Policy Research Working Paper 9052 Public Disclosure Authorized
Can We Rely on VIIRS Nightlights to Estimate the Short-Term Impacts of Natural Disasters?
Public Disclosure Authorized Evidence from Five Southeast Asian Countries
Emmanuel Skoufias Eric Strobl Thomas Tveit Public Disclosure Authorized Public Disclosure Authorized
Poverty and Equity Global Practice October 2019 Policy Research Working Paper 9052
Abstract Visible Infrared Imaging Radiometer Suite (VIIRS) night- studies of specific disasters, and (ii) fixed effect regression lights are used to model damage caused by earthquakes, models akin to the double difference method to determine floods, and typhoons in five Southeast Asian countries any effect that the different natural hazards might have had (Indonesia, Myanmar, the Philippines, Thailand, and Viet- on the nightlight value. The results show little to no signifi- nam). The data are used to examine the extent to which for cance regardless of the methodology used, most likely due each type of hazard there is a difference in nightlight inten- to noise in the nightlight data and the fact that the tropics sity between affected and nonaffected cells based on (i) case have only a few days per month with no cloud cover.
This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at [email protected].
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
Can We Rely on VIIRS Nightlights to Estimate the Short‐Term Impacts of Natural Disasters?
Evidence from Five Southeast Asian Countries
Emmanuel Skoufias* Eric Strobl** Thomas Tveit**
JEL Classification: Q54, C63, R11, R5,O18 Keywords: Remotely Sensed Data, Natural Disasters, Damage Index, Floods, Earthquakes, Typhoons
*The World Bank Group, Washington, DC USA ** University of Bern, Switzerland
Introduction
Natural hazards and the damages caused by them lead to headlines in the news now more than ever. Part of the reason for the increased reporting is due to the focus on climate, but it is also caused by larger economic damages than in prior years. The latter point is due to the growing global population and the increasing economic value in settled places. With this value increase comes a need for updated local data on where people live and the economic activity there. This is particularly true in the wake of a devastating natural hazard, when it can be difficult to assess damages quickly and accurately. A recent method to assess damages and economic activity is to use remotely sensed data. As a proxy for economic activity one has seen widespread use of nightlights (Henderson, Storeygard and Weil 2012) (Gillespie, et al. 2014) (Skoufias, Strobl and Tveit 2017) (Hodler and Raschky 2014) (Michalopoulos and Papaioannou 2014) (World Bank 2017). The modeling and measurement of natural hazards has also been extensively covered in the literature for different disaster types such as earthquakes (De Groeve, et al. 2008), typhoons (Emanuel 2011) and floods (Wu, Adler and Tian, et al. 2014) to name a few. The aim of this paper is to provide a refined model of a previous study (Skoufias, Strobl and Tveit 2017) where nightlights were used as a proxy for economic activity and then combined with damage indices to find the economic impact. The novelty is that a new, more spatially and temporally detailed data set for nightlights, the Visible Infrared Imaging Radiometer Suite (VIIRS) product from the National Oceanic and Atmospheric Administration (NOAA), following the work of Elvidge et al. (2017), is employed. In addition, settlement layers have been used to remove nightlight cells with zero population to reduce the impact that “empty” cells could have on the analysis. Five countries in Southeast Asia have been chosen as case studies due to their general lack of local level economic data and prevalence of different natural hazards: Indonesia, Myanmar, the Philippines, Thailand and Vietnam. These countries represent a large portion of the total GDP, population and area of the region. In addition, they experience earthquakes, typhoons and floods to a varying degree. The VIIRS data have a number of advantages over the widely used, but now discontinued, Defense Meteorological Satellite Program (DMSP). Firstly, they have a higher resolution, 450m by 450m compared to 1km by 1km for the DMSP. Furthermore, the VIIRS data are released monthly compared to the annual DMSP data, making it possible to compare month‐ on‐month effects of natural disasters. Finally, the DMSP data are normalized and capped at 63, whereas the VIIRS data have no upper limit. However, the VIIRS data exhibit a lot more noise than what is found in the DMSP data. Part of this is due to stray light corrections that can lead to negative light values for some months. Another issue, prevalent in both data sets, occurs due to the many days with cloud cover that happens in the tropics. As a matter of fact, several months have no days with a measurement. Importantly, both the DMSP and the VIIRS data have shown strong correlation with local GDP (World Bank 2017) (Henderson, Storeygard
2 and Weil 2012) and this paper will take advantage of this to use VIIRS as a proxy for economic activity. The settlement layers used in the analysis are from WorldPop (WorldPop 2013) (Worldpop 2016) and CIESIN (Facebook Connectivity Lab and Center for International Earth Science Information Network ‐ CIESIN ‐ Columbia University n.d.). Both sets are high resolution layers that show where people live through a combination of satellite images and census data. There are two reasons for including the settlement layers. Firstly, the data are used as a way to clean the VIIRS data by excluding cells that are unlikely to contain economic activity and population. The use of secondary data to improve recognition of human activity has previously been done for poverty (Jean, et al. 2016) and for urban markets (Baragwanath, et al. 2019). Secondly, the population data are used to distinguish between rural and urban areas. For the disaster modeling, we model three disaster types, earthquakes, typhoons and floods, by following closely the methodology from Skoufias, Strobl and Tveit (2017) for earthquakes and floods and Strobl (2012) for typhoons. These damage indices are then combined with the other data sets and two types of analysis are performed on this combined set. The first type is one case study for each hazard type. The choice fell on typhoon Haiyan, which hit the Philippines in 2013, the 2016 Aceh Earthquake and the 2017 floods in Southern Thailand. All three disasters caused significant local damage and were for the most part spatially and temporally separate from other disasters. The case studies consist of a simple graph showing the difference in mean nightlight values for the nightlight cells that were hit by the hazard and the cells that were not hit over a period starting 12 months before the hazard and until 12 months after. The second analysis moves the focus from specific events to a national level and hazards that occurred between June 2012 and March 2018. To find any short and medium‐term impacts that each disaster type has, a fixed effect regression with 12 monthly lags is performed. Overall, the results are not particularly significant for either analysis. Two case studies of typhoons in the Philippines exhibited the strongest results, where one found a clear drop in mean nightlight during the month that Haiyan struck and similarly a clear drop when Rammasun struck in July 2014. The 2017 floods in Thailand also exhibited a drop, but compared to noise in other months it was rather small, whereas the 2016 Aceh Earthquake exhibited no difference between cells hit by it versus those that were not affected. Part of the reason for these results could be that an influx of aid and personnel would lead to a net effect of zero, but given the noise that is exhibited overall in the VIIRS data it would be hard to draw any more confident conclusions in this regard. The issue of noise in the underlying monthly data has been well illustrated in Hu and Yao (2019), where the monthly VIIRS data showed much lower coorelation with GDP than annual VIIRS and DMSP data.1 The fixed effects regressions were also inconclusive, as there was little if any significance across disaster types and months. Several methods were tried to correct for the noise, such as dropping any negative values, setting any value below a certain threshold to zero (both for
1 The annual VIIRS data were not used in this study, as they are currently only available for 2 years
3 absolute and real values), running them only for cells with populations above a certain threshold as well as logaritmic hyberbolic sine transformations. However, the results did not improve or change. There could be several reasons for these results. The most likely being that the VIIRS data are noisy by nature and, due to the countries being close to the equator, many areas are often covered by clouds, reducing the quality of the measurements. The VIIRS data also consist of much noise near zero, meaning that a transformation or cleaning of the data was necessary, although perhaps not the best one was used. Another possibility is that the methodology or measurements for the different hazard types are inaccurate. Finally, the choice of regression model might also cause estimation problems. Overall, different methods were employed to alleviate these concerns, but none yielded the expected results. The remainder of the paper starts with an overview of the countries and the base data, followed by a general methodology overview, the three disaster types with results and finally a conclusion.
I. Overview, Common Data and Summary Statistics
A. Countries
We have chosen five different Southeast Asian countries: Indonesia, Myanmar, the Philippines, Thailand and Vietnam. These five countries have been chosen for several reasons. Firstly, they represent roughly 75 percent of the Southeast Asian GDP,2 providing a good representation of the region as a whole. Secondly, they constitute over 80% of the total landmass and 90% of the total population in Southeast Asia. Finally, they all represent different economies and their economies are affected differently by natural hazards.
2 According to the World Bank Group Research https://data.worldbank.org/indicator/ny.gdp.mktp.cd.
4 Figure 1: Countries in Southeast Asia
Indonesia. The largest country in terms of GDP, area and people is Indonesia. It has more than 40 percent of the total population in Southeast Asia, with more than 260 million inhabitants. It is also the country with the highest number of natural disasters. Geographically Indonesia is mostly made up of large and – in some cases – densely populated islands.
According to the Indonesian National Disaster Management Authority (BNPB), there were more than 19,000 natural hazards in the period 2001 ‐ 2015 (National Disaster Management Agency 2016), making Indonesia a useful country for any natural hazard analysis. The most frequent disasters are floods and landslides (52 percent), strong winds (21 percent) and fires (15 percent), while the most damaging ones are earthquakes, tsunamis and volcanic eruptions, which all cause major damage to buildings and infrastructure in addition to the human casualties.
The tropical climate of Indonesia often leads to annual floods. The BNPB data registered more than 10,000 incidents of floods or landslides leading to more than 3,500 fatalities from 2001
5 through 2015. During the period from 1985 to 2016, the Dartmouth Flood Observatory (DFO) registered 3,808 floods of magnitude 4 or more and 1,175 floods of magnitude 6 and up.3
Myanmar. The second country used in our analysis is Myanmar. It has the smallest GDP and population of our five countries, while at the same time being the second largest. During the period 2006‐2016, the UN Office for the Coordination of Humanitarian Affairs (OCHA) mentions several major natural hazards affecting Myanmar (OCHA 2016). There were three large earthquakes in 2011, 2012 and 2016, affecting more than 30,000 people and killing 100. Furthermore, a major flood occurred in 2015, displacing 1.7 million people in Western Myanmar and killing 172 people. Finally, more than 2.5 million people were affected by strong cyclones in 2008, 2010 and 2013.
Philippines. The Philippines is – like Indonesia – a country consisting of thousands of islands ranging in size from tiny and deserted to large and densely populated. It is the second smallest country in our sample, with the second highest population just in excess of 100 million. It has the third largest GDP. According to the Global Facility for Disaster Reduction and Recovery (GFDRR), the Philippines is at high risk from several types of natural hazards (GFDRR 2019). Prime among them are cyclones, where an average of 20 make landfall every year. In 2013, typhoon Yolanda led to 6,000 casualties and damaged more than 1.1 million houses. The Philippines are also exposed to earthquake and flood risks.
Thailand. As for Thailand, it is the third largest country, fourth most populous and boasts the second largest GDP of our countries. In terms of disasters, Thailand is mostly at risk to floods and typhoons, but earthquakes happen occasionally, with the largest recorded being a 6.1 magnitude earthquake that occurred in May 2014. The lower risk compared to our other countries is also mentioned by INFORM in their country profile of Thailand (INFORM 2019).
Vietnam. The final country is Vietnam, which is the smallest, but has the third highest population just below 100 million people. In terms of GDP, Vietnam rank fourth among our five countries. When it comes to disaster risk, they are on par with Thailand in that INFORM ranks them low and exposed primarily to floods and typhoons (INFORM 2019).
B. VIIRS
To find and identify areas with economic activity and their exposure to natural hazards, satellite images of nighttime lights will be used as a proxy. This is due to the highly localized nature of natural hazards and how they affect economic activity within populated areas. In an optimal scenario, one would prefer highly spatially disaggregated economic data, but lacking
3 Magnitude is defined as: M = log(D * S * AA), with D being the duration of the flood in days; S is the severity on a scale consisting of 1 (large event), 1.5 (very large event) and 2 (extreme event); and AA is the size of the affected area in square kilometers. Flood events registered by DFO have mainly been derived from news and governmental sources.
6 this, one can use nightlights. This methodology has been used in several papers already, with significant results (Henderson, Storeygard and Weil 2012) (Gillespie, et al. 2014) (Hodler and Raschky 2014) (Michalopoulos and Papaioannou 2014). In Henderson, Storeygard and Weil (2012), Indonesia is used as an example of using nightlights to capture an economic downturn following the Asian financial crisis in the late 1990s. Their results show that swings in GDP change can generally be captured. Nevertheless, one has to account for factors such as cultural differences in light usage, latitude and gas flares. In our case this is unlikely to affect our results since nightlights are used to capture exposure within a country rather than across countries. All the previously mentioned studies use a prior iteration of nightlight images, the DMSP satellite images. The approximate cell grid size of these images is 1 by 1 kilometer at the equator and the data provided are annual. This paper utilizes a newer nightlight data set, the VIIRS Day/Night Band (DNB) provided by The Earth Observations Group (EOG) at NOAA/NCEI. The data are produced following the methodology of Elvidge, et al. (2017). These data are provided starting in April 2012 and go till the present, making the time series much shorter than the DMSP data which start in 1992 and go through 2013. However, the VIIRS DNB images from EOG are monthly, whereas the DMSP data are annual. Furthermore, the images have 15 arcseconds grids (approximately 450 meters at the equator) compared to the 30 arcseconds of DMSP. The VIIRS coverage spans from 75N latitude to 65S around the entire globe, meaning that all of Southeast Asia is included. The VIIRS data have also seen usage in the economics and disaster literature. Firstly, Chen and Nordhaus’s (2015) analysis finds promising results for VIIRS as an economic and population indicator, also when comparing to the DMSP product. Secondly, Zhao, et al. (2018) used the underlying NPP‐VIIRS DNB Daily Data to analyze selected natural disasters. They found that the images were useful for detecting damages and power outages, but that cloud coverage was a major limitation in the assessment. More recent research utilizing the VIIRS nightlights also points to the limitation of monthly VIIRS in the detection of GDP and urban markets. Hu and Yao (2019) show how monthly VIIRS data have very low correlation for low income countries.4 The overall correlation is significantly better for middle income countries and annual, corrected data. Even when using the annual mean of monthly data, the results are worse than for annual VIIRS and DMSP. The data output consists of two variables: the average light radiance from DNB and the number of cloud free days. An important note regarding cloud free days is that for the tropical areas in Southeast Asia one will have months where there are no cloud free days and hence no radiance measurement. To account for months with no radiance value, this paper has interpolated between the month before and after. Furthermore, for cells with little light, one often finds negative light values due to airglow contamination (Uprety, et al. 2017). There is no established standard for how to interpret these measurements, but we did any analysis
4 If one includes all the low income countries, the correlation is negative (‐0.02), but it is heavily skewed by the poorest African nations. With the exclusion of the Central African Republic, the correlation improves to 0.12.
7 using different methodologies such as setting negative values to 0 or using absolute values with different thresholds for counting a value as 0. To provide some further details into the noise and distribution of the VIIRS data, Figures 2 and 3 show the distribution of VIIRS values below 25 for all countries and their populated areas and the number of cloud free days per month. Looking at Figure 2 first, one sees a clustering on both sides of 0, with the bin containing 0 and very small values constituting more than 15 percent. Furthermore, more than 13 percent of the total points have a negative value and 46 percent more have a value between 0 and 0.3, meaning that almost 60 percent of all points have negative or very small values. It is also worth noting again that these are only points that have been identified as being populated,6 meaning that they should be relatively free of disturbance from non‐human sources. Despite this, there are still very small deviations from 0. Even when limiting the light to cells with population over 50 (approximately 30 percent of populated cells), the distribution stays similar with 85 percent of light values being below 2, 75 percent below 1, 46 percent below 0.3 and 5 percent being negative. This distribution pattern follows through to high population numbers. When looking at population numbers above 1,000 (less than 12 percent of the total) 17 percent of the points still have nightlight values below 0.3 albeit only 0.3 of points have values below 0. The question will then be if VIIRS is only useful for very urbanized areas, if at all.
5 Values below 2 constitute approximately 95 percent of the total. This was done due to the maximum VIIRS value being 17,699 and hence a density graph would be meaningless going from ‐1.5 to 17,699. Also, any bin above 2 is very small and would not contribute to the graph. 6 See the next section for details on population layers.
8 Figure 1: VIIRS Value Distribution for VIIRS values below 2
Source: Authors’ estimates based on VIIRS and population layer data (see text for details)
Figure 3 provides an overview of the distribution of cloud free days. The first thing of note is that there are no points or months that have more than 20 cloud free days, meaning that no month has used more than 2/3 of the days to get a monthly mean value. The median is 4 days and the mean is 6, and in almost 11 percent of months there were no cloud free days. Overall, the implication is that the monthly values in reality consist of a rather small subsample of the monthly light. It might not matter much for a non‐climate related hazard such as earthquakes, but for a flood, which is highly correlated with clouds, it might mean that the mean value is based on values before or long after the event has occurred.
9 Figure 3: Days with No Cloud Cover in a Month
Source: Authors’ estimates based on VIIRS and population layer data (see text for details)
C. Population Layers
An important aspect of capturing economic activity through nightlights is to identify areas where there are people, hence avoiding stray lights or other sources of light that are not connected to human activity. For example, Baragwanath et al. (2019) used daily MODIS data to identify the extent of urban markets in India. To find these areas, two sources of human settlement layers are utilized. The WorldPop data sets have been used to identify settlement areas in Myanmar and Vietnam (WorldPop 2013) (Worldpop 2016). These data sets have a spatial grid cell resolution of approximately 100 meters at the equator and estimate the number of persons per square. Estimates are provided for 2010, 2015 and 2020, but for this paper we have only used the 2015 estimates. To construct the population files, national totals have been adjusted to match UN population estimates. To do this adjustment, a Random Forest model has been used to construct a weighted population density layer, which is used as a basis to place the population as closely as possible to its real geographical distribution.
10 For Indonesia, the Philippines and Thailand high resolution settlement layers from Facebook Connectivity Lab and Center for International Earth Science Information Network ‐ CIESIN ‐ Columbia University was used. These layers are produced at a spatial resolution of 1 arcsecond (~30 meters). The underlying data are based on 2015 and combines census data with satellite images from DigitalGlobe. The population is allocated according to subdivision censuses once settlements have been identified from the satellite images. In addition to identifying economic activity, the settlement layers will be used to identify rural and urban areas based upon population density.
II. General Empirical Strategy
The goal is to analyze the effect that natural hazards might have on the economic activity in an area. To try to tease out this effect, a data set was constructed for each natural hazard type. These sets contain localized nightlight values and damage indices for each hazard type, the total population in each light cell and administrative boundary data linking each nightlight cell to different administrative levels. The new data sets are used in two different sets of analysis. First, a simple case study on a singular large event is performed. The case studies are done on the Aceh earthquake in 2016, the Haiyan Typhoon and finally for the 2017 floods in Thailand. All of these events had severe damages in the affected areas and one would expect that the nightlight values would change significantly following these events. As for the actual analysis, it is a graph of nightlight means of cells in affected areas, split between cells that experienced damage versus those that did not. The second analysis is a fixed effects regression for each country and natural hazard type to see if any national effects exist. The regressions are lagged for the 12 months following the hazards to allow for any short‐ and mid‐term effects to materialize. The general methodology can be seen in Figure 4.
11 Figure 4: Flow chart for methodology
A. Constructing the Data Sets
The approach we have chosen is somewhat similar to what was done in Skoufias, Strobl and Tveit (2017), where the authors constructed damage indices, weighted with nightlights and aggregated up to annual values at a province or municipality level. However, the method this time also provides significant differences. Firstly, the nightlight data used as a proxy for economic activity is monthly – and not yearly – providing a better picture of how nightlights are affected by natural hazards in the short term. Furthermore, the spatial resolution is much higher, with the VIIRS data providing details down to 450 meter squares at the equator compared to 1 kilometer squares for the DMSP.
12 Secondly, by using high resolution population layers to identify areas with economic activity some noise from adjacent non‐economic areas is removed. Hopefully providing a better data set to pinpoint where a change in light value would be due to a change in economic output and not due to stray light or other light sources such as forest fires. One potential downside of this is that some agricultural land will not be included in the analysis. Finally, instead of aggregating up to a provincial or municipality level, the data are kept at a cell level. More specifically with regards to the methodology, one started with localized damage data modeled from actual intensity measures. These values were matched with any intersecting VIIRS cell and assigned to the corresponding month. For light cells that intersected several damage cells an average value was used. As earthquake damage estimates were modeled based on centroids, only the centroid intersection was used. Once the nightlight and damage data had been merged, population data was included. The population data sets were aggregated up to the same cell size as the VIIRS data and then matched to find the total population in each nightlight cell. Due to slight differences between the HRSL and WorldPop data sets, two different cutoffs for populated cells were utilized. As HRSL specifically provide settled areas and set other areas to having 0 population, any VIIRS cell with no settled areas in it, would be excluded from the final data set. As for the WorldPop data, any VIIRS cell with less than 5 people in it was cut. Finally, for potential future analysis, the VIIRS cells were assigned to administrative subdivisions such as provinces and municipalities. In cases where a cell crossed province borders, the province that contained the centroid was used.
B. Case Study Methodology
Once the data sets were constructed, one case study for each natural hazard type was chosen to graph the effect the hazard would have on the local nightlight values. The choice of hazards was based on reported damage and “fame”. In the case of floods, they were also picked based on how “isolated” they were, i.e. some of the largest events were lasting for months on end potentially confounding the effect of floods in different months. To check for any effect from the disasters, two graphs were constructed for each event; one with the mean of the nightlight values of cells that were struck by the event and one with the cells that were not affected. Furthermore, the analysis focused on the affected regions – state or province level ‐ and not the national averages.
C. Regression Methodology
The second set of analysis consists of regressions which were run to determine any effect that the different natural hazards might have had on the nightlight value. The methodology chosen were that of fixed effects, correcting for time and spatial effects. To correct for potential heteroscedasticity, Driscoll‐Kraay standard errors are used (Driscoll and Kraay 1998).
13 𝐿 , 𝛽 Σ 𝛽 𝐸𝐷 , , 𝜃 𝑒 ,
where 𝐿 , is the light level in cell i in month t and 𝐸𝐷 , , represents the vulnerability curve value in the same cell and at the same time. Lags are allowed from month t to t‐12. β is the intercept, θ are the cell fixed effects and 𝑒 , is the error term.
III. Earthquakes
To construct damage indices for earthquakes, this paper uses the same methodology as in Skoufias, Strobl and Tveit (2017), which in turn is based on modeling from contour maps generated by seismological ground stations (International and Regional Development 2001) (Agency 2006) (De Groeve, et al. 2008). In short, these contour maps are ShakeMaps from USGS, which are automatically generated maps providing several key parameters following an earthquake, such as peak ground acceleration (PGA), peak ground velocity (PGV) and modified Mercalli intensity (MMI), are used as a base for localized impact. More specifically, the ShakeMaps use data from seismic stations that is interpolated using an algorithm which is similar to kriging. To model the intensity in a given coordinate, the model also takes into account ground conditions and the depth of earthquake. The ShakeMaps are interpolated grids with point coordinates spaced approximately 1.5 kilometers apart (0.0167 degrees). With regards to the output data, PGA is a measure of the maximum horizontal ground acceleration as a percentage of gravity, PGV is the maximum horizontal ground speed in centimeters per second and MMI is the perceived intensity of the earthquake, a subjective measure. It is assumed that damage starts at an MMI level of V and a PGA of 3.9 percent of gravity (g). These damage levels are found for California in (Wald, et al. 1999), but the relationship has been found for other areas in the United States in Atkinson and Kaka (2006) and Atkinson and Kaka (2007) and for places such as Costa Rica (Linkimer 2007) and Japan, Southern Europe and the Western United States (Murphy and OBrien 1977). It is worth noting that the numerical relationship differs from region to region. The different measures are largely interchangeable, and in the International and Regional Development 2001 report, they use PGA to measure damage, pointing to the fact that PGA, unlike MMI is an objective measure, implying that MMI is not easy to obtain reliably across the globe. Also, for large scale modeling, where it is unfeasible to model local conditions precisely, PGA serves as a good proxy for intensity of earthquakes. For the actual construction of the damage index, two types of data will be used; the intensity data ‐ expressed as PGA ‐ and building inventory data. The building type data stems from the USGS building inventory for earthquake assessment, which provides estimates of the proportions (based on total number of buildings) of building types observed by country; see
14 Jaiswal and Wald (2008). The data provide the share of 99 different building types within a country separately for urban and rural areas. For Indonesia the building type information was compiled from a World Housing Encyclopedia (WHE) survey. The WHE survey uses fraction of population who lives or works in buildings of different types as their definition of how the building mass is split up. Without any other information available, we use this as an indication of the distribution of building types and mass, but, necessarily, assume that the distribution is homogenous within urban and rural areas. Damage curves by building type are derived from the curves constructed by the Global Earthquake Safety Initiative (GESI) project; see (International and Regional Development 2001). More specifically, buildings are first divided into 9 different types across a wide range of building materials from wood and steel through to lightweight shacks. Each building type itself is then rated according to the quality of the design, the quality of construction, and the quality of materials. Total quality is measured on a scale of zero to seven, depending on the total accumulated points from all three categories. According to the type of building and the total points acquired through the quality classification, each building is then assigned one of nine damage curves which provides estimates of the percentage of building damage for a set of 28 peak ground acceleration intervals. In order to use these damage curves, we first allocated each of the 99 building types given in the USGS building inventory to one of the 9 more aggregate categories of the GESI building classification. However, to assign a building type its quality‐specific damage curve we would further need to determine its quality in terms of design, construction, and materials, an aspect for which we unfortunately have no further information. We instead assume that building quality is homogenous across building type and construct eight different damage curves with different quality rating scenarios (ranging from 0 to 7). Initial analysis showed some effect based on the quality assumption, but lacking any meaningful information we decided to use a mean quality assumption of 4 across the countries. To model estimated damage due to a particular earthquake event the data from the ShakeMaps and GESI are used. Then, one identifies the value of peak ground acceleration that each nightlight cell experiences by matching each earthquake point with its nearest nightlight cell. If the cell is further away than 1.5 kilometers or if it experiences shaking (PGA) of less than 0.05g the value is set to 0. In order to derive a cell i specific earthquake damage index, ED, the following equation is applied: