irrespective of the choice of the distance decay parameter. Incidentally, note that the adjusted R2 is highest when choosing θ = 40, which is our benchmark model.20 Second, one may be concerned that the association between market access and lights 19 Note that a higher distance decay parameter θ implies that relatively remote locations are given more weight in the market access measure. 20 Note that as θ increases the regression coefficient associated with log market access mechanically increases as the variation in log market access shrinks. As a consequence, the impact of a one standard deviation in log market access on log NTL per capita is arguably less sensitive to the choice of distance decay parameter. Multiplying the regression coefficients reported in Table 3 obtained for different distance decay parameters with one standard deviation of the corresponding log market access measure yields estimates that range between 0.13 and 0.17. A similar degree of stability of this effect with respect to varying the distance decay parameter is obtained by Jedwab and Storeygard (2017), who examine the effect of changes in market access on local population growth. 32 reflects population movements (which are potentially endogenous to economic prospects as alluded to in section 5.2). To address this concern, column 4 estimates a specification in which we hold population fixed in our market access measure such that all variation in market access results from changes in the configuration and severity of obstacles. The resulting coefficient estimate is nearly identical to the one in our preferred specification. Column 5 repeats this exercise but now using θ = 80; again, the results are very similar to those reported in column 3. Variation in market access is thus primarily driven by changes in obstacles rather than migration and/or differential population growth. Third, column 6 presents estimates for an augmented measure of market access in which access to external markets is proxied for by including Tel Aviv as an additional destination. We hold the population fixed (so that we don’t have to take a stance on the evolution of the size of the external market over time) and use a distance decay parameter of 80 such that our measure of market access is especially sensitive to changes in the accessibility of foreign markets. The resulting coefficient estimate, however, is very similar to the one presented in column 5 which only considers domestic markets. This suggests that our findings are robust to controlling for changes in external market access. Inflating the size of the external market does not alter this conclusion, as is shown in Appendix Table 6. Table 4 presents estimates of the change in the log of lights emissions between 2006 and 2010 on the change in market access between 2006 and 2010 as well as the change in our proxies for political violence, separately using (the more conventional) top-coded data (columns 1 and 2) and bottom-coded data (columns 3 and 4). Changes in market access are a strong and significant predictor of changes in lights, irrespective of which datasource is used; ceteris paribus, localities with bigger improvements in market access also experienced larger increases in lights per capita. Interestingly, the estimated coefficients are larger than in the specifications that exploit annual variation, perhaps reflecting that longer time lags have a higher signal to noise ratio. The estimated impact of improving market access is slightly lower when using bottom-coded data, hinting at the possibility that less developed localities (whose natural light emissions are best captured using top coded data) benefit more from improvements in market access than more prosperous towns (whose light emissions are best captured by bottom coded data). 33 Table 3: Robustness Checks (1) (2) (3) (4) (5) (6) ln NTL pc ln NTL pc ln NTL pc ln NTL pc ln NTL pc ln NTL pc IV IV IV IV IV IV MA=Market Access Varying Distance Decay Holding population fixed External MA ln MA (θ = 30) 0.184∗∗∗ (2.86) ln MA (θ = 60) 0.246∗∗∗ (2.66) ln MA (θ = 80) 0.367∗∗ (2.56) ln MA (θ = 40, pop ¯ ) 0.189∗∗∗ (2.81) ln MA (θ = 80, pop ¯ ) 0.364∗∗ (2.55) ln MA (incl. external) 0.357∗∗ ¯ ) (θ = 80, pop (2.53) ln(1+Pal.Fat.<5km) 0.0165∗ 0.0165∗ 0.0201∗∗ 0.0166∗ 0.0204∗∗ 0.0231∗∗ (1.81) (1.86) (2.14) (1.88) (2.16) (2.32) ln(1+Isr.Fat.<5km) 0.0169 0.0176 0.0162 0.0180 0.0162 0.0121 (1.01) (0.98) (0.87) (1.05) (0.86) (0.66) Year FE Yes Yes Yes Yes Yes Yes Locality FE Yes Yes Yes Yes Yes Yes N 1928 1928 1928 1928 1928 1928 adj. R2 0.342 0.310 0.261 0.345 0.260 0.259 Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Heteroscedasticity robust t -statistics clustered by locality in parentheses. NTL pc measures local (top-coded) Nighttime Lights emissions per capita (Source: NOAA) as described in Section 4. Market Access is calculated using: M Ait = j h(Tijt )Pjt , where P denotes destination population, T denotes travel time, and h(T ) = exp (− 1 2 T 2 /θ2 ) is a distance decay function. The choice of the distance decay parameter θ is indicated in the variable names (see Section 3 for details). In columns 1, 2 and 3, population counts are allowed to vary over time, whereas in columns 4, 5 and 6, they are kept constant (as is ¯ ), such that all variation in market access is driven by obstacle deployment. The augmented market access measure in indicated by pop column 6 adds Tel Aviv as a 12th destination (in addition to the 11 governorate capitals) and assigns it a population of 10 million in order to proxy for external market access. Pal.Fat.<5km (Isr.Fat.<5km) denotes the number of Palestinian (Israeli) fatalities within a 5km radius in a given calendar year (Source: B’Tselem). Market Access is instrumented using Checkpoints 10-25km, the number of checkpoints deployed in a radius between 10 and 25km away from the locality’s center (Source: UNOCHA). 34 Table 4: Difference-in-Differences Regressions (2006-2010) (1) (2) (3) (4) Top Coded Bottom Coded ∆ln NTL pc ∆ln NTL pc ∆ln NTL pc ∆ln NTL pc OLS IV OLS IV ∆ln Market Access 0.100** 0.225** 0.0773* 0.194** (2.19) (2.37) (1.73) (2.10) ∆ln(1+Pal.Fat.<5km) 0.0235 0.0277 0.0654*** 0.0693*** (1.28) (1.46) (3.65) (3.77) ∆ln(1+Isr.Fat.<5km) 0.0800 0.110** -0.0193 0.00888 (1.56) (1.97) (-0.39) (0.16) N 241 241 241 241 adj. R2 0.018 -0.012 0.050 0.023 Notes: * p < 0.10, ** p < 0.05, *** p < 0.01. Heteroscedasticity robust t -statistics in parentheses. ∆ ln NTL pc measures the change in the natural log of local lights emissions per capita between 2006 and 2010 (Source: NOAA) as described in section 4. Columns 1 and 2 use (conventional) top-coded data, while columns 3 and 4 use bottom- coded data (see section 4). ∆ln Market Access is the change in the natural log of market access between 2006 and 2010 calculated using: M Ait = j h(Tijt )Pjt , where P denotes destination population, T denotes travel time, and h(T ) = exp (− 1 2 T 2 /θ2 ) is a distance decay function with θ = 40 (see Section 3 for details). ∆ ln(1+Pal.Fat.<5km) (Isr.Fat.<5km) measures the change in the number of Palestinian (Israeli) fatalities within a 5km radius (Source: B’Tselem). In columns 2 and 4 ∆ln Market Access is instrumented using both the level and the first difference of Checkpoints 10-25km, the number of checkpoints deployed in a radius between 10 and 25km away from the locality’s center (Source: UNOCHA). 35 6.3 Counterfactual output estimates We conclude our analysis with a back-of-the-envelope calculation of the costs of mobility restrictions in terms of forgone output. Specifically, we simulate how much higher lights per capita would have been had obstacles not been present in the West Bank holding all else equal. The results are presented in Figure 14. Using our preferred distance decay parameter of 40, in 2005, lights per capita would have been 15.7 percentage points higher had there not been any mobility restrictions. Put differently, GDP would have been about 4.7% higher in 2005 assuming a GDP to lights elasticity of 0.3. Over time, however, restrictions have been alleviated, such that, by 2012, the cost of mobility restrictions amounted to approximately 2.6% of GDP. The easing of mobility restrictions has thus catalyzed growth. These estimates must be interpreted with caution, as they are sensitive to the choice of the distance decay parameter, with more permissive decay parameters yielding higher estimates; for instance, with a distance decay parameter of 80, the estimated GDP per capita loss amounts to 9.2% in 2005 and to 3.6% in 2012. Taking the estimates with θ = 40 as the lower bound and θ = 80 as the upper bound, over the period 2005- 2012 road closure obstacles reduced GDP per capita in the West Bank between 4.1% and 6.1% on average each year by repressing market access. These estimates are of course also sensitive the assumed elasticity of lights with respect to GDP. Moreover, the alleviation of obstacles is easy to reverse, and the context is marred by uncertainty and distortions (described in section 2), which may have dampened the output response to changes in market access. 7 Concluding remarks Assessing the impact of market access on economic performance is challenging because it usually evolves slowly and non-randomly. In this paper we exploit the repeated redeployments of Israeli road obstacles along the internal road network of the West Bank as plausibly exogenous spatio-temporally disaggregated shocks to market access of Palestinian towns. Complementary to other studies, which have mostly focused on large long-term shocks to market access resulting from once-in-a-generation transportation network extensions, our study examines the intensive margin, quantifying the economic losses and gains caused by year-to-year enhancements or deteriorations to inter-city transit. Furthermore, previous studies have focused on countries with large internal markets, such as China, India, and the United States. We study the West Bank, a territory roughly twice the size of the average American county. 36 Figure 14: Counterfactual output in the West Bank in the absence of restrictions .3 log difference in NTL per cap .15 .1.2 .25 2004 2006 2008 2010 2012 year θ=30 θ=40 θ=60 θ=80 Counterfactual output in the West Bank in the absence of restrictions (using different distance decay parameters θ) Despite these contextual differences, our findings resonate with this broader lit- erature, confirming the importance of spatial interdependence in the productivity of market economies. We find that the localized, temporary, and reversible shocks to road network connectivity significantly impact economic output (proxied for by nighttime lights). The association between market access and nighttime lights is robust to con- trolling for conflict. Moreover, it strengthens when we instrument market access using the count of obstacles located in a radius between 10 and 25km from the locality. The relationship is furthermore robust to using alternative market access and nighttime lights measures, and also obtains when estimating a difference-in-differences specifica- tion using 4-year changes in nighttime lights and market access. A back-of-the-envelope calculation suggests that over the period 2005-2012 market access constraints resulting from the deployment of road closure obstacles reduced GDP per capita in the West Bank between 4.1% and 6.1% on average each year. References Abrahams, Alexei (2018a). Hard traveling: Commuting costs and urban unemployment with deficient labor demand. ESOC Working Paper No. 8. Abrahams, Alexei (2018b). Not dark yet: the Israel-PA relationship 1993-2017. In 37 Proxy Wars (eds Eli Berman and David Lake), chapter 7. Cornell University Press. Adukia, Anjali, Asher, Sam and Novosad, Paul (2017). Educational investment re- sponses to economic opportunity: Evidence from Indian road construction. Unpub- lished manuscript. AIX Group (2013). Twenty years after Oslo and the Paris protocol: Changing course, averting crisis – from dependency to economic sovereignty. Technical Report. AIX Group. Alder, Simon (2017). Chinese roads in India: The effects of transport infrastructure on economic development. mimeo. Alesina, Alberto, Michalopoulos, Stelios and Papaioannou, Elias (2016). Ethnic in- equality. Journal of Political Economy, 124, number 2, 428–488. Allen, Treb and Arkolakis, Costas (2014). Trade and the topography of the spatial economy. The Quarterly Journal of Economics, 129, number 3, 1085–1140. Allen, Treb and Atkin, David (2017). Volatility and the gains from trade. mimeo. Allen, W Bruce, Liu, Dong and Singer, Scott (1993). Accesibility measures of US metropolitan areas. Transportation Research Part B: Methodological, 27, number 6, 439–449. Amodio, Francesco and Di Maio, Michele (2017). Making do with what you have: Conflict, input misallocation and firm performance. The Economic Journal. Anderson, James E and Van Wincoop, Eric (2004). Trade costs. Journal of Economic Literature, 42, number 3, 691–751. Arkolakis, Costas, Costinot, Arnaud and Rodriguez-Clare, Andres (2012). New trade models, same old gains? American Economic Review, 102, number 1, 94–130. Asher, Sam and Novosad, Paul (2017). Rural roads and structural transformation. mimeo. Banerjee, Abhijit, Duflo, Esther and Qian, Nancy (2012). On the road: Access to trans- portation infrastructure and economic growth in China. Technical Report. National Bureau of Economic Research. Baum-Snow, Nathaniel (2007). Did highways cause suburbanization? The Quarterly Journal of Economics, 122, number 2, 775–805. 38 Baum-Snow, Nathaniel, Brandt, Loren, Henderson, J Vernon, Turner, Matthew A and Zhang, Qinghua (2017). Roads, railroads, and decentralization of Chinese cities. Review of Economics and Statistics, 99, number 3, 435–448. Black, John and Conroy, M (1977). Accessibility measures and the social evaluation of urban structure. Environment and Planning A, 9, number 9, 1013–1031. Blankespoor, Brian, Bougna, Th´eophile, Garduno Rivera, Rafael and Selod, Harris (2017). Roads and the geography of economic activities in Mexico. World Bank Policy Research Working Paper No. 8226. B’Tselem (2007). Ground to a halt. Technical Report. B’Tselem. ı, Massimiliano and Miaari, Sami H (2018). The labor market impact of mobility Cal` restrictions: Evidence from the West Bank. Labour Economics, 51, 136–151. Chen, Xi and Nordhaus, William D (2011). Using luminosity data as a proxy for economic statistics. Proceedings of the National Academy of Sciences, 108, number 21, 8589–8594. Dalvi, M Quasim and Martin, KM (1976). The measurement of accessibility: some preliminary results. Transportation, 5, number 1, 17–42. Deichmann, Uwe (1997). Accessibility indicators in GIS. United Nations Statistics Division, Department for Economic and Policy Analysis, New York, NY, USA. Donaldson, Dave (2010). Railroads of the Raj: Estimating the impact of transportation infrastructure. Technical Report. National Bureau of Economic Research. Donaldson, Dave and Hornbeck, Richard (2016). Railroads and American economic growth: A market access approach. The Quarterly Journal of Economics, 131, number 2, 799–858. Duranton, Gilles and Turner, Matthew A (2012). Urban growth and transportation. Review of Economic Studies, 79, number 4, 1407–1440. Etkes, Haggay and Zimring, Assaf (2015). When trade stops: Lessons from the Gaza blockade 2007–2010. Journal of International Economics, 95, number 1, 16–27. Faber, Benjamin (2014). Trade integration, market size, and industrialization: evi- dence from China’s national trunk highway system. Review of Economic Studies, 81, number 3, 1046–1070. 39 Feyrer, James (2009). Distance, trade, and income-the 1967 to 1975 closing of the Suez Canal as a natural experiment. Technical Report. National Bureau of Economic Research. Geertman, Stan CM and Ritsema Van Eck, Jan R (1995). GIS and models of acces- sibility potential: an application in planning. International Journal of Geographical Information Systems, 9, number 1, 67–80. Geurs, Karst T and Van Wee, Bert (2004). Accessibility evaluation of land-use and transport strategies: review and research directions. Journal of Transport Geography, 12, number 2, 127–140. Gonzalez-Navarro, Marco and Quintana-Domeque, Climent (2016). Paving streets for the poor: Experimental analysis of infrastructure effects. Review of Economics and Statistics, 98, number 2, 254–267. Guy, Clifford M (1983). The assessment of access to local shopping opportunities: a comparison of accessibility measures. Environment and Planning B: Planning and Design, 10, number 2, 219–237. Hansen, Walter G (1959). How accessibility shapes land use. Journal of the American Institute of Planners, 25, number 2, 73–76. Harris, Chauncy D (1954). The market as a factor in the localization of industry in the United States. Annals of the Association of American Geographers, 44, number 4, 315–348. Head, Keith and Mayer, Thierry (2011). Gravity, market potential and economic development. Journal of Economic Geography, 11, 281–294. Heikkila, Eric J and Peiser, Richard B (1992). Urban sprawl, density, and accessibility. Papers in Regional Science, 71, number 2, 127–138. Henderson, J Vernon, Storeygard, Adam and Weil, David N (2011). A bright idea for measuring economic growth. American Economic Review, Papers and Proceedings, 101, number 3, 194–199. Henderson, J Vernon, Storeygard, Adam and Weil, David N (2012). Measuring eco- nomic growth from outer space. American Economic Review, 102, number 2, 994– 1028. 40 Hsu, Feng-Chi, Baugh, Kimberly E, Ghosh, Tilottama, Zhizhin, Mikhail and Elvidge, Christopher D (2015). DMSP-OLS radiance calibrated nighttime lights time series with intercalibration. Remote Sensing, 7, number 2, 1855–1876. Ingram, David R (1971). The concept of accessibility: a search for an operational form. Regional Studies, 5, number 2, 101–107. Jaeger, David A and Paserman, M Daniele (2008). The cycle of violence? an empirical analysis of fatalities in the Palestinian-Israeli conflict. American Economic Review, 98, number 4, 1591–1604. Jedwab, Remi, Kerby, Edward and Moradi, Alexander (2017). History, path dependence and development: Evidence from colonial railways, settlers and cities in Kenya. The Economic Journal, 127, number 603, 1467–1494. Jedwab, Remi and Moradi, Alexander (2016). The permanent effects of transporta- tion revolutions in poor countries: evidence from Africa. Review of Economics and Statistics, 98, number 2, 268–284. Jedwab, Remi and Storeygard, Adam (2017). The average and heterogeneous effects of transportation investments: Evidence from Sub-Saharan Africa 1960-2010. mimeo. Koenig, Jean-Gerard (1980). Indicators of urban accessibility: theory and application. Transportation, 9, number 2, 145–172. Kwan, Mei-Po (1998). Space-time and integral measures of individual accessibility: a comparative analysis using a point-based framework. Geographical Analysis, 30, number 3, 191–216. Min, Brian (2015). Power and the vote: Elections and electricity in the developing world. Cambridge University Press. Niksic, Orhan, Eddin, Nur Nasser and Cal` ı, Massimiliano (2014). Area C and the Future of the Palestinian Economy. World Bank Publications. Pinkovskiy, Maxim L (2013). Economic discontinuities at borders: Evidence from satellite data on lights at night. Unpublished manuscript, Massachusetts Institute of Technology. Redding, Stephen J and Sturm, Daniel M (2008). The costs of remoteness: Evidence from German division and reunification. American Economic Review, 98, number 5, 1766–97. 41 Redding, Stephen J. and Turner, Matthew A. (2015). Chapter 20 - transportation costs and the spatial organization of economic activity. In Handbook of Regional and Urban Economics (eds Gilles Duranton, J. Vernon Henderson and William C. Strange), Handbook of Regional and Urban Economics, volume 5, pp. 1339 – 1398. Elsevier. Song, Shunfeng (1996). Some tests of alternative accessibility measures: A population density approach. Land Economics, 72, 474–482. Storeygard, Adam (2016). Farther on down the road: transport costs, trade and urban growth in Sub-Saharan Africa. The Review of Economic Studies, 83, number 3, 1263–1295. Tuttle, Benjamin T, Anderson, Sharolyn J, Sutton, Paul C, Elvidge, Christopher D and Baugh, Kim (2013). It used to be dark here. Photogrammetric Engineering & Remote Sensing, 79, number 3, 287–297. UN-OCHA (2011). Barrier update: Special focus. UN-OCHA report. Wachs, Martin and Kumagai, T Gordon (1973). Physical accessibility as a social indicator. Socio-Economic Planning Sciences, 7, number 5, 437–456. Wu, Jiansheng, He, Shengbin, Peng, Jian, Li, Weifeng and Zhong, Xiaohong (2013). Intercalibration of DMSP-OLS night-time light data by the invariant region method. International Journal of Remote Sensing, 34, number 20, 7356–7368. 42 8 Appendix Table 5: Summary statistics Mean Std. Dev. Min Max N ln NTL pc -3.115 0.713 -5.405 -0.896 1928 ln NTL pc - deblurred -3.024 0.940 -7.195 -0.308 1928 ln Market Access (θ = 40)(default) 11.079 0.702 7.904 12.351 1928 ln Market Access (θ = 30) 10.585 0.938 5.498 12.18 1928 ln Market Access (θ = 60) 11.615 0.513 9.664 12.558 1928 ln Market Access (θ = 80) 11.946 0.413 10.635 12.69 1928 ¯ ) ln Market Access (θ = 40, pop 11.054 0.680 7.893 12.232 1928 ¯ ) ln Market Access (θ = 80, pop 11.921 0.369 10.625 12.57 1928 ¯ ) 12.003 ln Market Access (incl. external) (θ = 80, pop 0.391 10.625 12.935 1928 ln(1+Pal.Fat.<5km) 0.375 0.671 0 3.850 1928 ln(1+Isr.Fat.<5km) 0.0673 0.248 0 1.946 1928 Checkpoints 10-25km 17.384 8.586 0.417 41.083 1928 Notes:NTL pc measures local (top-coded) Nighttime Lights emissions per capita (Source: NOAA) as described in Section 4. Market Access is calculated using: M Ait = j h(Tijt )Pjt , where P denotes destination pop- 1 2 2 ulation, T denotes travel time, and h(T ) = exp (− 2 T /θ ) is a distance decay function. The choice of the distance decay parameter θ is indicated in parentheses, pop¯ denotes that the population is held constant such that all variation in market access is driven by obstacle deployment (see Section 3 for details). Pal.Fat.<5km (Isr.Fat.<5km) denotes the number of Palestinian (Israeli) fatalities within a 5km radius in a given calendar year (Source: B’Tselem). Checkpoints 10-25km is the number of checkpoints deployed in a radius between 10 and 25km away from the locality’s center (Source: UNOCHA). 43 Table 6: Alternative Measures of External Market Access (1) (2) (3) (4) ln NTL pc ln NTL pc ln NTL pc ln NTL pc IV IV IV IV External market (millions of people) 0 10 25 50 ln Market Access (incl. external) 0.364∗∗ 0.357∗∗ 0.348∗∗ 0.338∗∗ ¯ ) (θ = 80, pop (2.55) (2.53) (2.51) (2.48) ln(1+Pal.Fat.<5km) 0.0204∗∗ 0.0231∗∗ 0.0260∗∗ 0.0294∗∗ (2.16) (2.32) (2.43) (2.48) ln(1+Isr.Fat.<5km) 0.0162 0.0121 0.00982 0.00816 (0.86) (0.66) (0.54) (0.44) Year FE Yes Yes Yes Yes Locality FE Yes Yes Yes Yes N 1928 1928 1928 1928 adj. R2 0.260 0.259 0.239 0.208 Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Heteroscedasticity robust t -statistics clustered by locality in parentheses. NTL pc measures local (top-coded) Nighttime Lights emissions per capita (Source: NOAA) as described in Section 4. Market Access is calculated using: M Ait = j h(Tijt )Pjt , where P denotes 1 2 2 destination population, T denotes travel time, and h(T ) = exp (− 2 T /θ ) is a distance decay function. The set of destinations comprises the 11 governorate capitals and Tel Aviv, which serves to represent external markets. The size of the external market varies depending on the population assigned to Tel Aviv, indicated in the column headings. The choice of the distance decay parameter θ is 80 to allow for maximum impact of external market access (see Section 3 for details). Population counts are kept constant, such that all variation in market access is driven by obstacle deployment. Pal.Fat.<5km (Isr.Fat.<5km) denotes the number of Palestinian (Israeli) fatalities within a 5km radius in a given calendar year (Source: B’Tselem). Market Access is instrumented using Checkpoints 10-25km, the number of checkpoints deployed in a radius between 10 and 25km away from the locality’s center (Source: UNOCHA). 44