Bright investments: Measuring the impact of transport infrastructure using luminosity data in Haiti

Mitnik, Oscar A.; Sanchez, Raul; Yañez-Pagans, Patricia Working Paper Bright investments: Measuring the impact of transport infrastructure using luminosity data in Haiti IDB Working Paper Series, No. IDB-WP-935 Provided in Cooperation with: Inter-American Development Bank (IDB), Washington, DC Suggested Citation: Mitnik, Oscar A.; Sanchez, Raul; Yañez-Pagans, Patricia (2018) : Bright investments: Measuring the impact of transport infrastructure using luminosity data in Haiti, IDB Working Paper Series, No. IDB-WP-935, Inter-American Development Bank (IDB), Washington, DC, https://doi.org/10.18235/0001474 This Version is available at: https://hdl.handle.net/10419/208140 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc-nd/3.0/igo/legalcode IDB WORKING PAPER SERIES Nº IDB-WP-00935 Bright investments: Measuring the impact of transport infrastructure using luminosity data in Haiti Oscar A. Mitnik Raul Sanchez Patricia Yañez-Pagans Inter-American Development Bank Office of Strategic Planning and Development Effectiveness IDB Invest - Strategy and Development Department December 2018 December 2018 Bright investments: Measuring the impact of transport infrastructure using luminosity data in Haiti Oscar A. Mitnik Raul Sanchez Patricia Yañez-Pagans Cataloging-in-Publication data provided by the Inter-American Development Bank Felipe Herrera Library Mitnik, Oscar Alberto. Bright investments: measuring the impact of transport infrastructure using luminosity data in Haiti / Oscar A. Mitnik, Raul Sanchez, Patricia Yáñez. p. cm. — (IDB Working Paper Series ; 935) Includes bibliographic references. 1. Electric lighting-Haiti. 2. Transportation-Haiti-Finance. 3. Infrastructure (Economics)-Haiti-Finance. I. Sanchez, Raul. II. Yáñez, Patricia. III. Inter-American Development Bank. Office of Strategic Planning and Development Effectiveness. IV. IDB Invest. V. Title. VI. Series. IDB-WP-935 1300 New York Ave NW, Washington DC 20577 Copyright © Inter-American Development Bank. This work is licensed under a Creative Commons IGO 3.0 AttributionNonCommercial-NoDerivatives (CC-IGO BY-NC-ND 3.0 IGO) license (http://creativecommons.org/licenses/by-nc-nd/3.0/igo/ legalcode) and may be reproduced with attribution to the IDB and for any non-commercial purpose, as provided below. No derivative work is allowed. Any dispute related to the use of the works of the IDB that cannot be settled amicably shall be submitted to arbitration pursuant to the UNCITRAL rules. The use of the IDB's name for any purpose other than for attribution, and the use of IDB's logo shall be subject to a separate written license agreement between the IDB and the user and is not authorized as part of this CC-IGO license. Following a peer review process, and with previous written consent by the Inter-American Development Bank (IDB), a revised version of this work may also be reproduced in any academic journal, including those indexed by the American Economic Association's EconLit, provided that the IDB is credited and that the author(s) receive no income from the publication. Therefore, the restriction to receive income from such publication shall only extend to the publication's author(s). With regard to such restriction, in case of any inconsistency between the Creative Commons IGO 3.0 Attribution-NonCommercial-NoDerivatives license and these statements, the latter shall prevail. Note that link provided above includes additional terms and conditions of the license. The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank, its Board of Directors, or the countries they represent. http://www.iadb.org 2018 Bright investments: Measuring the impact of transport infrastructure using luminosity data in Haiti ∗ Oscar A. Mitnik†Raul Sanchez‡Patricia Ya˜ nez-Pagans§ December 2018 Abstract This paper quantifies the impacts of transport infrastructure investments on economic activity in Haiti, using satellite night-light luminosity as a proxy measure. Our identification strategy exploits the differential timing of rehabilitation projects across various road segments of the primary road network. We combine multiple sources of non-traditional data and carefully address concerns related to unobserved heterogeneity. The results obtained across multiple specifications consistently indicate that receiving a road rehabilitation project leads to an increase in luminosity values of between 6% and 26% at the communal section level. Taking into account the national level elasticity between luminosity values and GDP, we approximate that these interventions translate into communal section-GDP increases of between 0.5% and 2.1%, for communal sections benefited by a transport infrastructure project. We observe temporal and spatial variation in results, and crucially that the larger impacts appear once projects are completed and are concentrated within 2 km buffers around the intervened roads. Neither the richest or the poorest communities reap the benefits from road improvements, with gains accruing to those in the middle of the ranking of communal sections, based on unsatisfied basic needs. Our findings provide novel evidence on the role of transport investments in promoting economic activity in developing countries. Keywords: Haiti; night-time luminosity, road investments JEL codes: O1; O47; R4; D04 ∗We thank useful comments and suggestions from two anonymous IDB reviewers, Eduardo Cavallo, Rene Cortes, Marco Gonzalez-Navarro, Pablo Guerrero, Manuel Pastor, Maja Schling, Adam Storeygard, and from participants in presentations at CAF, IDB, IDB Invest, and the Annual LACEA Meetings. We also thank Naijun Zhou for his work in pre-processing some of the satellite data. †Inter-American Development Bank and IZA. Email: [email protected]. ‡IDB Invest. Email: [email protected]. §IDB Invest. Email: [email protected]. 1 Introduction Roads can have an important role in alleviating poverty (Gertler et al.,2014;GonzalezNavarro and Quintana-Domeque,2016). By reducing isolation, better roads should increase the accessibility to basic services (such as health and education), and to markets and employment centers, thus helping to reduce vulnerability and income variability (van de Walle and Cratty,2002). Despite several studies about the effects of road improvements on socio-economic outcomes, there is still limited evidence for Latin America and the Caribbean (LAC) countries. There is even less evidence in highly poor and vulnerable settings where data limitations make it difficult to conduct rigorous causal analyses. In this paper we exploit night-light satellite luminosity data, as well as detailed historical administrative information, to evaluate the impact of transport investments on economic activity in Haiti. In recent years a growing number of studies have relied on non-traditional sources of data for impact evaluation purposes (Alix-Garcia et al.,2015;Khanna,2016;Gendron-Carrier et al.,2018). Micro-satellite data holds particular promise given the availability of historical information, ample geographic coverage, and its granularity. For these reasons, it is increasingly being used for poverty mapping and economic analysis (IPA,2016). Yet the potential of satellite data cannot be realized without overcoming substantial technical challenges, that we highlight and address in this study. The case of Haiti is quite unique. It is the poorest country in the Western Hemisphere, with 59% of the population living below the poverty line (The World Bank, 2012). The country also faces deep regional economic imbalances, with 75% of the rural population being poor (UNDP,2013) and with its capital, Port-au-Prince, accounting for 80% of the country’s industrial, commercial, and financial activities. Fostering economic development outside of the capital has been a priority of Haiti’s government, and of multilateral development organizations working in the country. Road improvements are deemed as fundamental mechanisms that can help attain this objective, as road transport is the leading mode of transportation for cargo and passengers in the country. In 2010, Haiti experienced one of the strongest earthquakes in its history leaving almost three million people affected and large economic losses (CBS,2010; Cavallo et al.,2010). This event spurred an unprecedented program of foreign financial assistance to help rebuild the country’s infrastructure and to promote eco1 nomic development. Between 2010 and 2014, US$13.5 billion dollars had been invested or pledged for the country by multiple international organizations and through private charitable contributions (U.S. Congress,2014), roughly twice the size of the country’s GDP in 2010 (The World Bank,2018).1Despite the large dependence on financial aid, and the need for robust empirical evidence to better guide policy making, almost no rigorous impact evaluation studies have been conducted on the country, which is partly due to the limited availability of statistical information, and difficulty and cost of producing them (CEPR,2012). The main objective of this paper is to quantify the impacts of transport infrastructure investments on economic activity in Haiti, proxied by night-time satellite luminosity data (from now on referred as luminosity). For this, we generate a novel geo-referenced panel data set for the country, exploiting multiple sources of satellite information, secondary data, and detailed historical administrative information on infrastructure interventions in Haiti’s national road network, which have been funded by multilateral development institutions between 2004 and 2013.2We take advantage of the differential timing of road rehabilitation projects and compare changes in luminosity occurring in buffers around road segments that received a rehabilitation project (“treated”) versus those observed around segments that did not receive an intervention (“controls”). We estimate a variety of fixed-effects models at the communal section and pixel-level and conduct multiple robustness and placebo checks to reduce any concerns of unobserved heterogeneity. Although there are prior studies addressing the link between road infrastructure and economic activity, there are still relatively few papers that rigorously establish causal links, and the majority of them have been concentrated in developed countries or in Asian or African countries. Among these papers, early work by Chandra and Thompson (2000) exploits county-level industry data to show that counties next to a US Interstate Highway increase their level of economic activity, while those adjacent counties not directly on the highway see a decrease in economic activity. Datta (2012) and Ghani et al. (2016) evaluate the upgrade of a central highway network in India, finding that manufacturing grows disproportionately along the road network and that firms close to improved roads reduce their average stock on inventories and re-optimize their choice of suppliers. Banerjee et al. (2012) find 1Only taking into account multilateral or direct country aid, Haiti received disbursements for US$8.4 billion from 2011 to 2016 (IDB,2017). 2Given the availability of data, we focus on interventions from the Inter-American Development Bank, the World Bank, and the European Union. 2 that road construction in Indian villages results in greater access to government services, lower consumer prices, higher agricultural prices, increased employment outside of agriculture and less daily migration. Casaburi et al. (2013) show that village feeder roads in Sierra Leone contribute to reduce market prices of local agricultural goods. The use of luminosity data is relatively recent in the economics field and follows the work of Henderson et al. (2012) that shows that country-level mean light intensities are a good proxy for GDP. In the transport sector, Storeygard (2016) uses city-level luminosity values, interacted with global oil price shocks and distances to the nearest port, to show that there is a significant inverse relationship between transport costs and urban economic output in multiple African countries. Gonzalez-Navarro and Quintana-Domeque (2016) show a more dispersed distribution of luminosity data in cities that have implemented subway systems suggesting a decentralization of economic activity. Alder (2017) uses a general equilibrium trade framework to compare the transport network configuration strategy followed by India, of building a highway connecting the four largest economic centers of the country (Golden Quadrilateral), versus the Chinese strategy, of connecting intermediate-sized cities. Using district-level data he finds that the Chinese strategy can lead to further gains and less unequal effects in economic activity when compared to the Indian strategy. Finally, Khanna (2016) explores the impacts on economic activity of transport infrastructure investments in the Golden Quadrilateral of India. Connecting nodal cities with straight lines as instruments for the endogenous placement of road networks he shows evidence of spatial spillovers as a result of road investments using luminosity data. Despite the increased popularity of satellite imagery in recent research applications, there are almost no studies oriented to study the impact of interventions using luminosity data in the LAC region and none in the transport sector. The only study in the LAC region that has used this data is Corral and Schling (2017) that applies synthetic control methods to show that shoreline stabilization investments have beneficial medium-term effects in economic growth in Barbados. Our aim with this work is not only to evaluate for the first time the impacts of road investments in Haiti and in LAC, but to showcase how non-traditional sources of data may be more widely used for impact evaluation in transport and in areas where access to information may be a limitation. We believe that this exercise could be usefully replicated in multiple countries, where transport investments are an important part 3 of infrastructure investments and no evidence about its effectiveness is available. From a methodological perspective, this paper also moves several steps further by carefully addressing multiple of the concerns related to unobserved heterogeneity. These concerns are central in the literature given the non-random placement of infrastructure investments (Yanez-Pagans et al.,2018). In particular, we run alternative specifications that rely on introducing multiple fixed effects. We test parallel-trend assumptions, present several placebo tests, and a series of robustness checks. In addition, although our preferred analysis is at the communalsection level, we also explore effects within one squared-kilometer areas (pixel level).3This allows for a better understanding of how localized or dispersed impacts can be. Moreover, by exploiting historical data for a period of more than ten years, our focus is both on short and medium-term effects, which is relevant for transport investments that usually may take some time to deliver effects. One the main limitations of previous transport evaluations that have used primary data is that they only cover two periods of time (baseline and follow-up) and cannot uncover dynamic effects or test model assumptions (Valdivia,2011). Our main result indicates that roads generate approximately between a 6% (considering the preand post-completion periods) and 26% (considering only the post-completion period) increase in night-time luminosity at the communal-section level. Taking into account the national level elasticity between luminosity values and GDP, we approximate that this type of transport interventions translate in communal section-GDP increases of at least 0.5% after investment approval, and possibly as high as 2.1%, after road rehabilitation completion. These average effects hide some important heterogeneity. First, communal sections that gain the most from these investments appear to be those in the middle of the income distribution, while we do not see any significant effects in the richest or poorest areas. Second, our results indicate that most impacts appear four or more years after project approval, at a similar level to the effects for investments completed during our analysis period. This makes clear that roads need to be fully operational before households and communities start realizing large benefits. Third, we find no evidence that those communities experiencing the largest gains in transport cost savings and accessibility (i.e. those that are further away from main cities) are necessarily those 3As is discussed below, there could be drawbacks to relying on pixel-level luminosity data, and this is why we consider pixel-level analyses only as additional useful information, but they are not our main focus. 4 This information, however, is only available for cross-sections years 1995, 2000, 2005 and 2010. GRC data are less ready to recognize dim light sources, yet can quantify variety in light inside locales that are top-coded in the DMSP/OLS version (Gonzalez-Navarro and Turner,2016).11 5 Results 5.1 Baseline descriptive statistics Table 4presents summary statistics for the communal sections in the sample. We divide them in two groups. The first is composed by those communal sections that were never treated during the study period, this means none of the buffers of road segments that received a rehabilitation project touch the boundaries of these communal sections. The second are those communal sections that received a road improvement intervention at any given year within the time frame of the analysis. We can see that there are no significant differences across some of the variables, such as the greenness of the vegetation (NDVI) or the number of other types of projects (energy, roads, housing, etc.) that they received. There are however, some differences worth pointing out. Ever treated communal sections have smaller populations and are exposed to more rainfall. As mentioned above, the implementation of a DID method does not require equality in variable levels at baseline, but parallel trends. We conduct tests to show that this is the case. It should be noted that the deblurring process removes the upper bound of the raw DMSP/OLS data. 5.2 Main effects and interpretation Table 5presents, step by step, how we arrive to our preferred specification. Using the IHS of deblurred luminosity at the communal section level (with a 2.5 km buffer of influence) as the dependent variable, columns (1) to (4) gradually introduce multiple fixed effects in the model. Without covariates or fixed effects (column 1), the magnitude of the coefficient of interest is 0.201, which means that, on average, a communal section that received a road improvement intervention sees 11See Ziskin et al. (2010) for an explanation of the underlying methodology. 11 an increase of 22.3% in its luminosity levels12. Introducing only communal section fixed effects (column 2) makes the effect even stronger to almost 35%, but introducing year fixed effects (column 3) or both year and satellite fixed effects (column 4) reduces the impacts to 11%. In column (5) we include contemporaneous time-varying controls considering other infrastructure projects implemented in the country that could also affect the luminosity values recorded. More specifically, we control for energy projects, other road improvement projects in the secondary or tertiary network, housing and urban development projects13. We also control for population, natural disasters occurrence and a dummy variable for areas affected by the 2010 earthquake. In column (6) we include not only the contemporaneous effects but also one lag for annual levels of precipitation, NDVI, and natural disaster occurrence. As mentioned above, the AIC indicates that one lag is the optimal lag specification. Even though there are no large differences between columns (5) and (6), our preferred specification is presented in column (6) where we include the complete set of covariates and where impacts are around 7%.14 As discussed above, we use deblurred luminosity to account for potential measurement error in luminosity levels. However, in Table 6we explore alternative measures of luminosity and see how they might affect the obtained results. In column (1) we replicate the regression in column (6) from Table 5, which is our preferred specification using deblurred luminosity. In columns (2) and (3) we estimate the same specification using two alternative measures of luminosity. In column (2) we use the IHS of the DMSP-OLS raw luminosity value, as provided by NOAA. Using this variable we find an even larger effect of 11%. In column (3) we use the GRC measure that is not upper-bounded. This alternative measure of luminosity suggests that the impact of road interventions are in the order of 15%. The fact that 12When regression models have log transformed outcomes the impact of a one-unit change in a covariate (X) is calculated by exponentiating the coefficient. In this case it will be (exp(β1)−1) = exp(0.201) −1 = 0.223. When the estimated coefficient is less than 0.10 the interpretation that a unit increase in X is associated with an average of 100 ∗β1percent increase in Y works well. We refer to the exponetiated coefficients throughout the text, unless we specify otherwise) 13The list or number of other projects included is limited to IDB and World Bank projects, given that it was constructed based on publicly available information. If it were to be the case that there are control areas projects for which we do not have information, this would create a downward bias, making our results conservative estimates of the road improvement projects impacts. 14Given that we are (implicitly) averaging observations in years where have two satellites reporting data, as an additional robustness check, we run regressions including only the luminosity values of the newest satellite each year, while still keeping the satellite fixed effects to capture the changes in technology. Results are consistent and the preferred model (column 6 in Table 5) shows an average impact of 8%. 12 this last result is larger than the one obtained in our preferred specification might be telling us that there are also gains in economic activity in urbanized and highly dense areas that are not being accounted for by the bounded measure. 5.3 Robustness and placebo tests Table 7presents several robustness checks. Column (1) shows again our preferred specification from column (6) of Table 5, which is based on a buffer of influence of 2.5 km at each side of the road. In columns (2) and (3) we report results obtained when changing the buffer to 3.5 and 5 km to each side of the road, respectively. These changes have a direct effect on the communal sections that are selected in the sample, as shown in the last line of the Table 7, where we can see the gains in sample size. Despite these changes, results remain close to our preferred specification, between the 6-7% range. In columns (4) to (6) of Table 7we explore the robustness of the results to eliminating the largest populated areas. As some of the treated road segments serve to connect some larger cities, such as Portau-Prince and Cap-Ha¨ ıtien, we want to rule out the possibility that the luminosity gains that we observe are concentrated in those urban areas or are the result of agglomeration. For this, we exclude all those communal sections that are part of Port-au-Prince (column 4) or Cap-Ha¨ ıtien (column 5), or both (column 6). Results remain stable in (4) and are equal to 7% and have just a marginal decrease in (5) and (6) to 6%. As discussed above, a key identification assumption of the DID regressions is that the treated and control areas do not exhibit differences in trends before the interventions. We test this parallel trends assumption in two ways, as shown in Table 8. In column (1) we add a dummy variable identifying those communal sections that will be treated in the future, prior to approval. The coefficient on that dummy variable is not statistically significant, and the treatment effect post-approval of a project goes up to 12%. In column (2) we split the pre-treatment dummy in subperiods prior to approval of a project (one, two, three, and four or more years prior to approval). None of those coefficients are significant, as expected, and a Wald test of joint significance of all four coefficients is not significant either. The treatment effect in this case is also equal to 12%. Columns (3) and (4) of Table 8present alternative tests of our identification strategy. In particular, we construct two placebo treatments, for which we do not 13 expect to find a significant treatment effect. In column (3) we take luminosity values from 1992 to 1999 and use them to replace the actual luminosity values observed from 2006 to 2013. As we do not have enough historic information to replace 2004 and 2005 luminosity values (the first two years for which there is a treatment), we drop those two years from the regressions. We should not expect to see any impact of treatment when using those lagged years as output variables, and indeed the coefficient is close to zero and not statistically significant. We also conduct tests (not reported here15) replacing the values of other years with the historic data, for example, replacing values between 2004 and 2011 and dropping 2012 and 2013. In all cases the coefficient is close to zero and not statistically significant. In column (4) we create a different placebo treatment, this time randomizing the timing of the treatment (prior to actual road construction), and repeating this exercise for 200 replications. Here again, as expected, the results show there is no treatment effect. 5.4 Heterogeneous treatment effects We test for heterogeneous treatment effects to further understand how impacts are distributed across space and time. Exploiting data from Haiti’s Poverty Map from 2003 and looking at the distribution of the measure of Unsatisfied Basic Needs (UBN) across communal sections reported in this source, columns (1), (2), and (3) in Table 9divide communal sections in three groups: poorest, poor, and least poor.16 We observe that impacts are coming from communities in the middle part of the distribution (15% effect), which indicates that while the richest communities are not benefiting from these road improvement projects, the poorest of the poor are not gaining either. A test of equality of coefficients across groups, confirms that this difference is significant. This finding highlights the need to provide complementary policies in other areas (e.g. education, health, poverty reduction programs) to support those in the base of the pyramid, and that could allow them to take full advantage of the improvement in accessibility and transport connectivity. In columns (4) and (5) of Table 9we test whether distance to the main cities (i.e. Port-Au-Prince or Cap-Ha¨ ıtien) has any role in explaining the impacts observed. 15Results can be requested from the authors. 16MPCE (2004) classifies the population according to their level of access to basic social services. Based on the official classification of the UBN, the poorest population corresponds to very weak and extremely weak, poor is the population with low access, and least poor corresponds to less weak and moderately weak. 14 One hypothesis, based on the notion of agglomeration effects, is that those communal sections that are closer to main urban areas might be the ones that exhibit the largest growth in economic activity (i.e. luminosity values). As opposed, if impacts are driven by transport costs savings and gains in accessibility, one would expect that those communal sections further away should be experiencing higher impacts. Results seem to indicate that there are no differences in impacts across distance. Although the estimated coefficient for communal sections that are further away is statistically significant and the one for those that are closer is not, the magnitudes are very similar and not statistically different. Variations in statistical significance across coefficients seems more related to lack of statistical power in the case of communal sections that are closer to the cities. Finally, in columns (6) and (7) of Table 9we test whether there are heterogeneous impacts at different years after project approval. This serves to test what the short, medium and long run effects of these investments are. In column (6) we interact the treatment dummy with number of years since project approval. We can see that most impacts appear after 4 or more years after project approval and the estimated impact during this period is close to 25%. This finding seems reasonable to the extent that households and communities might not experience any gains during construction, but can only get the benefits once the improved road is fully operational. We further test this hypothesis in column (7) by dividing the treatment variable in two periods, construction and after completion. Results show that impacts during construction are marginally significant and smaller (6%)than those that appear after completion (26%). The fact that our average effect is closer in magnitude to the impact computed here during the construction phase is due to observing only a few projects after completion in our sample as shown in Table 1. 5.5 Pixel-level results In all the analyses so far we have relied on luminosity values aggregated at the communal section level. We have done this as we believe it is the more appropriate unit of measure to provide an economic interpretation of the results. However, we also estimate the model using pixel-level data. We do not see this exercise as a way to estimate changes in economic activity at the pixel level, since nightlights might not be appropriate proxies for GDP for very small areas. Rather, the analysis gives us as a way to explore whether luminosity impacts are concentrated only around 15 the intervened roads or not, and how far away from the roads the impacts might materialize. Table 10 serves as way to establish the comparability of the prior communal section results and the pixel ones, for different specifications of the buffer of influence (columns 1 to 3), eliminating the pixels associated to Port-au-Prince (column 4), Cap-Ha¨ ıtien (column 5), or both (column 6 ), and finally utilizing the GRC measure of luminosity (column 7). The pixel-level effects appear smaller than the communal section-level ones, in the order of 6%. Column (7) meanwhile, suggests large effects (17%) when using the GRC luminosity measure, which is in line with the results in Table 6which also showed much higher effects using the GRC luminosity measure17. Table 11 exploits the pixel-level regressions to explore heterogeneous treatment effects across different buffers of distance. As mentioned before, this is the central objective of this section and allows to have some evidence on whether there isa gradient of impacts by distance to the intervened road. For each definition of the buffer of influence, three regressions are run: the first one only uses pixels within less than a 1 km from the road, the second uses only pixels between 1 and 2 km from the road, and the third one uses pixels above 2 km, and up to the boundary of the buffer of influence (2.5, 3.5 or 5 km). The results show that treatment effects on pixels within 1 km of the road are around 11%, those for pixels between 1 and 2 km are smaller but still significant, between 7% for the 2.5 km buffer to 10% for the 5 km buffer. Finally, the treatment effects pretty much disappear after 2 km (they are only significant at the 10% level for the 5 km buffer). This suggests that the effects may be even larger than those estimated using communal section level data, but that they drop off fairly quickly, being concentrated in a buffer of 2 km in either side of the roads. 5.6 Night-light luminosity data as a proxy for economic activity An important assumption in this analysis is that the outcome variable (luminosity values) is a good proxy for economic activity and development. To empirically test this assumption, we use national level data and compute the elasticity between the 17The fact that the estimated coefficient with the GRC measure is larger than the one estimated with the upper-bounded measure reflects the fact that the GRC data source, by combining very small values with unbounded values, is reducing the weight or ignoring really small values while also highlighting or putting more weight on larger values that were previously not available. 16 luminosity value and the Gross Domestic Product (GDP). Table 12 reports these elasticities for both the raw luminosity values and the deblurred luminosity. We start with the most basic specification, without including any controls, and obtain elasticities between 0.06 and 0.07 for raw and deblurred luminosity, respectively. We then start adding satellite fixed effects to take into account changes in technology that might affect the values reported within a given year and a fixed-effect for the 2010 earthquake. In the most complete specification, reported in columns (3) and (6), we obtained elasticities of 0.06 and 0.08 for raw and deblurred luminosity, respectively. One of the most challenging aspects of working with luminosity data is the economic interpretation of results. To provide a first approximation, we take into account that the national level elasticity we estimate for the deblurred luminosity values and GDP is around 0.08. If we assume that this elasticity also holds at the communal-section level, and considering that the impact or receiving a road improvement intervention on luminosity values that we consistently estimate lies within 6% and 26% (for the communal section regressions), this would imply that road investments could have generated between 0.5% and 2.1% increase in the communal-section GDP in Haiti. We also exploit household-level data from the Demographic and Health Survey (DHS) to compute the elasticity between the average household asset index and the luminosity values computed at the communal section level. Results are reported in Table 13. Column (1) shows that a one percent increase in luminosity at the communal section iis associated with a two percent increase in the average asset index (DHS a) of the communal section. Column (2) includes fixed effects at the communal section level. This specification obtains an elasticity of 0.05. We also include year (column 3) and satellite (column 4) fixed effects, and the elasticities are around 0.02. By combining these results with the estimated impacts, we suggest that receiving a road intervention could lead to an increase of 0.1% and 0.5% in the average household asset index at the communal section level. 6 Conclusion We provide novel evidence on the impacts that transport investments have had in promoting economic activity, proxied by satellite luminosity data, in Haiti. Given the lack of information and fragile conditions in the country, there are still very few causal studies providing empirical evidence on the impacts generated by the large 17 package of financial assistance that the country has received in the past years, particularly since the 2010 earthquake. Beyond the contributions to the discussion around the effectiveness of financial aid in Haiti, this is also the first study providing evidence on the impacts of transport infrastructure investments using satellite data in Latin America and the Caribbean and constitutes also one of the very few causal analysis on this topic. As it was shown throughout the study, from a methodological perspective, this paper also moves several steps further in the literature by carefully addressing multiple of the concerns related to unobserved heterogeneity. These concerns are key in the related literature given the non-random placement of infrastructure investments and the inherent identification challenges that this brings. The results we consistently obtain across multiple specifications indicate that a road rehabilitation project leads to an increase in luminosity values between 6% and 26% at the communal section level (the preferred level of analysis for an economic interpretation of results). Taking into account the national level elasticity between luminosity and GDP, we approximate that transport investments have generated between 0.5% and 2.1% increase in the communal-section GDP of the intervened communal sections, during the period of analysis. The average effects we observe hide some important heterogeneity. First, communal sections that gain the most from these investments are those in the middle of the distribution of an Unsatisfied Basic Needs indicator, while we do not see any significant effects in the richest or poorest areas according to this same metric. This implies that in order to tackle poverty reduction and promote inclusive growth in the country, transport investments will not be impactful entirely by themselves, but rather there is need for complementary policies across different sectors (e.g. education, health, poverty reduction programs) to support and lift those in the base of the pyramid. Second, our results indicate that most impacts appear four or more years after project approval. This is consistent with the notion that roads need to be fully operational before households and communities start seeing any benefits. The full or more longer term impacts are actually much larger and could be close to a 26% increase in luminosity values (i.e. 2.1% increase in GDP). Third, we find no evidence that those communities experiencing the largest gains in transport cost savings and accessibility are necessarily those attracting more economic activity. This is reflected in the fact that regardless of the distance to the main cities, effects remain constant. Finally, our pixel-level analysis suggest that impacts are consis18 tently concentrated in a buffer of 2 km on either side of the intervened roads, and drop off fairly quickly after that. The context of Haiti is quite unique, given the high levels of poverty in the country and the sizable natural disasters experienced over the past years, particularly the 2010 earthquake. Our empirical strategy explicitly controls for all natural disasters and specifically for the 2010 earthquake. In addition, robustness checks excluding large cities that were affected by the earthquake, such as Port Au Prince, still provide similar conclusions. Thus, we believe our results are not contaminated or just capturing the effects of the earthquake recovery effort. Nevertheless, it is still possible that the large and positive impacts observed in this case might not necessarily occur in other (more developed) settings. Future research replicating this approach for other countries might provide useful insights on the heterogeneity of effects across areas with different institutional and socioeconomic characteristics, and baseline development levels. It is reasonable to expect that the economic multipliers of infrastructure investments are larger in contexts with higher levels of poverty and lower development level. Moving forward, this work opens multiple avenues or opportunities for evaluation research. The methodological approach we propose in this study can be useful for public agencies, international organizations, and other institutions seeking to evaluate the impacts of transport investments, particularly in settings where data availability may be limited. Important challenges still remain and have to do with the appropriate economic interpretation of these results, and a better understanding of what nigh-time lights are really measuring. The literature has done some progress along these lines recently (Machemedze et al.,2017;Chen and Nordhaus,2011; Klemens et al.,2015), but in contexts with limited other data available, these correlations will still need to be taken as given. 19 References Abrahams, A., Lozano-Gracia, N., and Oram, C. (2016). Deblurring DMSP Nighttime Lights. Technical report, Working Paper. Alder, S. (2017). Chinese Roads in India: The Effect of Transport Infrastructure on Economic Development. University of North Carolina at Chapel Hill. Alix-Garcia, J. M., Sims, K. R., and Ya˜ nez-Pagans, P. (2015). Only one tree from each seed? Environmental effectiveness and poverty alleviation in Mexico’s Payments for Ecosystem Services Program. American Economic Journal: Economic Policy, 7(4):1–40. Angrist, J. and Pischke, J. (2009). Mostly Harmless Econometrics. An Empiricist’s Companion. Princeton University Press. Banerjee, A., Kumar, S., and Pande, R. (2012). Connectivity and Rural Development: Examining India’s Rural Road Building Scheme. BID (2018). Haiti invests heavility in rebuilding roads. https://www.iadb.org/en/news/webstories/2009-04-13/ haiti-invests-heavily-in-rebuilding-roads%2C5339.html. Web Stories. Accessed July 10, 2018. Casaburi, L., Glennerster, R., and Suri, T. (2013). Rural roads and intermediated trade: Regression discontinuity evidence from Sierra Leone. Cavallo, E., Powell, A., and Becerra, O. (2010). Estimating the direct economic damages of the earthquake in haiti. The Economic Journal, 120(546):F298– F312. CBS (2010). Red cross: 3m haitians affected by quake. https://www.cbsnews. com/news/red-cross-3m-haitians-affected-by-quake/. Accessed June 8, 2018. CEPR (2012). Center for Economic and Policy Research. Lack of Data Prevents Measurement of Aid Effectiveness, Impact. http://cepr.net/blogs/haiti-relief-and-reconstruction-watch/ lack-of-data-prevents-measurement-of-aid-effectiveness-impact. Accessed June 5, 2017. 20 Table 1 – Continued from previous page Financing Road Section Year Year Amount Length institution approved finished (in US$ M) (in km) WB1RN3 Barriere Battant - Carrefour la Mort 2006 2013 16 8.0 WB2RN2 Carrefour - Miragoˆ ane 2010 2017 65 86.1 RN4 Carrefour Dufort - Jacmel 44.6 Financing institutions: Inter American Development Bank (IDB): IDB1(HA0087); IDB2(1922/GR-HA); IDB3(HA-L1046); IDB4(HA-L1054). The European Union (EU): EU1(FED/2005/017-548); EU2(FED/2005/017-548); EU3(unknown). The World bank (WB): WB1(P095523); WB2(P120895). 27 Table 2: List of natural disasters Year Description Location 1998 09/23, Hurricane Georges Sud-Est and Nord-Ouest 2002 05/24-05/27, tropical storms and flooding Camp Perrin, L’Asile, Anse-` a-Veau. 2004 05/23–05/24, torrential rains Mapou, Belle-Anse, Bodary, and Fonds-Verrettes: Sud-Est d´ epartement. 09/10, Hurricane Ivan Southern peninsula and west coast. 09/18-09/19, Hurricane Jeanne Artibonite. Gona¨ ıves. 2005 07/06-07/07, Hurricane Dennis Bainet, Grand-Goˆ ave, Les Cayes. 10/04, floods P´ etion-Ville and Grand-Goˆ ave in the Ouest d´ epartement. 10/17–10/18, Hurricane Wilma West and South of Haiti. 10/23, tropical Storm Alpha Grand’Anse and Nippes. 10/25, flooding. Port-de-Paix, Bassin-Bleu, Anse-` a-Foleur. 2006 11/22-11/23, heavy rain Grand’Anse Department and the Nippes and Nord-Ouest d´ epartements. 2007 03/17, floods. Grand’Anse, J´ er´ emie, Abricots, Bonbon, Les Irois Continued on next page 28 Table 2 – Continued from previous page Year Description Location Sud-Est: Jacmel, Ouest, Cit´ e Soleil, Delmas, Port-au-Prince (Carrefour-Feuilles, Canap´ e Vert) Nord-Ouest:, Port-de-Paix, Saint-Louis du Nord, Anse-` a-Foleur, Cap-Ha¨ ıtien, Nord-Est: Ferrier, Ouanaminthe. 05/08-05/09, torrential rain Nord, Nord-Est and Sud d´ epartements. 2008 08/26, Hurricane Gustav Sud and Grand’Anse d´ epartements. 09/01, Hurricane Hanna Artibonite and Nord-Est d´ epartements. 09/06, Hurricane Ike Nord, Ouest and Nord-Ouest d´ epartements. 2009 10/20, heavy rain Carrefour. 2010 01/12, earthquake of magnitude 7.0 Port-au-Prince. 01/20, a second earthquake of magnitude 6.1 Port-au-Prince. 11/05, Hurricane Tomas South-west. 2016 10/03-10/04, Hurricane Matthew southwestern Haiti near Les Anglais. 29 Table 3: List of other infrastructure projects Sector Financing Description Year Amount institution approved (in US$ M) Energy IDB5P´ eligre Hydroelectric Plant Rehabilitation Program 2008 12.5 IDB6Rehabilitation of Electricity Distr. System in Port-au-Prince 2010 15.7 IDB7Rehabilitation of Electricity Distr. System in Port-au-Prince, Phase II 2010 14 IDB8Supplementary Financing for the Peligre Hydroelectric Plant 2011 20 IDB9Rehabilitation of the P´ eligre Transmission Line 2014 7.7 Urban development IDB10 Urban Rehabilitation Program 2005 50 IDB11 Support to the Shelter Sector Response Plan 2010 30 IDB12 Infrastructure Program 2011 55 IDB13 Productive Infrastructure Program 2012 50 IDB14 Productive Infrastructure Program II 2013 40.5 IDB15 Water Management Program in the Artibonite Basin 2013 25 IDB16 Productive Infrastructure Program III 2014 55 IDB17 Sustainable Coastal Tourism Program 2014 36 IDB18 Productive Infrastructure Program IV 2015 41 WB3Rural Community Driven Development - Additional Financing II 2010 15 Continued on next page 30 Table 3 – Continued from previous page Sector Financing Description Year Amount institution approved (in US$ M) WB4Urban Community Driven Development Project 2011 30 WB5Port-au-Prince Neighborhood Housing Reconstruction 2011 65 Other transport IDB19 Pont-Sonde - Mirebalais Highways and Access Roads 1990 53 IDB20 Support for Transport Sector in Haiti II 2012 53 IDB21 Emergency Road Rehabilitation Program in Response to Hurricane Sandy 2012 17.5 WB6AF Infrastructure & Institutions Emergency Recovery 2012 35 Financing institutions: Inter American Development Bank (IDB): IDB5(HA-L1032); IDB6(HA-L1014); IDB7(HA-L1035); IDB8(HA-L1038); IDB9(HA-L1100); IDB10 (HA-L1002); IDB11 (HA-L1048); IDB12 (HA-L1055); IDB13 (HA-L1076); IDB14 (HA-L1081); IDB15 (HA-L1087); IDB16 (HA-L1091); IDB17 (HA-L1095); IDB18 (HA-L1101); IDB19 (HA0049); IDB20 (HA-L1058); IDB21 (HA-L1086). The World bank (WB): WB3(P118139); WB4(P106699); WB5(P125805); WB6(P130749). 31 Table 4: Communal section covariates, averages of years 2000-2003 Never treated Ever treated p-value Variable Mean Std. Dev. Mean Std. Dev. difference Deblurred luminosity 104.36 339.06 49.08 147.58 0.01 IHS of deblurred luminosty 0.59 1.47 0.49 0.99 0.20 Normalized Difference Vegetation Index (NDVI) Average indext0.66 0.16 0.63 0.15 0.01 Total rainfall (in MM, 1000) Total rainfallt8.23 7.86 10.71 7.37 0.00 Population (1000 hab.) 36.05 66.99 18.77 24.06 0.00 Number of energy projects 0.00 0.00 0.00 0.00 - Number of other road projects 0.04 0.20 0.04 0.21 0.90 Number of other types of projects 0.00 0.00 0.00 0.00 - Natural disasters Number of Natural disasterst0.00 0.00 0.01 0.11 0.01 32 Table 5: Communal-section level treatment effect on luminosity using 2.5km buffer Deblurred luminosity (1) (2) (3) (4) (5) (6) Treatment 0.201* 0.298*** 0.101*** 0.101*** 0.073** 0.066** (0.105) (0.036) (0.035) (0.035) (0.033) (0.033) Observations 4,554 4,554 4,554 4,554 4,554 4,554 Number of CS 207 207 207 207 207 207 R-squared 0.005 0.059 0.144 0.146 0.177 0.178 Communal section FE NO YES YES YES YES YES Year FE NO NO YES YES YES YES Satellite FE NO NO NO YES YES YES Contemporaneous covariates NO NO NO NO YES YES Lagged covariates NO NO NO NO NO t−1 SE clustered at the communal section level between parentheses. *p<0.10; **p<0.05; ***p<0.01. Treatment represents the coefficient β1from the regression following equation (1). 33 Table 6: Communal-section level treatment effect using alternative luminosity measures Deblurred data Raw data GRC (1) (2) (3) Treatment 0.066** 0.104*** 0.137** (0.033) (0.030) (0.055) Observations 4,554 4,554 1,035 Number of CS 207 207 207 R-squared 0.178 0.477 0.498 Communal section FE YES YES YES Year FE YES YES YES Satellite FE YES YES NO Contemporaneous covariates YES YES YES Lagged covariates t−1t−1t−1 SE clustered at the communal section level between parentheses. *p<0.10; **p<0.05; ***p<0.01. Column (1) computes the treatment effect (β1) from the regression following equation (1) using the deblerrud data according to the methodology proposed by Abrahams et al. (2016). Column (2) uses the data from the Defense Meteorological Satellite Program - Operational Linescan System (DMSP-OLS). The products are 30 arc second grids, spanning -180 to 180 degrees longitude and -65 to 75 degrees latitude. Available at https://ngdc.noaa.gov/eog/ dmsp/downloadV4composites.html. Column (3) uses an alternative measure of luminosity (Global radiance calibrated, GRC). https://ngdc.noaa.gov/eog/dmsp/download_ radcal.html. 34 Table 7: Communal-section level treatment effect on luminosity using alternative specifications Alternative buffers No PaP 2.5km 3.5km 5km No PaP No CH and no CH (1) (2) (3) (4) (5) (6) Treatment 0.066** 0.072** 0.058** 0.063** 0.058* 0.058** (0.033) (0.029) (0.026) (0.030) (0.030) (0.029) Observations 4,554 5,126 5,632 4,488 4,488 4,422 Number of CS 207 233 256 204 204 201 R-squared 0.178 0.168 0.166 0.187 0.180 0.191 Communal section FE YES YES YES YES YES YES Year FE YES YES YES YES YES YES Satellite FE YES YES YES YES YES YES Contemporaneous covariates YES YES YES YES YES YES Lagged covariates t−1t−1t−1t−1t−1t−1 SE clustered at the communal section level between parentheses. *p<0.10; **p<0.05; ***p<0.01. Treatment represents the coefficient β1from the regression following equation (1). Specifications (1)-(3) use alternative buffers. Specifications (4)-(6) eliminate the communal sections that correspond to Port-au-Prince (PaP), Cap-Ha¨ ıtien (CH), and both. 35 Table 8: Specification tests Parallel trend Placebo outputs Lag 90’s output Random (no 2004-2005) timing (1) (2) (3) (4) Treatment 0.111* 0.106* 0.069 0.001 (0.060) (0.061) (0.059) (0.003) Ever treated before approval 0.051 (0.044) Placebo treatment date t−10.071 (0.054) Placebo treatment date t−20.027 (0.049) Placebo treatment date t−30.014 (0.051) Placebo treatment date t≤40.068 (0.048) Observations 4,554 4,554 3,726 4,554 Number of CS 207 207 207 207 R-squared 0.178 0.179 0.409 - Communal section FE YES YES YES YES Year FE YES YES YES YES Satellite FE YES YES YES YES Contemporaneous covariates YES YES YES YES Lagged covariates t−1t−1t−1t−1 P-value 0.884 SE clustered at the communal section level between parentheses. *p<0.10; **p<0.05; ***p<0.01. Treatment represents the coefficient β1from the regression following equation (1). Column (4) indicates a random allocation of the timing of the treatments. 36