Are Disaster Impact Estimates Distorted by Errors in Popular Night-Time Lights Data?

Abstract

Satellite-detected night lights data are widely used to evaluate economic impacts of disasters. Growing evidence from elsewhere in applied economics suggests that impact estimates are potentially distorted when popular Defense Meteorological Satellite Program (DMSP) night lights data are used. The low resolution DMSP sensor provides blurred and top-coded images compared to those from the newer and more precise Visible Infrared Imaging Radiometer Suite (VIIRS) images. Despite this, some disaster impact studies continue to use DMSP data, which have also been given a new lease of life through the use of harmonized series linking DMSP and VIIRS data to provide a longer sample. We examine whether use of DMSP data affects evaluations, using expected typhoon damages in the Philippines from 2012–19 as our case study. With DMSP data, negative impacts on economic activity from expected damages at the municipality level appear over 50% larger than when the VIIRS data are used. The DMSP data give the appearance of greater spatial autocorrelation in luminosity and through this channel may tend to spread apparent spatial impacts of shocks. Harmonized data that adjust VIIRS images to be like the DMSP data also have this spurious autocorrelation so researchers should be cautious in using these data for disaster assessments.

Introduction

Practitioners and researchers can now rapidly evaluate impacts of disasters and monitor the recovery by using satellite-detected earth observation data. Readily accessible night-time lights (NTL) data especially enable this research on disasters (Feeny et al. 2022; Nguyen and Noy 2020). Yet even as this literature expands, questions are raised elsewhere about using NTL data as a proxy for changes in local economic activity. These data seem to only weakly relate to changes in traditional measures, such as commune-level employment and household expenditures (Goldblatt et al. 2020) or county-level GDP (Zhang and Gibson 2022), even in countries where economic activity-luminosity relationships seem to strengthen as the data are spatially aggregated (Gibson et al. 2021). The caveats from these studies coming from outside of the disasters literature matter because it is changes in economic activity that are especially needed for estimating disaster impacts; hence, the cross-sectional relationships between luminosity and economic activity do not necessarily imply that there are equally useful relationships for measuring the changes in activity (Chen and Nordhaus 2019).

Moreover, it is especially in the low density rural areas of developing countries where relationships between NTL data and local economic activity are weak (Gibson et al. 2021). In such places there is a scarcity of alternative data for evaluating disaster impacts unlike the situation for urban areas. Yet luminosity is especially associated with economic activity in urban areas, and in recognition of this fact some researchers suggest that the NTL data may only be suitable for studying the impacts of (or recovery from) disasters that mainly affect urban areas (Akter 2023). If this caveat is agreed upon more widely it would tend to limit the social usefulness of the rapidly growing literature that estimates impacts of disasters.

One potential reason for why NTL data may be a poor proxy for changes in the local economic activity is that the most popular data are from the Defense Meteorological Satellite Program (DMSP), launched by the US Department of Defense in the 1960s to observe clouds for short-term weather forecasts. Hence, these satellites and sensors were not designed for long-term earth observation, nor for research purposes, and so findings based on these data are somewhat serendipitous. It should be no surprise that it is urban areas whose density and scale allows them to be detected from space even with fairly crude sensors and that the ability to detect changes in local economic activity may be limited. In contrast, newer, more accurate and precise luminosity measures are available since 2012 from the research-oriented Visible Infrared Imaging Radiometer Suite (VIIRS) that detects a far wider range of lighting and thus should provide a better basis for measuring changes in local economic activity.

It might be expected that research on the economic impact of disasters would quickly switch to using these newer VIIRS data. Yet studies on the impacts of disasters continue to be produced using DMSP data, as we show below.¹ Moreover, the DMSP data have been given a new lease of life by so-called ‘harmonized’ data where VIIRS data and DMSP data are linked to provide one long time-series. This harmonization essentially requires degrading the more precise VIIRS images to make them like the low-resolution DMSP data (Nechaev et al. 2021) and so, indirectly, some of the flaws in the DMSP data still enter into the post-2012 record even when the more precise and accurate VIIRS data are available.

Given the on-going use of the DMSP data in disasters research some evaluation of the sensitivity of estimated impacts to the use of different NTL data may be helpful. We therefore examine whether use of DMSP data affects disaster evaluations, using expected typhoon damages in the Philippines from 2012–19 as our case study. With the DMSP data, the municipal level (negative) impacts on economic activity appear over 50% larger than when the VIIRS data are used. The potential overstatement of the economic shocks is especially apparent with the spatial models that allow for spillovers; most likely this is because the DMSP data give the appearance of greater spatial autocorrelation in luminosity and through this channel they will tend to spread the apparent spatial impacts of shocks. Harmonized data that degrade VIIRS images to be like the DMSP data also have this spurious autocorrelation so researchers should be cautious in using these data for disaster assessments.

Our focus is on use of NTL data as a proxy for local economic activity in studies of disaster impacts. We pay far more attention to the left-hand side variable in our econometric models than to the key right-hand side variable, which is an expected damage function based on a wind speed model. This same hazard indicator is used in other studies of the economic growth impacts of hurricanes in the United States (Strobl 2011) and Central America (Strobl 2012; Ishizawa et al. 2019), typhoons in China (Elliott et al. 2015) and most pertinently, for typhoon impacts on local economic activity in the Philippines (Strobl 2019). However, there is evidence linking typhoon fatalities in the Philippines to intense rainfall rather wind damage (Yonson et al. 2018), albeit with vulnerability and exposure seeming more important than the destructive characteristics of the hazard itself. Our findings should apply to models that use alternative right-hand side variables, such as rainfall, because the errors we examine are from the left-hand side NTL variable that we (and many other studies) use although it would need to be confirmed with further study using alternative natural hazards indicators.²

In the next section we discuss the persistence in the use of DMSP data in the disasters literature. That section also serves as our literature review and highlights some of the main differences between DMSP and VIIRS. Background to our case study is given in Section III, mapping the patterns of luminosity and expected typhoon damages in the Philippines. The data on damages and luminosity, and the econometric methods are discussed in Section IV. The impact estimates are in Section V and Section VI has our conclusions.

Why are we still talking about DMSP in 2024?

There are at least three reasons for ongoing use of DMSP data in the disasters literature even 12 years after newer, more accurate and precise, VIIRS data became available. First, science advances slowly – especially for results published in economics journals – so some disasters with impact estimates now being published were from before VIIRS (pre-2012). The top panel of Table 1 has details on articles and working papers studying disaster impacts and using DMSP data, published from 2019 onwards (7 years after first availability of the VIIRS data). We found 16 papers on disasters using DMSP data that came out in the last five years. Some studied particular disasters, such as Hurricane Katrina, that were before availability of VIIRS data. But many studied multiple disasters within a particular sample period, especially the 1992–2013 period that corresponds to availability of the usual DMSP data. Such research designs could instead be applied to, say, the 2012–23 period with VIIRS data.

Table 1 a: Recent Papers Studying Disasters using DMSP data

Moreover, even papers studying pre-2012 disasters could use VIIRS data to examine possible biases from using DMSP data. First, one can examine the proportion of ‘false zeros’ from places where DMSP annual composites show no luminosity yet VIIRS shows positive luminosity in the same year. False zeros are a non-classical measurement error that can bias impact estimates (Kim et al. 2024). For example, about one-quarter of municipalities in the Philippines show zero luminosity in the DMSP annual composites; almost 3-times higher than what VIIRS shows for the same years. Growth over time in lighting makes true zeros increasingly less likely, so if DMSP lacks sensitivity to pick up newly illuminated but dimly lit places that VIIRS detects, false zeros should be more frequent post-2012. So a researcher studying a pre-2012 disaster could compare post-2012 VIIRS and DMSP data to flag false zeros and form an upper bound for possible bias in their impact estimates from this problem. Likewise, the more spatially precise VIIRS data can also be used to see if DMSP data give a misleading picture of the similarity in luminosity of nearby areas, which is something that we examine below. Our review found no studies with data exercises along these lines.

A second reason for ongoing using of DMSP data is that a researcher may be studying a post-2012 disaster but wants the longest possible time-series. For example, if using a difference-in-differences framework they likely want to test for parallel trends in the pre-disaster period and DMSP data allow one to go back 20 years earlier than VIIRS data allow. Yet, if this was a primary reason for on-going use of DMSP data we would expect to see discussion in these papers of the trade-offs faced when choosing a NTL data source – a more accurate one but with a shorter time-series versus a more error-prone one but with a longer time-series – and we did not find this sort of discussion in the papers we reviewed.

Another reason DMSP data remain relevant to disasters research in 2024 is more subtle, and relates to the growing use of so-called ‘harmonized’ data where the DMSP data and the VIIRS data are linked to create one long time-series from 1992 to the present. Several studies provide these harmonized data, with Li et al. (2020) perhaps the most widely cited, while Nechaev et al. (2021) is a very thorough example by the authors who are the most familiar with processing both DMSP and VIIRS images. We have listed some of the recent studies of disaster impacts that use these harmonized data in the bottom panel of Table 1. Challenges in trying to make consistent inferences from the inherently different VIIRS and DMSP data include:

• Differences in timing of earth observation alter the quantum and composition of measured luminosity. DMSP nominally observed earth between 8.30-10pm while VIIRS observes at 1.30am.³ Economic activities after midnight differ from those in the mid-evening.

• Large differences in spatial precision; at nadir the DMSP ground footprint is 25 km² but it is 45-times smaller for VIIRS at 0.55 km² (Elvidge et al. 2013).⁴ Moreover, the precision advantage for VIIRS is even larger off-nadir because DMSP does not adjust for angular viewing effects, especially blurring images at scan edges (Abrahams et al. 2018).

• Differences in coarseness of quantization; DMSP Digital Numbers (DN) theoretically go from 0-63 but non-zero values below five are rare due to bottom-coding, while top-coding (saturation) may affect pixels with DN values for annual composites as low as 55 (Bluhm and Krause 2022).⁵ In contrast to this approximate 50-point scale, the 14-bit quantization of VIIRS allows 16,384 different values and there are no saturation problems.

Given these differences, attempted at harmonization involves degrading the more precise VIIRS images, so some of the DMSP flaws enter into the post-2012 record even with more precise and accurate VIIRS data available. Nechaev et al. (2021) note that while it is possible to smooth (i.e., coarsen) high resolution VIIRS images there is no deterministic way to add detail to the coarse DMSP images. Likewise, it is easy to apply a censor threshold to VIIRS radiances to create a top-coded series but the reverse does not hold; there is nowhere to take brightness values that exceeded the DMSP saturation threshold in order to assign them to equivalent VIIRS values. Hence, it is best to think of the harmonization as inherently creating DMSP-like adjusted VIIRS data rather than VIIRS-like adjusted DMSP data.

Figure 1 shows some evidence of the way that the harmonized data can re-introduce DMSP flaws into post-2012 data. The figure has Moran’s I spatial autocorrelation statistics for night-time lights data for the Philippines in 2013.⁶ Intuitively, these statistics show the apparent similarity between an area and adjacent areas, with higher values showing greater similarity (up to a maximum of 1). Spatial autocorrelation is especially important for studying disasters such as typhoons, where the path of storm damage will create correlations between neighbouring spatial units. The left-hand panel has results for municipalities, which have a mean area of 184 km², while on the right are results for barangays, whose mean area is 7 km². With DMSP data the Moran’s I statistic is 0.43 for municipalities, slightly over double what VIIRS data show for the same year (the difference is statistically significant at the p < 0.01 level). In other words, DMSP overstates similarity of neighbouring areas, which is another perspective on the acknowledged pattern of DMSP data understating spatial inequality due to their spatially mean-reverting errors (Zhang et al. 2023; Mathen et al. 2024).⁷

Fig. 1

Moran’s I statistics for spatial autocorrelation in night-time lights data from the Philippines (annual composites for 2013) [Notes: DMSP is the Defense Meteorological Satellite Program, VNL is VIIRS night lights, DVNL is DMSP-like night time lights derived from VNL. The Moran statistics are calculated using a first-order contiguity spatial weighting matrix and the error bars show 95% confidence intervals.]

What happens when VIIRS data are adjusted to be DMSP-like? The third bar in each chart shows the Moran’s I statistic based on the DVNL harmonized data that are described by Nechaev et al. (2021). The adjustments to VIIRS data to make them DMSP-like increase the Moran’s I statistic by 77% in the municipality-level results, thereby largely undoing gains in spatial precision achieved by using VIIRS. In other words, the harmonized data re-introduce a distortion that affected the DMSP data. This same pattern is evident in the barangay results, although the baseline level of spatial autocorrelation is higher because a given output of light is being split into smaller areas so there naturally appears to be greater similarity between the adjacent small areas. If we carried on with this exercise all the way down to the (output) pixel scale of 30 arc-seconds used in some studies of disaster impacts (e.g. Strobl 2019) we likely would still see this pattern whereby, for the same set of on-the-ground lights, there is more apparent spatial autocorrelation after the spatially precise VIIRS data have been adjusted (degraded) to be like DMSP.

Why does overstated spatial autocorrelation matter to estimates of disaster impacts? By making adjacent areas look more similar, apparent impacts of a disaster may be spread more widely and this could potentially bias econometric estimates. The patterns in Fig. 1 also make it logically inconsistent to highlight the greatly improved spatial precision of VIIRS while using the DMSP-like VIIRS data (as Joseph (2022) does in a study of the Haiti earthquake) because the adjustments to create DMSP-like VIIRS data largely reverse these gains in spatial precision.

Patterns of Luminosity and of Typhoon Damages in the Philippines

We start with luminosity maps to highlight some key differences between DMSP and VIIRS. The map in the left panel of Fig. 2 shows the apparent location of night-time lights in the Philippines in 2013, according to DMSP annual composites. These same data are mapped as Fig. 1 in Strobl (2019), who repeats the commonly made claim that DMSP has a resolution of approximately one km near the equator. As noted above, this claim is ambiguous because it conflates sensor resolution, whose ground footprint is, at best, 25 km² at the nadir, with the scale of the grid on to which the data are allocated (Zhang et al. 2023). The map is dominated by a large illuminated area centred on the National Capital Region, with lit area that seems to extend north through much of Central Luzon and into the Ilocos region and south almost to the island of Mindoro. There are also large patches of lights in Central Visayas, centred on Cebu City, in Western Visayas, centred on Iloilo and Bacolod City, and in several parts of Mindanao, especially around Davao City, Cagayan de Oro, and Zamboanga City.

Fig. 2

Night-time lights in the Philippines (annual composites for 2013) according to DMSP (left) and VIIRS (right) [Notes: The DMSP annual composite is from satellite F18. The VIIRS night lights are the average masked series. The box in the image at right is the area mapped in Fig. 3.]

The broad picture provided by DMSP is quite different to what is seen in the VIIRS annual composite for 2013, in the map on the right of Fig. 2. According to the VIIRS data, lights are much more concentrated within the National Capital Region, within Metro Cebu but not spreading over the rest of the island, within Davao City, and in one or two other places. In other words, the DMSP images diffuse the concentrated sources of light to give an exaggerated sense of the scale of brightly lit areas. The remote sensing literature has known for over two decades that the DMSP images may exaggerate city boundaries by a factor of up to ten (Henderson et al. 2003), in a pattern often described as ‘blooming’ but which is more accurately termed as ‘blurring’ (Abrahams et al. 2018). This blurring is due to inherent flaws in the DMSP sensor and data management, including pixel aggregation (smoothing) that is done to conserve memory, geolocation errors, and uncorrected angular viewing effects where the expanded ground footprint far away from the nadir overstates the light coming from particular pixels. These features of DMSP have previously been shown to exaggerate the illuminated area of towns and cities in other countries (Gibson 2021; Zhang et al. 2023).⁸

The Philippines-wide maps in Fig. 2 do not allow a full discussion of some of the flaws in DMSP because the scale is too large. So in Fig. 3 we zoom into part of the Ilocos and Cordillera regions of northern Luzon, covering about six percent of the Philippines land area.⁹ Once again, the maps have the DMSP image on the left and the VIIRS image on the right, for annual composites for 2013. There are three points of interest (labelled as A, B, and C on the VIIRS map) that help show the potential distortions that may affect the DMSP data. The specific distortions we examine are examples of the general distortions that underlie the exaggerated spatial autocorrelation in DMSP data that was shown in Fig. 1.

Fig. 3

Night-time lights in a selected part of northern Luzon (annual composites for 2013) according to DMSP (left) and VIIRS (right) [Notes: The area mapped is the box in the image at right of Fig. 2. The A, B, and C in the figure at right refer to points of interest noted in the text. For other notes, please see Fig. 2.]

The first point of interest (A) is San Fernando, the capital city of La Union province. The administrative boundaries of this city enclose an area of 100 km² although much of that is forested hillsides, with only about one-third as built-up area. The VIIRS image reflects this pattern, with clearly illuminated pixels covering about 30 km² and a clear demarcation from the surrounding hinterland. In contrast, the DMSP image makes the lit area appear far larger, extending about 15 km inland and about 50 km north and south along the coast. The second point of interest (B) is the mountain city of Baguio, that was extensively damaged in the 1990 Luzon earthquake. This city with a population of about 370,000 has a highly urbanized area of about 60 km², and VIIRS images are consistent with this; a rectangle of roughly 10 km in the north–south direction and 8 km in the east–west direction would cover not just lights from the city but also parts of the adjacent municipalities of La Trinidad and Tublay. In contrast, in the blurred DMSP image Baguio seems to have a highly illuminated area of about 500 km² and so city area is overstated by a factor greater than six. The third point of interest is in the Lingayen Gulf (C), just offshore from Binmaley Beach where DMSP images make it seem that lights extend for about five kilometres over the water – while it is true that in parts of the Philippines houses are over water on stilts, photographs of this area show long sandy beaches with no houses. Instead it is due to the distortion from the DMSP blurring that makes it seem that lights are coming from over the water.

Why are these features of DMSP data relevant to the estimation of disaster impacts? Consider a shock that affects a city, such as a typhoon hitting San Fernando or an earthquake damaging Baguio. The direct effect of the shock may reduce luminosity because of damage to electricity generation and transmission infrastructure, or damage to buildings causing a fall in demand for lighting. Indirect effects may then result from harm to livelihoods, creating an economic shock that also leads to lower demand for lighting. If these effects last for a period of weeks or months the annual composites would be likely to show a reduction in luminosity. However there may be uncertainty about the spatial extent of this shock, especially for the indirect effects. Hence, if lights emanating from a concentrated urban area are diffused in the blurred DMSP images, any dimming of city lights in the annual composite will also show up as a dimming in the adjacent areas. Of course these areas may truly be affected by any of the economic channels that link them to the city (potentially positively if activity displaces to undamaged hinterland or negatively if there is a fall in the demand for hinterland labour and output), so with blurred DMSP data it may be quite difficult to get an accurate estimate of the impacts. In contrast, the more spatially precise VIIRS data should allow these impacts to be observed just where they occur and not in adjacent areas if they truly are unaffected.

It will require formal econometric evidence, in Section V below, to ascertain whether features of DMSP data illustrated in Figs. 2 and 3 cause detectable bias when the data are aggregated to the municipality level. There is evidence from elsewhere at similar spatial scale that these flaws in DMSP data do matter. Specifically, Kim et al. (2024) study the closure of an industrial zone in North Korea located in a municipality whose area is 180 km² (so similar to the mean area of municipalities in the Philippines) divided into two districts. The district with the industrial zone was at the 99th percentile of North Korea’s luminosity distribution in both VIIRS and DMSP (prior to the closure) but there was a big gap between where the two data sources ranked the other district; at the 29th percentile in VIIRS and the 79th percentile in DMSP. In other words, the luminosity gap between the two districts was greatly attenuated in DMSP data (seeming to be only 20 percentiles rather than 70 percentiles) because the blurred DMSP images attributed some lights from the industrial zone to the adjacent district. This flaw in these data carried over into difference-in-differences evaluation of the impact of the closure; VIIRS data yielded a precisely estimated 50% fall in luminosity after closure of the industrial zone (using the other district as the control group) while DMSP data implausibly implied that closing the industrial zone caused luminosity to increase due to some of the fall in luminosity from the treated district being wrongly attributed to the comparison district.

An important aspect of studying these potential econometric biases is to use models that allow for the spatial autocorrelation that is a likely feature of natural hazards (whereas economic shocks, such as closing an industrial zone, can be highly localized). In Fig. 4 we show a damage index calculated at the municipality level from expected damage estimates for each year from 2012 to 2019, where these are derived from the storm track of typhoons.¹⁰ There are clear bands of expected damage, lying on an east-to-west orientation with a slight tilt towards the northwest; these bands are especially apparent in northern Luzon, but also in the Eastern Visayas through the Bicol Peninsula and up towards Manila. While Mindanao is typically less affected, there is one damage band going across Mindano and into the southern end of Negros island. With these clear spatial patterns, the usual assumption that regression errors for each observation are independent of errors for neighbouring observations may not be expected to hold. Moreover, in addition to the unobserved components (i.e. the errors) of changes in economic activity being spatially correlated there are also likely to be spillovers from the fact that municipalities are engaged in economic interactions with their neighbours. For these reasons, our modelling strategy starts with the most general spatial econometrics model that can allow for both local and global spillovers, and then imposes restrictions to consider various special cases, including the usual aspatial panel data model.

Fig. 4

Damage index derived from typhoon paths, 2012–2019

Data and Methods

In order to examine the sensitivity of estimated disaster impacts to the use of different NTL data we need at least two sources of data on night-time lights and also estimates of exposure to disasters. We use expected typhoon damages in the Philippines from 2012 to 2019 as our case study and primarily rely upon two sources of annual NTL data: DMSP stable lights and VIIRS night lights (VNL) version 2.1 masked average radiances.¹¹ Key references on how these annual composites of the night-time lights data are formed are Baugh et al. (2010) and Ghosh et al. (2021) for DMSP and Elvidge et al. (2017, 2021) for VIIRS.

The DMSP data are 6-bit digital numbers (DN) ranging from 0 to 63 (lower values for less brightly lit areas), for each 30 arc-second output grid (roughly 0.9 × 0.9 km at the latitude of the Philippines). In the annual composites, ‘stable’ simply means that ephemeral lights, from sources like fires and gas flaring, are removed before the annual composite is built from nightly images. An alternative meaning of ‘stable’, in terms of temporal consistency, does not apply as there is no in-built calibration of the DMSP data to adjust for the constant changes in sensor amplification (that are made without any record kept, where amplification changes are so cloud-tops can be viewed with similar brightness across light and dark parts of the lunar cycle). Consequently, it is best to think of the DN value as a relative measure of brightness, because the same DN value in different years could correspond to different radiance values (Doll 2008). While there are radiance-calibrated DMSP data (Elvidge et al. 1999), these are not available since 2010, and so cannot be used for comparing with the VIIRS data.

The VNL data are version 2.1 annual composites from 2012 to 2019, from Elvidge et al. (2021), derived from monthly cloud-free radiance averages from the Suomi/NPP satellite. These data undergo an initial filtering to remove extraneous features such as fires and aurora before the resulting rough annual composites are subjected to further outlier removal procedures. The lit grid cells are isolated from background noise using thresholds that apply across years, making these data better for change detection than earlier vintages of the VIIRS annual composites (that used year-specific thresholds). The data are in units of nano Watts per square centimeter per steradian (nW/cm²/sr) presented on a 15 arc-second output grid (which is equivalent to 0.47 × 0.45 km at the latitude of the Philippines).

Our expected damage index from typhoons is for all municipalities in the Philippines, over the period from 2012 to 2019. We use the same approach as described in Strobl (2019), except that rather than calculating at the municipality level Strobl used DMSP output grid cells (of 30 arc-seconds) as the spatial units and 1992–2013 as the study period. Our concern with a grid-cell level analysis is that DMSP sensor resolution is far coarser than the output grid (at best, having a ground footprint of 25 km² at nadir) and so the combination of a coarse sensor and a finer (resampled) output grid is likely to create spurious spatial autocorrelation in the luminosity estimates for grid cells. In contrast, with a mean area for the municipalities of 184 km² (which is far larger than the sensor resolution) this concern is mitigated.

The original source of data used to calculate the expected damage index is the Best Track Archive for Climate Stewardship (IBTrACS) provided by the National Oceanic and Atmospheric Administration.¹² It has details on the geospatial position, maximum sustained wind speed, radius of maximum winds, and other characteristics at 6-h frequency for each storm. The calculation involves two steps. First, a wind field model developed in Holland (1980) and updated in Boose et al. (2004) is used to estimate the wind speed experienced in a location at a particular time. It is a function of maximum sustained wind velocity anywhere in the typhoon, the radius of maximum winds, the distance from the centre of the storm to the point of interest, the forward velocity of the storm, the clockwise angle between the forward path of the storm and a radial line from the storm centre to the point of interest, and the gust factor.

Second, we apply the damage function that is proposed in Emanuel (2011), which links the maximum wind speed from a storm experienced at point of interest, above a certain threshold, to the expected damage. Following Strobl (2019), we assume that maximum wind speed below 92 km/hr would cause no damage to the location. This step yields a damage index between 0 and 1. For a municipality that is hit by several storms in a single year, we take the maximum of the multiple estimates of expected damages. For more details on the formula, data sources and parameters used in our calculations, please refer to Appendix A.

Estimation Framework

Spatial econometric methods allow for non-randomness in space of phenomena and account for possible spillovers. A key element making these models feasible is that possible interactions between spatial units are summarized with a $N\times N$ spatial weights matrix, W. In this study we use a row normalized contiguity weights matrix with ones for neighbours and zero otherwise, with a diagonal of zeros because a municipality cannot neighbour itself. The average municipality in the Philippines has five neighbours.

In what follows, luminosity in municipality i in year t is denoted as ${L}_{it}$, the expected damages are ${D}_{it},$ the ${\mu }_{i}$ are time-invariant fixed effects for each municipality, the ${\vartheta }_{t}$ are year fixed effects, and ${e}_{it}$ is a random error. By using the spatial weights matrix we can allow for spatial lags, which are averages of these variables over the neighbouring units. Thus, with the addition of a few extra variables (the spatial lags) it becomes possible to summarize effects of many complicated interactions (if each potential spillover was to be separately estimated the number of parameters would exceed the number of observations).

Our starting point is the very general spatial autoregressive model with spatial autoregressive errors (SARAR).¹³ This model allows for spatially lagged dependent variables, spatially lagged independent variables, and spatially lagged errors:

$${L}_{it}=\uplambda {WL}_{it}+{{\beta }_{1}D}_{it}+{{\beta }_{2}WD}_{it}+{\mu }_{i}+{\vartheta }_{t}+{\rho Wu}_{it}+{e}_{it}$$

(1)

The SARAR model allows for changes in an outcome variable in a given area to have effects on contemporaneous outcomes in other areas (via the autoregressive spatial lag of the dependent variable, if $\uplambda \ne 0$). It also allows changes in independent variables to affect not only own-area outcomes but also outcomes in neighbouring areas (if ${\beta }_{2}\ne 0)$. The ${\rho Wu}_{it}$ term allows for spatial autocorrelation, where errors for a given municipality correlate (ρ) with a weighted average of the errors from surrounding municipalities. Equation (1) nests a spatial Durbin model if $\rho =0,$ a spatial auto-regressive model (aka spatial lag model) where only the dependent variable is spatially lagged if ${\beta }_{2}=\rho =0$, a spatial error model where only the errors are spatially lagged (if λ $={\beta }_{2}=0)$, and the most restrictive of all, which is an aspatial model with no spatial lags (if ${\uplambda =\beta }_{2}=\rho =0)$. Another feature of Eq. (1) is that including municipality fixed effects provides a ‘within’ estimator interpretation, which is the appropriate framework for considering how changes in luminosity relate to the expected damages index (Zhang and Gibson 2022).

An important feature of spatial econometric models is that lags of either the outcome variable or of independent variables (but not of errors) mean that total effects of changes in an independent variable—e.g. higher expected damages—may be quite different to what the usual regression coefficient shows. While ${\widehat{\beta }}_{1}$ is the object of interest in the typical model without spatial lags, in the spatial models when either the spatial lags of outcomes or the spatial lags of independent variables are non-zero then ${\widehat{\beta }}_{1}$ does not capture the total effect of a change in the independent variable. A useful decomposition of the more complex spatial relationships that occur relies on rewriting Eq. (1) in matrix notation (for simplicity, subscripts are dropped and fixed effects and error terms combined in v because the errors do not affect this decomposition) as:

$$L={\left(I-\uplambda W\right)}^{-1}\left(D{\beta }_{1}+WD{\beta }_{2}\right)+{\left(I-\uplambda W\right)}^{-1}v$$

(2)

Following Elhorst (2012), the $N\times N$ matrix of partial derivatives can be written (noting that diagonal elements of W are zero) as:

$$\frac{\partial L}{{\partial D}_{k}}={\left(I-\uplambda W\right)}^{-1}\left({\beta }_{1k}{I}_{N}+{\beta }_{2k}W\right)$$

(3)

where D_k is the damage index in municipality k. The total marginal effect on luminosity from expected damages has two components, a direct one and an indirect one, that may both vary over space. The estimator that we use follows LeSage and Pace (2009) in reporting a single direct effect, that averages the diagonal elements of the matrix in (3) and a single indirect effect that averages the row sums of the non-diagonal elements of that matrix. Indirect effects arise not just from adjacent area units, if ${\beta }_{2k}\ne 0,$ but also from (potentially) all areas through the spatial autoregressive effect if $\uplambda \ne 0$. Thus, there can be both local and global spillovers and when these are accounted for, averages from the matrix of derivatives $\partial L/{\partial D}_{k}$ may be quite different to the estimated direct impact effect, ${\widehat{\beta }}_{1}$.

Results

In order to the use the data described in Section IV some transformations are needed because the units of DMSP (digital numbers) and of VNL (radiances) are not comparable. We take the logarithm of the sum of lights in each municipality in each year, using the inverse-hyperbolic sine (IHS) transformation to allow for the inclusion of any zeros.¹⁴ The impact estimates can then be interpreted as approximate percentage changes in economic activity (to the extent that this change is proxied by changes in luminosity) for the damages index changing from 0 to 1. The descriptive statistics in Table 2 show that such a movement is more than ten standard deviations, if all municipalities are considered, and nine standard deviations if attention is restricted just to the municipalities (47% of the total) where expected damages were greater than zero over 2012–19. Hence, a more useful interpretation may be to consider one-tenth of the estimated impact, corresponding to an approximately one standard deviation change in the damage index. The descriptive statistics also show that the municipalities that ever faced typhoon damage are more economically active, with about 15% higher luminosity than the never damaged municipalities.

Table 2 Descriptive statistics for expected damages and log luminosity

The results of estimating Eq. (1) and then imposing the various parameter restrictions and estimating the nested models are given in Table 3 (for DMSP) and Table 4 (for VNL). The nesting restrictions are rejected in all cases, so that the SARAR models appear to be the most data-acceptable models in both cases. In other words, the spatial Durbin model, the spatial lag model, the spatial error model, and the aspatial model that does not have any lags all rely on restrictions that are rejected by the data. The discussion therefore concentrates mostly on the results in column (1) for the SARAR model, which allows for spatial lags of the outcome (luminosity as a proxy for economic activity), of the independent treatment (the expected typhoon damages), and of the errors. Some attention is also paid to the column (5) results, for the typical panel data model with area and time period fixed effects that does not have any spatial lags because models of this type are frequently found in the literature using NTL data to estimate disaster impacts.

Table 3 Typhoon damage effects on luminosity (log DMSP): Municipalities in the Philippines, 2012 to 2019

Table 4 Typhoon damage effects on luminosity (log VIIRS night lights): Municipalities in the Philippines, 2012 to 2019

Given the statistical relevance of the spatial lags of the outcomes and of the treatment, it is important to note that the usual regression coefficients do not tell the full story and so the matrices of marginal effects based on Eq. (3) need to be used. The results of these marginal effects calculations are reported in the “average impacts” rows of Tables 3 and 4, using the decomposition due to LeSage and Pace (2009). As a preliminary comparison, the “total” row of Table 3 that is based on using the DMSP NTL data is almost 70% larger than the corresponding value in Table 4 using VNL data, for the SARAR model that is the most data acceptable of the five models. This same pattern, of larger negative average total impacts when using DMSP data, holds across all five models and so even if a researcher approached these data without using the most general spatial econometric models they would find that the DMSP data was yielding apparently larger negative impacts from the typhoon damages.

This difference in the apparent impacts is more plausibly attributed to overstatement when using the DMSP data rather than understatement when using VNL data, although we cannot be too firm in this claim due to the lack an independent benchmark estimate of actual changes in economic activity.¹⁵ Nevertheless, the VIIRS sensors have better properties and the prior evidence (including in settings with benchmarks such as county-level GDP) also is consistent with the claim that the VNL data are more accurate and precise and give results that are closer to the truth (Zhang and Gibson 2022; Zhang et al. 2023). The exaggerated spatial autocorrelation shown by the DMSP data in Fig. 1 provides one mechanism for the spatial impacts of a shock to apparently spread more widely. In this regard, it is especially notable that the indirect impacts exceed the direct impacts, except when constrained models (as in column (3)) are used that a priori rule out including spatial lags of the damages index. The map of the damages index patterns in Fig. 4 suggests that these spatial lags are likely to be important, given that they are based on the typhoon tracks, and so for disasters with this type of damages pattern, the exaggerated spatial autocorrelation in luminosity seen with the DMSP data may prove to be especially distorting to the impact estimates.

In terms of the magnitude of the estimated impacts, using the results in Table 4, for a standard deviation increase in the expected damages index, VNL luminosity will decline by about four percent according to the SARAR model. If one used the aspatial model, the fall in luminosity would be about three percent. While the first of these estimates allows for spatial spillovers neither of them allow for temporal lags. In the previous results for the Philippines reported by Strobl (2019) the temporal lags were all imprecisely estimated and only the contemporaneous effects on luminosity in the same year as the expected damages showed up as statistically significant.¹⁶ We therefore have concentrated on spatial effects, which are especially salient in light of the exaggerated spatial autocorrelation shown with the DMSP data (and with VIIRS data that have been adjusted to be like DMSP).

Discussion and Conclusions

In this study we have provided some evidence on the sensitivity of estimated disaster impacts to the use of different night-time lights data. Although the newer, more accurate and precise, VIIRS data have been available for over a decade, there is still some ongoing use of the older and less suitable DMSP data in the disasters literature. Moreover, the attempt to create longer time-series with various ‘harmonized’ data that splice together DMSP data and VIIRS data also makes some of the features (or flaws) of DMSP data relevant in the current era because these harmonized data inherently involved degrading the more precise VIIRS data. We show one aspect of this by using spatial autocorrelation statistics where the harmonized VIIRS data are very different than the original VNL data (with a Moran’s I statistic that is 77% higher), and are much more like DMSP data in showing a high degree of similarity of adjacent areas. In other words, the harmonized data give the appearance of greater spatial autocorrelation.

In our case study of typhoons in the Philippines over the 2012–19 period we found that the estimated negative impacts on luminosity were larger when using DMSP data than when using VNL data for the same period. These damages were especially expressed through local spillovers from spatial lags of the covariates, which is consistent with damages that are derived from the storm path of typhoons. In this case, the tendency of DMSP data to overstate the degree of spatial autocorrelation, which is the flipside of these data understating local inequality, seems to cause estimated impacts to spread. We also use imagery from selected areas of the Philippines which provides examples of this blurring of lights from the highly concentrated sources such as cities, and these blurring patterns have also been shown in other countries. While more evidence is needed, to the extent that it is possible to study disasters in other settings for years where there is overlapping availability of different sources of night-time lights data, it would be useful to repeat the comparisons carried out here.

We focussed on the blurring and spatial autocorrelation aspects of night-time lights data because damages from natural hazards like typhoons are often spatially autocorrelated. Other aspects of different types of night-time lights data, including overpass time, also likely affect usefulness for impact evaluations. While DMSP is usually described as observing earth between 8.30-10pm, when much consumption occurs, unstable satellite orbits saw overpass times become progressively earlier as satellites aged (even as early as late afternoon). Hence, to extend the DMSP time-series, Ghosh et al. (2021) resorted to using data from overpasses that were 12 h earlier, effectively giving a reading at about 3.30am—so the same activity that is observed by VIIRS at about 1.30am should be seen. Thus, our comparisons should not be affected by time-of-observation issues. However, a future study might use data from new satellites, such as China’s SDGSAT-1 (https://sdg.casearth.cn/en/datas/SDGSAT) that have a more favourable overpass time of ca. 9pm and have spatial resolution that is about 20-times finer than for VIIRS.

Of course our comparisons are less informative than is ideal because we do not have a true measure of on-the-ground damages. We also lack conventional subnational economic data, such as GDP, at municipality level in the Philippines. Two options for future study of these issues would be to either focus on a highly insured setting, where individual claims data might provide a measure of actual damages (but behaviour might differ from the less-insured settings of interest in developing countries), or on a setting with benchmark economic data at a sufficiently disaggregated level. For example, China has GDP estimates for county-level units (the third sub-national level), and for a province like Guangdong that is frequently hit by typhoons, third-level units are less than an order of magnitude larger than municipalities in the Philippines. However, these economic data are only available annually and in places that are able to recover rapidly from natural hazard damages an annual timescale may not be fine enough, and higher frequency daily or weekly data might be better.

Appendix A: Details on the Damage Index Calculations (Data and Parameters)

• 1. Data Definitions

Variable Name	Units	Description
Sid		Storm identifier
Year	year
Number		Cardinal number of the system for that season (year)
Basin		Western North Pacific
Name		Name provided by the agency
iso_time		Format is YYYY-MM-DD. Most points are in 6-h intervals and some are in 3 h
Nature		Combined storm type This is assigned based on all available storm types. They include: DS—Disturbance TS—Tropical ET—Extratropical SS—Subtropical NR—Not reported MX—Mixture (contradicting nature reports from different agencies)
Lat	Deg north	Latitude of storm center
Lon	Deg east	Longitude of storm center
dist2land	Km	Distance to land from the current position
Landfall	Km	Nearest location to land within next 6 h This can be thought of a landfall flag:  = 0 – Landfall within 6 h  > 0 – No landfall within next 6 h *Calculations are based on storm center (columns 9,10). Values less than 60 nmile likely are impacted by the system even though the center of the system is not over land
usa_wind	knots	Maximum sustained wind speed in knots: 0—300 kts
usa_pres	mb	Minimum sea level pressure, 850—1050 mb
usa_sshs		Saffir-Simpson Hurricane Scale information based on the wind speed provided by the US agency wind speed (US agencies provide 1-min wind speeds) -5 = Unknown -4 = Post-tropical -3 = Miscellaneous disturbances -2 = Subtropical Tropical systems classified based on wind speeds -1 = Tropical depression (W < 34) 0 = Tropical storm [34 < W < 64] 1 = Category 1 [64 <  = W < 83] 2 = Category 2 [83 <  = W < 96] 3 = Category 3 [96 <  = W < 113] 4 = Category 4 [113 <  = W < 137] 5 = Category 5 [W >  = 137]
usa_poci	mb	pressure in millibars of the last closed isobar, 900—1050 mb (not reanalyzed or not Best Tracked)
usa_roci	N. miles	radius of the last closed isobar, 0—999 n miles (not reanalyzed or not Best Tracked)
usa_rmw	N. miles	radius of max winds, 0—999 n miles (not reanalyzed or not Best Tracked)
tokyo_r50_long	N. miles	The longest radius of 50kt winds or greater
tokyo_r50_short	N. miles	The shortest radius of 50kt winds or greater
tokyo_r30_long	N. miles	The longest radius of 30kt winds or greater
tokyo_r30_short	N. miles	The shortest radius of 30kt winds or greater
storm_speed	knots	Translation speed of the system as calculated from the positions in LAT and LON
storm_dir	degrees	Translation direction of the system as calculated from the positions in LAT and LON. Direction is moving toward the vector pointing in degrees east of north [range = 0–360 deg]

• 2. Wind Field Model

To estimate the local winds experienced by a municipality we estimate ${V}_{ijt}$ which is the wind experienced in point i (municipality) amidst typhoon j, at time t. The wind field model takes the following form:

$${V}_{ijt}=GF \left[{Vm}_{ jt}-S\left(1-sin\left({T}_{ijt}\right)\right)\frac{{Vh}_{jt}}{2}\right]{\left[{\left(\frac{{Rm}_{jt}}{{R}_{it}}\right)}^{{B}_{jt}} exp\left(1-{\left[\frac{{Rm}_{jt}}{{R}_{it}}\right]}^{{B}_{jt}}\right)\right]}^{1/2}$$

Our data fills in some of the necessary parameters. However, there are some parameters that will be assumed (based on previous literature) or estimated.

Data var	Parameter	Description
	Vijt	wind experienced at time t due to typhoon j at any point (P = i)
usa_wind	Vm	maximum wind speed
storm_dir*	T	clockwise angle between the typhoon’s forward path and radial line between typhoon center and point of interest (P = i)
storm_speed	Vh	forward velocity
usa_rmw	Rm	radius of maximum winds
estimated	R	radial distance from typhoon center to point (P = i)
Assumed	G	gust factor
Assumed	F	surface friction
Assumed	S	asymmetry due to forward motion of the storm
Estimated	B	shape of wind profile curve

How each factor will be estimated or assumed:

• R and T is calculated factoring in the location of municipality centroid or P = i

• G (gust) is assumed to be 1.5 (Paulsen and Schroeder 2005)

• F (surface friction) linearly decreases further inland

• Vickery et al.. (2009) suggest a reduction factor of 0.7, thus a 14% reduction on the coast and 28% reduction 50km inland.

• S (asymmetry) is assumed to be 1 (Boose et al.. 2004)

• B (wind profile) is estimated using approximation method by Holland (2008);

• $B= {b}_{s}{\left(\frac{{v}_{mg}}{{v}_{m}}\right)}^{2} \approx 1.5{b}_{s}$ 

• $\frac{{V}_{mg}}{{V}_{m}}$is a conversion factor from gradient to surface wind

• ${b}_{s}=-4.4x{10}^{-5}{\Delta p}^{2}+ 0.01\Delta p+0.03\frac{{\partial p}_{c}}{\partial t}-0.014\psi +0.15{V}_{T}^{x}+1.0$ 

• $x=0.6\left(1-\frac{\Delta p}{215}\right)$ 

• $\Delta p$ is the pressure drop from the typhoon center.

• $\frac{{\partial p}_{c}}{\partial t}$ is the intensity change

• $\psi$ is the absolute value of latitude

• ${V}_{t}$ is the typhoon transition speed

  

• Rm is given by variable usa_rmw¹⁷ but can also be estimated with a parametric model Xiao et al.. (2009)

$$ln{R}_{m}=5.3259+\left(-0.0249\right)\Delta p-0.0161\psi$$

• 3. Damage Function

After estimating local wind experience, ${V}_{ijt}$, we can estimate the damage function. Typhoon damage comes in three forms: (1) wind destruction, (2) flooding, and (3) storm surge. Since these factors are all correlated, wind speed can be used as a general proxy for potential damage.

To estimate potential damage, index by Emanuel (2011) is used.¹⁸

$${f}_{ijt}=\frac{{v}_{n,ijt }^{3}}{1+{v}_{n,ijt}^{3}}$$

where,

$${v}_{n,ijt}=\frac{MAX\left[\left({V}_{ijt}-{V}_{thresh}\right), 0\right]}{{V}_{half}-{V}_{thresh}}$$

On top of the local wind estimation, we need to establish wind thresholds, details of each factor outlined below:

• ${v}_{ijt}$ is wind experienced at time t due to typhoon j at any point (P = i)

• ${V}_{thresh}$ is the threshold below which point i does not incur damage.

• Emanuel (2011) uses 50kts

• ${V}_{half}$ is the threshold where half the property in point i is damaged.

• Emanuel (2011) uses 150kts