Lost in Translation? A Critical Review of Economics Research Using Nighttime Lights Data

Abstract

In the three decades since a digital archive of satellite-detected night-time lights (NTL) data was created, thousands of scholarly articles have been published using these data. An important change in the last decade saw a significant share of highly cited articles with NTL data now written by economists. The way that economists treat the literature in other disciplines potentially interferes with the diffusion of updated findings on NTL data. Our bibliometric analysis finds that many economics studies using NTL data, especially highly cited ones, ignore studies by the remote sensing scientists who help provide the NTL data. This review considers two implications of the growing distance in the literature between economists using NTL data and remote sensing scientists. First, newer, more accurate and precise NTL data from sources like VIIRS (Visible Infrared Imaging Radiometer Suite) have slower uptake in economics, perhaps due to a lack of awareness. Yet, economists using NTL data increasingly work with spatially disaggregated units, for which the older, coarser, DMSP data are less suited. Second, a misunderstanding of DMSP spatial resolution leads to pixel-level regression studies in economics that are potentially subject to measurement error bias, for which we provide two case studies. Overall, the full value of NTL-based research may not be realized due to these weak connections.

1. Introduction

The growth in research using satellite-detected night-time light (NTL) data shows the success of remote sensing approaches to measuring economic activity and anthropogenic change. In over three decades since the National Oceanic and Atmospheric Administration (NOAA) created a digital archive of NTL data from the Defence Meteorological Satellite Program (DMSP), thousands of scholarly articles have been published using NTL data. In the first two decades of this era, most of these articles were written by researchers with a background in either remote sensing, geography and geographic information systems (GISs) or environmental sciences, even when those articles appeared in journals for other disciplines, such as economics. However, an important change occurred in the last decade that has not been previously been studied: a significant share of highly cited papers using NTL data are now written by economists. Some quantification of this change is provided in the next section of this paper.

Some features of the way that economists treat the literature in other disciplines may potentially interfere with the diffusion of updated knowledge about NTL data. There is a growing trend for economics studies using NTL data, and especially those economics studies that are highly cited, to make little to no reference to either the remote sensing literature or to the researchers who have specialized in the development and analysis of NTL data. This trend may have been enabled by the very success of remote sensing researchers in making the NTL data easily and widely available but it also stems from particular features of the economics discipline.

This review considers two implications of this growing distance between economics researchers using these NTL data and the scientists whose work is making the NTL data available. First, research in economics seems to be slower to shift from using the older and less accurate NTL data coming from DMSP to some of the newer, more accurate and precise NTL data coming from sources such as the Visible Infrared Imaging Radiometer Suite (VIIRS). Outside of economics, it has been more than four years since the annual output of new articles based on VIIRS luminosity data exceeded the annual output of new articles based on DMSP night-time lights data. In contrast, that switchover point is yet to occur for economics articles using luminosity data. Some quantitative evidence on this disciplinary gap is provided in Section 3, and three hypotheses about the continued reliance that economists have on the DMSP data are considered.

A second, and related, consequence of the distance between economists using NTL data and the scientists whose work makes the NTL data available concerns the potential misunderstanding in some economics studies regarding the spatial resolution of DMSP. In Section 3, we show that, over time, economics studies increasingly use DMSP data for analyses that rely on ever more local spatial units (e.g., counties or villages rather than countries or provinces). In some cases, each DMSP output pixel is treated as an independent observation in regression-based studies in economics because it is assumed that the DMSP sensors provide separate images for each 30-arc-second (ca. 1 km²) pixel. These pixel-level studies ignore the spatially mean-reverting errors in DMSP data that result from the sensors having coarser spatial resolution than the scale of the output grid and consequently have results that are potentially subject to measurement error bias. Two case studies are provided in Section 4 out of the published economics articles whose econometric results may be misleading because of these biases—one where the NTL data are the outcome (or left-hand-side) variable in regressions and one where they are the explanatory (or right-hand-side) variable—to show the potential ubiquity of these biases.

Overall, the evidence presented in this review provides grounds for concern that the full value of NTL data for research is not being realized, in part due to weak connections between economics researchers using NTL data and the scientists making the data available. If the studies by economists were largely ignored in the scholarly literature, this situation would perhaps matter less. However, given that some of the most cited studies using NTL data are by economists, it becomes more important to understand this issue and to have examples of where analyses may go astray because of these weak connections.

2. Highly Cited Articles Using NTL Data

The top 10 most-cited articles that use NTL data, according to a Google Scholar search in November 2024, are listed in

Table 1. Top 10 most-cited articles using NTL data.

Rank	Cites	Title	Journal	Year	Prior Top10	Economists
1	3229	Measuring economic growth from outer space	AmerEconRev	2012	No	Yes
2	1628	The new world atlas of artificial night sky brightness	SciAdv	2016	No	No
3	1559	Pre-colonial ethnic institutions and contemporary African development	Econometrica	2013	No	Yes
4	1407	Using luminosity data as a proxy for economic statistics	PNAS	2011	No	Yes
5	1147	Relation between satellite observed visible-near infrared emissions, population, economicactivity and electric power consumption	IntJRemSens	1997	#6	No
6	1090	The first world atlas of artificial night sky brightness	NotRoyAstSoc	2001	#5	No
7	1124	Regional favoritism	QuartJEcon	2014	No	Yes
8	984	Mapping city lights with nighttime data from the DMSP operational linescan system	PhotoEngRS	1997	#2	No
9	891	VIIRS night-time lights	IntJRemSens	2017	No	No
10	861	Ethnic inequality	JPolitEcon	2016	No	Yes

Notes: The citations are as of 5 November 2024, in Google Scholar. The prior top 10 refers to papers listed in Table 5 of [11]. The rank for each article provides the number in the reference list where bibliographic details can be found.

[1,2,3,4,5,6,7,8,9,10]. These articles had at least 860 citations (with a maximum of 3230 citations), at the time. The composition of the group of highly cited articles has changed considerably compared to a similar ranking that was made a decade ago [11]. Only 3 of the previous top 10 articles are in the current list, with the other 7 articles from the previous top 10 list now each having between 350 and 760 citations.

A major change from the top 10 ranking a decade ago is that one-half of these highly cited articles using NTL data have economists as authors (for the articles ranked 1st, 3rd, 4th, 7th, and 10th). The previous top 10 ranking had no articles with economist authors (1 previous top 10 article was in Ecological Economics [12], but with authors primarily from environmental science and geography). Moreover, four of the five articles with economist authors in Table 1 were published in ‘top five’ journals in economics (American Economic Review, Econometrica, Journal of Political Economy, Quarterly Journal of Economics, Review of Economic Studies); articles in these journals provide disproportionately large academic labor market rewards and therefore substantially influence the choices made by the next generation of economists [13]. In contrast, Ecological Economics is neither an “elite” nor even an “excellent” economics journal in a recent ranking study [14], making articles in that journal less likely to influence the research agenda of economists.

The key change that is highlighted in Table 1 reflects one aspect of the maturation of research using NTL data. Previously, most articles using NTL data were written by researchers with a background in either remote sensing, geography and GIS, or environmental sciences, and were typically published in remote sensing-related journals [11,15]. However, in the last decade, a wider range of disciplines started to use NTL data, with some articles from those disciplines becoming very prominent. While this reflects a success in NTL-based research, it also carries a risk of weakened connections with the literature, where the scientists who are experts on NTL data publish their updated findings. For example, the 10th paper in Table 1, which is written by economists, cites 2 of the other Table 1 papers with economist authors [1,4], but it cites no papers by authors who are experts on NTL data (such as the authors of [2,5,6,8,9]) or any other articles from remote sensing journals. This is common. The next section provides bibliometric evidence that a large and growing share of economics articles using NTL data do neither cite any remote sensing studies nor any of the experts on NTL data (such as the set of ‘most productive’ authors listed in Table 3 of [11]).

Four other aspects of Table 1 are notable. First, the economist-authored articles are fairly recent (2012 onwards) while some of the other highly cited articles are older. If citations are normalized by the years elapsed since publication, a substantially faster citation rate for the economics articles is apparent, averaging 150 citations per year versus 90 citations per year for the other five articles in Table 1. Hence, if this ranking is updated in future, the articles with economist authors will likely still be in the top 10 while some of the older articles may drop out of the list. Second, using Google Scholar rather than other citations databases was a deliberate choice. Journal review processes in economics are slow. The time from article submission to article acceptance averages 25 months at top economics journals; more than four times the submission-to-acceptance time for top science journals [16]. Given this delay, economists almost always release results as working papers which provide a large fraction of the total dissemination of their research [17]. Thus, it is important to use citations databases that count citations from working papers (like Google Scholar) because much of the economics literature is published in this form before it is awaiting publication in journals. Third, the five articles with economist authors only use DMSP data but two of the articles by NTL experts use VIIRS data [2,9]. The slower uptake of VIIRS data in economics is examined in the next section. Fourth, the table does not include highly cited articles that use data from either DMSP or VIIRS, except for where the focus is not specifically on luminosity, such as studies of fire [18] or gas-flaring [19].

3. Delayed Awareness of VIIRS NTL Data in Economics Versus Other Disciplines

Figure 1 shows the annual output (for 2012 to 2023) of journal articles in the economics subject area using either DMSP or VIIRS luminosity data. The data are from Scopus, which has the subject area as a search field (while Google Scholar now lacks this feature). From 2012 to 2016, no economics articles used VIIRS but 17 used DMSP; in the next four years, 11 articles used VIIRS and 55 used DMSP. In the last three years, the averages were 18 articles per year using VIIRS and 27 per year using DMSP, so DMSP still appears to be the more popular data source in economics. Notably, in the last three years, 21 (out of 134) articles have used both DMSP and VIIRS data, so these are scored as 0.5 articles for each source in the figure.

In contrast to the situation in economics, when all other non-economics disciplines are combined, it appears that the annual output of new articles based on VIIRS overtook the annual production of new articles based on DMSP four years ago. The trend shown in Figure 2 is based on articles that uniquely mention either DMSP or VIIRS (and ‘night’ to ensure a focus on luminosity), but the same pattern is also apparent if articles that mention both data sources are included. While articles mentioning DMSP reached a plateau and seem to now be declining, there is continued, strong growth in articles mentioning VIIRS.

There are several reasons why economists seems to be slower to recognize VIIRS NTL data compared to researchers in other disciplines. We consider three hypotheses related to the type of spatial units used in economics studies with NTL data, the length of the time series used by economics studies, and the possibility that economists are less aware of alternatives to DMSP. One reason for this lack of awareness is the tendency of some economics studies using NTL data to not cite the remote sensing literature and instead to simply cite prominent economics studies (such as [1]) that use NTL data, of which the 10th ranked paper in Table 1 is an example. This sort of citation pattern could ‘lock in’ knowledge about NTL data as it existed ca. 2010 (when some of the prominent economics studies were first drafted) whereas a discipline receiving more guidance from remote sensing studies would be able to update knowledge about new developments in NTL data.

3.1. Economists Are Mainly Using NTL Data for Aggregate-Level Studies

Our first hypothesis is that the reason for the continued reliance on DMSP data in economics studies is because these studies are mostly for aggregate spatial units, like countries, provinces or regions. The DMSP data have spatially mean-reverting errors [20,21,22], due especially to blurred images that result from pixel aggregation to save onboard memory, from geolocation errors and from angular viewing effects [23,24,25], but also due to top-coding that makes brightly lit central cities seem no brighter than dimmer suburbs and towns [26]. Radiance-calibrated DMSP data are available for 6 (mostly non-adjacent) of the 15 years from 2006 to 2010 [27], as one way to avoid top-coding and saturation problems, but few economics studies use these data [28], perhaps due to the short and discontinuous time series. Therefore, in this review we will continue to include top-coding as one contributor to the spatially mean-reverting errors in the ‘usual’ DMSP stable lights annual composites (the time series with no gaps beginning in 1992 [29]).

Evidence for mean reversion is from direct comparisons with more spatially precise NTL data (see below, and [20,22]), from lower estimates of local spatial inequality [30,31] to higher spatial autocorrelation indicators [32], if DMSP data are used rather than GDP data or more spatially precise VIIRS data. This feature of DMSP data is generally a disadvantage, as seen by various attempts to ‘correct’ for top-coding, including using certain statistical adjustments [26], but there are some situations where spatially mean-reverting measurement errors might aid researchers. For example, if one wants a measure of central tendency, such as the mean level of economic activity at a national or provincial level, mean-reverting errors may be advantageous; by suppressing local variation, prediction equations for means (or totals) for aggregated units will have a better fit [33]. This is one reason why the DMSP data are found to be comparatively better predictors of economic activity at more spatially aggregated levels than they are for spatially disaggregated units [34].

To assess the hypothesis that the reason for economists’ persistence in using DMSP data is because their research mainly uses aggregate spatial units, we considered a field of economics that uses both micro-level and macro-level data—development economics (fields that only use macro data or only use micro data may, a priori, not show changes over time). This is also a field where research relies heavily on NTL data due to a dearth of other data on economic activity, such as from surveys and administrative sources, in developing countries. Use of DMSP data for these countries also reflects concerns about the accuracy of whatever conventional economic activity data they do report, partly due to the greater share of the agricultural and informal sector in their economies [35], and due to a recognition of their generally weaker statistical capacity [1,4,36]. We systematically searched applied articles in four journals with a focus on developing countries, Journal of Development Economics, World Development, World Bank Economic Review, Economic Development and Cultural Change, and articles based on developing country data from two general interest journals (American Economic Journal: Applied Economics, and Quarterly Journal of Economics). This yielded a sample of 30 articles, with details on these provided in Appendix A.

We classified each article on a six-level scale, in terms of spatial units used in their main analyses: national; region/province/state; district; sub-district/city/county/municipality; town/village or grid larger than 5 km²; and pixel/micro-grid (<5 km²). The first four levels match the Global Administrative Areas database (www.gadm.org, accessed on 16 March 2025). Despite the hypothesis, only one-sixth of the studies use national or first sub-national level spatial units (Figure 3). The most common level of aggregation was the use of districts as spatial units, while just under one-half of the studies used even more finely grained spatial units.

The studies were coded from 1 for the most aggregated (at the national or cross-country level) to 6 for the most disaggregated (at pixel-level or for micro-grids). The average was 3.6 on that scale. To test for time trends, the value on that 1 to 6 scale was regressed on the year of publication. The more recently published studies used more finely grained spatial units, with the average value on the scale rising (i.e., moving from national level towards pixel-level) by 5.8% per year (p < 0.1). Thus, it is not true to say that economists mainly use DMSP data for studies of spatial aggregates as an explanation for their slower recognition of VIIRS NTL data. In fact, the hypothesis that economics night-time lights studies rely on spatially aggregated data is becoming even less true over time.

3.2. Economics Studies Need a Long Time Series

Another hypothesis about economists’ continued reliance on DMSP data is that they need long time series in their studies. There are two aspects to this. First, a long time series may allow for the study of specific events or interventions that occurred prior to the availability of VIIRS data. For example, several studies use DMSP data to examine the impacts of disasters, such as Hurricane Katrina [37] or the Boxing Day tsunami [38,39], that occurred before 2012 and so cannot be evaluated using VIIRS data. However, there are three counter-points to this aspect of the hypothesis. First, events in the pre-VIIRS era are also studied by non-economists without preventing the balance of the literature falling outside of the economics discipline shifting from DMSP to VIIRS (as seen in Figure 2). Second, even for events that occurred prior to VIIRS data being available, analyses could still use VIIRS data supplementally, especially to help assess the robustness of findings. For example, bottom-coding of DMSP data [25] creates ‘false zeros’ that bias impact evaluations in dimly lit places [22]; by comparing VIIRS and DMSP images for the same places but after 2012 [40], a lower bound on the extent of bottom-coding can be formed by reasonably assuming that temporal growth in luminosity makes true zeros (i.e., completely unlit places) less likely in later years. None of the studies covered in Figure 3 had these sorts of exercises that used VIIRS data. Third, many economics studies on disasters are not for specific events, but rather consider all disasters (e.g., all earthquakes, or all hurricanes) that occurred in a certain time window, such as 1992−2013, which coincides with the usual availability of DMSP annual composites [41,42]. It would be straightforward for that same research design to instead apply to, say, the 2012–2023 period using VIIRS data rather than DMSP data.

The second reason for a long time series is not so much that particular years are included, but rather so that there are enough temporal observations to allow changes to be detected (given that much of economics is focused on growth). For example, only five of the articles in the sample for Figure 3 use cross-sectional research designs; for the others, the time series averages 14 years, and the modal length of their time series is 22 years, corresponding to the widely used DMSP stable lights annual composites for 1992−2013 [8,29]. In contrast, the available time series is much shorter for VIIRS annual composites, given that it only begins in 2012 [43,44]. However, even in this case, some discussion of the tradeoffs of using a longer time series of less accurate NTL data versus a shorter time series of more accurate (and more recent) NTL data might be expected. None of the studies covered in Figure 3 had this sort of discussion. Also, a longer DMSP time series increases the threat of inconsistent luminosity measurement, as in the case of the unstable orbits that were observed earlier in the evening as satellites aged (given the 12 hr revisit time, this feature was utilized by the ‘extended’ DMSP time series that is based on pre-dawn observations [45]) and from inter-satellite differences. Notably, few economics studies use any inter-calibrated DMSP data [19,46,47,48] and instead argue that year fixed effects (separate intercepts for each year when using panel data with repeated annual observations in the same places) soak up the inter-satellite variation [1]. However, these year fixed effects (and the related averaging of digital number (DN) values for the years when two DMSP satellites provide the data) are only appropriate if satellite effects are random. However, the evidence from alternative approaches that also introduce satellite fixed effects [4,49] is that the satellite-specific measurement effects are non-random (given the statistical significance of the fixed effects for each satellite). Thus, while the length of time series explanation for why economics studies continue to be based on DMSP data may have some validity, it also has some important caveats.

3.3. Lack of Awareness of Alternative NTL Data Sources

The ongoing reliance on DMSP data in economics may also reflect disciplinary barriers to diffusion of knowledge about sources of newer and more accurate NTL data. Scientometric studies note that economics is a discipline that is insular, hierarchical, and highly concentrated in terms of authors, journals, institutions, and regions [50,51,52,53]. In terms of insularity, within-discipline citation rates greatly exceed those of other social sciences, reflecting a discipline less open to influences from the outside [50]. As noted above, this sort of insularity is exemplified by the 10th paper in Table 1, in terms of it lacking references to remote sensing studies [10]. In terms of the high degree of concentration, institutions in three micro locations (specific ZIP codes) in the United States contribute over 40% of articles in the top five economics journals and those articles received one-half of all citations, far exceeding the concentration even in bench sciences where specialized laboratories might account for specific locations being prominent. This concentration in economics publishing and citations is rising over time, even as it is declining in other disciplines [53].

This insularity showed up in an earlier review of 18 economics studies using NTL data, which found more than one-quarter of them did not cite any remote sensing studies, while attention to hierarchy was displayed by over 90% of those same studies citing [1], which was published in a ‘top 5’ journal in economics [49]. This was in spite of the NTL data being fairly new to the economics discipline at the time and these data being based on collection methods (remote sensing) that typical economics doctoral training does not cover. With this lack of reliance on the remote sensing literature, and particularly on articles published by NTL experts, it is plausible that developments, such as in the availability of newer, more accurate and precise, VIIRS data, might take longer to be recognized in economics than in disciplines that are more open to outside influences.

Table 2. Bibliometric analysis of referencing patterns for economics studies.

		Time Trend (Annual Change)
	Average Share of Papers	Articles Given Equal Weight	Citation-Weighted
Economics papers that:
Cite no articles in remote sensing	0.425	0.025 *	0.086 ***
Journals		(0.014)	(0.014)
Cite no articles by C. Elvidge	0.301	0.022 *	0.030 **
		(0.013)	(0.013)
Cite no articles by P. Sutton	0.581	0.032 **	0.008
		(0.015)	(0.034)
Cite no articles by K. Baugh	0.419	0.021	0.058 ***
		(0.014)	(0.020)
Cite no articles by either Elvidge,	0.296	0.022 *	0.028 **
Sutton or Baugh		(0.013)	(0.013)

Note: The sample is 183 journal articles published from 2013 to 2023, in “economics, econometrics and finance” fields, as identified from a Scopus search on 23 October 2024, using the following terms: ALL (“DMSP” + “night”) AND SUBJAREA (econ). Robust standard errors in ( ), * p < 0.1, ** p < 0.05, *** p < 0.01.

shows the results from a greatly expanded analysis of referencing patterns in economics to examine reliance on remote sensing studies and on studies written by NTL experts, as evidence for the hypothesis about economists lacking awareness. The sample is 183 journal articles, identified by a Scopus search covering publication dates from 2013 to 2023, where “DMSP” and “night” are mentioned in the text, for the subject area of “economics, econometrics and finance” (the finest level classification Scopus provides). In other words, this search is for economics articles that were published following what is considered to be the seminal article in economics using NTL data [1]. To ensure that the focus was just on economics journals, articles from the search that are in journals that are not in the RePEc (Research Papers in Economics) bibliographic database (https://ideas.repec.org/, accessed on 16 March 2025) were omitted. The details on these 183 articles are provided in Appendix B.

From the total number of 11,083 references in these 183 journal articles, just 360 are to remote sensing sources (3.3% of all references). A higher proportion (43%) of these 183 articles cite no remote sensing studies, compared to the fraction of non-citing studies in the earlier (less extensive) review of 18 economics articles [49]. Moreover, the proportion of economics papers based on DMSP data but not referring to any remote sensing studies is rising over time, as seen from the time trend results in the last two columns of Table 2. These time trends are estimated from a regression where an indicator variable (coded as ‘1’ for a study with no remote sensing references and ‘0’ otherwise) is regressed on the year of publication. Thus, the coefficient of 0.025 means that the share of these economics journal articles that cite no studies in remote sensing journals is rising by 2.5 percentage points per year (around a mean share of 42.5 percentage points). The results in the last column account for the fact that scholarly influence is highly unequal, by weighting each of the 183 economics articles by the citations those articles have received (as of October 2024). The weighted results show that the trend to increasingly ignore remote sensing studies in reference lists is much stronger for the highly cited economics studies. Given that more highly cited journal articles have more influence on future scholarship, it can be expected that the trend of economics studies being based on NTL data but not referring to any studies in the remote sensing literature will strengthen over time.

Not all discussion of NTL data occurs in remote sensing journals, so another way to examine the issue of economics authors ignoring guidance from DMSP experts is to see if key authors are referenced, irrespective of the outlet where those experts published. A previous systematic literature review of application of DMSP/OLS night-time light images found that three researchers—C. Elvidge, P. Sutton, and K. Baugh—were authors of more than 10% of the 144 articles (published from 1992−2013) that were covered [11]. Indeed, one author, C. Elvidge, was on more than 30% of those papers reflecting his wide-ranging contributions. However, turning to the 183 economics articles in the current sample, 30%, 58% and 42% of them make no reference to any papers by Elvidge, Sutton, or Baugh, respectively. Given that these three researchers sometimes co-author on the same papers, we also consider the combination of all three authors. The results reported in the last row of Table 2 indicate that 29.6% of the economics papers in this sample cite none of either Elvidge, Sutton, or Baugh. Moreover, the statistical significance of the coefficients in the last two columns shows that the trend of ignoring these DMSP experts is becoming stronger over time, especially for the economics papers that are highly cited (and as noted above, such studies are the ones that are more likely to influence future authors).

4. Cases of Economics Studies Misunderstanding Spatial Precision of DMSP Data

The slower uptake of VIIRS NTL data in economics suggests a foregone opportunity for analyses to be grounded in more spatially precise and temporally consistent luminosity measures. However, even though economics articles using NTL data may be poorly informed due to their tendency to ignore remote sensing articles (and articles by NTL experts published elsewhere), nothing in the results in Section 3 suggests that studies may consequently be misleading. In this section, we go further, examining the specific issue of how the spatial resolution of DMSP data is interpreted in some economics studies, and how this may undermine analyses. Some economics studies treat each DMSP output pixel as an independent observation in regressions, assuming that DMSP sensors provide separate images for each 30-arc-second (ca. 1 km²) pixel. These pixel-level studies ignore the spatially mean-reverting errors in DMSP data [20,21,22] that result from the sensors having coarser spatial resolution than the scale of the output grid [23,25]. Consequently, these studies have reported results that are potentially subject to measurement error bias.

4.1. Textural and Citation-Based Analysis

A distinction between the underlying spatial resolution of the DMSP sensor and the fineness of the output grid for the generated NTL data products is made in some remote sensing papers. For example, a recent study linking NTL data to human mobility in Africa [54] notes (p. 2) that, in comparison to DMSP, the VIIRS instrument has the following:

“increased spatial resolution of both the Ground Instantaneous Field of View (i.e., 0.55 versus 25 km² at Nadir) and the corresponding generated global grids (i.e., 15 versus 30 arc-second grid cell corresponding to ~500 m versus ~1 km at the equator)”

In other words, while the output grid for the DMSP data products is twice as coarse as that for VIIRS data products, these data products are derived from images made by instruments where the ground footprint for the DMSP sensor is 45 times as coarse as that for the VIIRS sensor [23]. This distinction should especially matter to pixel-level econometric analyses of DMSP data if the lack of independence between the DN values for adjacent output pixels (given the far coarser initial image) is not taken into account.

However, this distinction between the spatial scale of the output grid and the resolution of the sensor was absent in all of the economics studies that we reviewed. To provide some textural examples, we bold and underline key verbs in the following quotations taken from some of these economics articles that use DMSP data:

“these images record average light output at the 30 arc second level, equivalent to about 1 km² at the equator” [55] (p. 66)

“light intensity is measured at 30 arc second resolution (equivalent to 0.86 km² at the equator) on a scale from 0 to 63” [56] (p. 258)

“these satellite-based remote sensors collect daily night-time light intensity data from every location on the planet at about a 1 km² resolution” [57] (p. 65)

“satellites …[are]…sending images of every location between 65 degrees south latitude and 65 degrees north latitude at resolution of 30 arcseconds” [58] (p. 587)

These quotations give the impression that the authors believe that the DMSP data sent to earth are for pixels of a 30-arc-second resolution. In none of these studies, or in others we reviewed, is there recognition that the data transmitted to earth are 5 × 5 blocks of the original (‘fine’) pixels, where this blocking is needed in order to conserve data storage. Absent of any geolocation errors, these 5 × 5 blocks may be ca. 7 km² (2.7 km × 2.7 km) at the nadir of the 3000 km swath of the DMSP sensor. However, random geolocation errors found in an experiment by Tuttle and others [24] may further expand the DMSP ground footprint.

A clear discussion of these issues was given by Elvidge and others in 2013, in a short paper entitled “Why VIIRS data are superior to DMSP for mapping nighttime lights” [23]. This paper noted (p. 65) that the data collected by DMSP are the “product of a five by five averaging of the native resolution ‘fine’ data. This results in pixel footprints that are five kilometers on a side at nadir and the footprints expand as the scan moves toward the edge of scan.” This clear discussion was accompanied by an equally clear figure showing that DMSP at nadir has a 5 km × 5 km footprint (after on-board averaging), and notes for the figure explain that it then expands toward the edge of scan. However, this paper, and particularly this clear description of the footprint, seems to be largely ignored by economists. For example, all 30 economics studies reviewed for Figure 3 were published after the Elvidge et al. (2013) paper came out, but just two of them cite it [59,60]. Likewise, none of the four economics papers whose quotations above suggested a far smaller footprint [55,56,57,58] cite the Elvidge et al. paper, despite being published some years later (from 2015 to 2020).

Figure 4 has a more complete analysis of citation patterns for the Elvidge et al. (2013) paper [23]. The 533 citations (as of November 2024) of this paper in Scopus were separated by year and subject field (which Scopus allows but Google Scholar does not). For the first seven years, citations of this paper by journal articles that are classified as economics averaged less than one per year, even though economics articles based on NTL data grew strongly over this period, as seen in Figure 1. Meanwhile, in all other fields the annual citations of the Elvidge et al. (2013) paper [23] grew strongly, from 3 in 2013 to 59 in 2019.

There has been somewhat more recognition of this paper in economics since 2020, with an average of just under four citations per year. However, this is still comparatively low given the growing number of papers in economics based on NTL data (e.g., Figure 1 shows about 40 economics papers per year based on DMSP in those years) or compared to the citations accruing for this paper from other disciplines (averaging about 65 citations per year, outside of economics). Perhaps if the economics papers covered in the samples for Figure 1 and Figure 3 and Table 2 had considered the clear guidance offered by the Elvidge et al. (2013) paper [23], some of them may have amended their research approach.

4.2. Visual Evidence on the Coarse DMSP Spatial Resolution

The idea that DMSP sends images with a 1 km² resolution, so each 30-arc-second pixel has a DN value separate from DN values of adjacent pixels, seems sufficiently embedded amongst economists that some visual evidence to examine this idea may help. Along these lines, Figure 5 and Figure 6 map an area around Washington DC, from Dulles Airport in the west to the Port of Baltimore in the east, and from Frederick in the north to Manassas in the south (giving dimensions of ca. 100 km east–west and 80 km north–south). This urban area has a lot of local variation in luminosity, due to green space (e.g., National Arboretum, Rock Creek Park, Arlington National Cemetery, Patuxent Refuge), and building height restrictions pushing clusters of brightly lit high-rises to outer areas (e.g., Tysons). There are several brightly lit airports (Dulles, IAD; Reagan National, DCA; and Baltimore-Washington, BWI) but also extensive unlit areas with at least 30% tree cover.

Figure 5 uses the Black Marble near-nadir annual composite for 2013 (near-nadir images should have less shadowing, while composing from fewer nights should matter less for permanently lit urban areas). The local variation in luminosity is clearly seen. The light output from brightly lit central areas of the two big cities (e.g., near Metro Center in Washington DC or in downtown Baltimore) reaches ca. 4500 nW/cm²/sr. The three airports and the Port of Baltimore are clearly distinguished. The strip of outlying towns (some with population over 50,000) along the I-270 highway going northwest to Frederick are also apparent. Also, dimly lit areas, like Rock Creek Park (ca. 60 nW/cm²/sr) or the National Arboretum (ca. 90 nW/cm²/sr), are distinct from their more luminous neighbors.

Figure 6 uses the DMSP stable lights annual composite for 2013. If the DMSP sensor truly observes at a 1 km² resolution, as claimed in economics studies, points of interest from Figure 5 should also be apparent, as each one of them is larger than 1 km². In fact, all of the District of Columbia, which covers ca. 180 km², shows no variation in the DMSP data. Across the entire map, which covers ca. 8000 km², about two-thirds of the pixels have DN values of 55 or above; a threshold used by some for identifying pixels subject to top-coding on at least some nights of the year [26]. The very blurred image in Figure 6 does not allow one to see where cities stop and rural hinterland starts, yet several of the studies reviewed for Figure 3 are based on samples for urban areas [61] and some of them claim that DMSP data provide a superior measure for distinguishing urban areas from non-urban ones [57].

The inability of DMSP data to show luminosity differences at a 1 km² level is not just apparent in big cities. Figure 7 maps the Kano campus of the Nigerian Law School, in which the dormitories and classrooms, for about 1000 students, are situated in an isolated location beside Lake Bagauda (surface area of 2 km²). On the other side of the lake, the town of Wak has about 25,000 people. The rural hinterland has no concentrated sources of night lights. If DMSP data truly detect differences in lights for areas as small as 1 km², the two lit areas should show up separately and the remaining area should show no luminosity.

The map in Figure 7 shows that VIIRS data meet the standard of detecting differences in luminosity for areas of approximately 1 km²; the Law School campus has 12 lit pixels (ca. 2.5 km²) and Wak has 36 lit pixels (ca. 7.5 km²), with no light recorded over the lake or from the rural hinterland. In contrast, DMSP suggests a contiguous lit area of 78 km², including the lake and much of the rural hinterland, with no separation seen between the Law School and Wak township. In other words, the DMSP data do not allow one to see NTL differences for areas as small as 1 km² contrary to claims made in economics articles.

4.3. DMSP Measurement Errors on the Left-Hand and Right-Hand Side of Regressions

The Ground Instantaneous Field of View resolution of DMSP, as clearly described by Elvidge et al. (2013) [23], makes some of the economics studies using pixel-level regressions with DMSP data questionable, in terms of the validity of their results. Specifically, their regressions will be affected by spatially mean-reverting measurement error bias. The intuition for why the errors are mean-reverting is easiest to explain using the example of on-board averaging (for conserving memory). Consider an area with one brightly lit pixel surrounded by 24 unlit pixels. After on-board averaging, the luminosity attributed to the brightly lit pixel will be dragged down to the mean (of the 25 pixels), while the 24 unlit pixels will have their values pulled up to the mean. The other sources of blurring in DMSP images also contribute towards these mean-reverting error biases [20,25].

4.3.1. A Measurement Error Framework

Consider a regression model, $y=\alpha +\beta x+u,$ for an outcome variable, with y explained by x, with intercept $\alpha$ and response coefficient β, where u is a pure random error. Rather than having data on true values of the outcome variable, researchers use an error-ridden observed value y* that is related to the true value via the following:

$$y*=\theta +\lambda y+v$$

(1)

The textbook case (also known as classical measurement error) assumes that $\theta =0,\lambda =1,$ and that $E\left(v\right)=cov\left(y,v\right)=cov\left(x,v\right)=cov\left(u,v\right)=0.$ With these assumptions, just white noise is added to the true value, and the Ordinary Least Squares (OLS) regression estimator produces unbiased estimates of the response coefficient: $E\left({\widehat{\beta }}_{OLS}\right)=\beta .$

Evidence against these assumptions has come from various sources. For example, from self-reported wages [62,63], from recalled household consumption [64], and from self-reported farm size and crop productivity [65,66]. This evidence finds that measurement errors are mean-reverting, $0<\lambda <1,$ rather than white noise, and it is found by empirically estimating Equation (1) using right-hand-side measures that are more accurate than the typical data (that are on the left-hand side). For example, self-reported wages are regressed on administrative records from the employer [62,63]. This same framework has been used with DMSP estimates as the mis-measured variable, and VIIRS data as the more accurate variable; for example, for European regions, $\widehat{\lambda }=0.7$ [20], and at the county level in North Korea, $\widehat{\lambda }=0.4$ [22]. The importance of these estimates is that when measurement errors are mean-reverting the OLS estimator is no longer unbiased, and instead yields the following:

$${\beta }_{y*x}={\displaystyle \frac{cov\left(y*, x\right)}{var\left(x\right)}}={\displaystyle \frac{cov\left(\lambda \alpha +\lambda \beta x+\lambda u-v,x\right)}{var\left(x\right)}}=\lambda \beta .$$

(2)

In other words, the estimated coefficient is some fraction of the true response coefficient with the degree of scaling towards zero depending on the strength of the mean-reversion (i.e., there is greater attenuation bias, the closer to zero is $\lambda ).$

The second case has an error-free outcome variable, $y$, but the observed value of the explanatory variable, $x*$, is an error-ridden measure of the true variable, $x$:

$$x*=\theta +\lambda x+v$$

(3)

In this case, the OLS estimator of the response coefficient yields the following:

$${\beta }_{yx*}={\displaystyle \frac{cov\left(y, x*\right)}{var\left(x*\right)}}=\beta {\displaystyle \frac{\lambda {\sigma }_x^2}{{\lambda }^2{\sigma }_x^2+{\sigma }_v^2}}$$

(4)

First, note that in the special case of classical measurement error, with $\lambda =1,$ Equation (4) simply gives the standard result; the estimated response coefficient is attenuated in proportion to the reliability ratio (the variance in the signal over the sum of variances in the signal and noise) of the mis-measured variable [67]. This ‘iron law of econometrics’ [68], implies that the estimated OLS coefficient may be a lower bound to the true effect. But with sufficiently strong mean-reversion (i.e., $\lambda$ approaching 0), this direction of bias may reverse, so that the estimated response coefficient exaggerates the true effect. This exaggeration occurs if the smaller first term in the denominator from multiplying by${\lambda }^2$ (for $0<\lambda <1$) outweighs the effects of adding the variance of the random noise term $\left({\sigma }_v^2\right).$

From this framework, we can predict that the mean-reverting measurement errors in DMSP data will attenuate OLS estimates of response coefficients when the DMSP data are the left-hand-side variable of a regression. When DMSP data are, instead, the right-hand-side variable, the response coefficients may be exaggerated, depending on how severe the degree of mean reversion is (and this likely depends on granularity of the spatial units).

4.3.2. An Example with Errors in the Left-Hand-Side Variable

A recent economics article used the 2012 DMSP stable lights annual composite, with the logarithm of DN values for 3.6 million pixels (from ca. 35,000 urban areas, mostly in developing countries) regressed on an indicator for the pixel being below 10 m elevation, as a proxy for being flood-prone [61]. The aim of the regression was to highlight the economic importance of flooding, due to the concentration of economic activity in low-lying urban areas. Elsewhere in the article, the authors used pixel-level data from DMSP annual composites for 2003 to 2008 to see how 53 floods affected luminosity. The article does not cite Elvidge et al. (2013) [23], nor does it mention anywhere that spatial resolution of the DMSP sensor is far coarser than the scale of the output grid. There is no mention of the fact that pixel-level DMSP data will be subject to spatially mean-reverting errors.

We used coordinates of these 3.6 million pixels from the replication files [61] to obtain 2012 annual composites for DMSP DN values and for Black Marble near-nadir radiances. The first column of

Table 3. Pixel-level regressions: DMSP, Black Marble, and low-lying elevation.

Dependent Variable:	ln(DN) DMSP	ln(DN) DMSP	ln(Radiance) Black Marble
Elevation < 10 m	0.18 ***		0.62 ***
	(0.04)		(0.08)
ln (Black Marble radiance)		0.28 ***
		(0.01)
R-squared	0.10	0.63	0.08

Note: N = 3,637,696. Coordinates for the grid points and the elevation indicator are from the replication files of [61]. Regressions also include country fixed effects. Robust standard errors clustered at country level in ( ), *** p < 0.01, ** p < 0.05, * p < 0.10.

reports our results from regressing (log) DMSP DN values on the low elevation indicator variable; the coefficient of 0.18 matches that reported in the original paper (Table 3, col 1). This coefficient implies that luminosity values in low-lying areas are about 20% higher than national mean values (with a standard error of 4.4 percentage points using the approximate unbiased variance estimator for a dummy variable in a semilogarithmic equation [69]). However, there are grounds to doubt the validity of this estimate; according to Equation (2), the OLS regression coefficient will be attenuated in proportion to the degree of mean-reversion ($\widehat{\lambda }$) in the pixel-level DMSP data.

The second column of Table 3 reports estimates for Equation (1), where the DMSP (log) DN values for each pixel are regressed on the radiance for that pixel from VIIRS. The initial paper already excluded DMSP pixels with DN values of 0 (as the logarithm for these is undefined) when creating their sample [61]. However, 18.3% of pixels in their sample that appear lit according to DMSP have zero radiance with VIIRS data (just as in Figure 6, with Washington DC lights spreading over unlit rural parts of northern Virginia and Maryland in the DMSP image). To incorporate these pixels, the inverse-hyperbolic sine transformation is used; this is equivalent to logging non-zero values while allowing zeros to be included [70]. The mean-reverting error coefficient is estimated as $\widehat{\lambda }=0.28;$ the null hypothesis needed for the classical measurement error, of $\lambda =1,$ is soundly rejected (p < 0.001).

According to Equation (2), the mean-reverting error in DMSP DN values will cause the response coefficient in a regression with DN values on the left-hand side to be attenuated to less than one-third of the true value. We verify this in the third column of Table 3, which reports the regression of the (log) radiances from Black Marble on the indicator variable for the pixel being below a 10 m elevation. The coefficient is 0.62, which is more than three times larger than the coefficient that was estimated with DMSP data for the same places in the same year. The semi-logarithmic functional form means that some transformation is needed to calculate results in percentage terms [69], which are shown in Figure 8. According to regressions with DMSP data, low-lying pixels have luminosity that is 20% above the national mean (with a standard error of 4.4%). Yet the actual figure, based on the more accurate Black Marble data, shows that average radiance from low-lying pixels is 85% above the national mean (standard error of 15.3%). Evidently, the results published in [61] greatly understate the potential economic costs from flooded cities because the analysis did not consider the spatially mean-reverting errors in DMSP data. While we can only speculate, a likely reason that this issue was ignored is that the authors were not guided by the very clear discussion in Elvidge et al. (2013) [23] on DMSP spatial resolution.

4.3.3. An Example with Errors in the Right-Hand-Side Variable

A recent economics article linked pixel-level DMSP DN values for 2008 and 2013 to reported latitude and longitude coordinates for each enumeration area (‘cluster’) from the Demographic and Health Survey (DHS) for Nigeria in those two years [57]. The outcomes in the regressions were anthropometric measures (height-for-age z-score, HAZ; weight-for-height z-score, WHZ; weight-for-age z-score, WAZ) for each child aged 5 and below. The regressions include child characteristics (age, birth order, gender), parent characteristics (mother age and education, father education) and household characteristics (wealth quintile, having a TV, reading newspapers, and visiting family planning agents) as control variables on the right-hand side. The key right-hand-side variable was the pixel DN value, which the authors argued allowed them to study effects of urbanization on child nutritional outcomes along an urbanization continuum rather than working with traditional urban/non-urban indicators from the survey. The (mis)understanding that the authors had about DMSP spatial resolution can be seen from their quotation above, that the DMSP sensor collects daily night-time light intensity at about a 1 km² resolution [57].

In regressions with NTL data on the right-hand side, spatially mean-reverting measurement errors may cause coefficients to be exaggerated if mean-reversion is severe enough, as shown in Equation (4). The Table 3 result, showing that $\widehat{\lambda }=0.28$, meets this threshold. An intuitive way to see this is a comparison of the coefficient of variation for (log) DMSP of 0.189, to that of (log) VIIRS, which is 0.613. If errors were random, the variance in a mis-measured variable would exceed that of a more accurate variable, giving a ‘reliability ratio’ (with variance of the mis-measured variable as the denominator) below one. But with strongly mean-reverting data, the denominator is smaller than the numerator (so the ‘reliability ratio’ exceeds one), due to the ‘shrinkage’ of the denominator from $\lambda$ being so close to 0. This is the Equation (4) condition needed for regression coefficients to be exaggerated.

Unlike the flooded cities example in Section 4.3.2, there was no replication file for the pixel-level anthropometric study in Nigeria [57]. Moreover, the sample selection process reported did not yield the same sized sample when we repeat what the authors described (see Appendix C for details). Thus, for purposes of providing an example, we used the full DHS sample for 2013, of ca. 23,000 children from 895 clusters. We obtained the DN values from the 2013 DMSP stable lights annual composite for the pixels containing the coordinates that are given for each of these DHS clusters. We built a comparison dataset with VIIRS radiances for the same pixels, forming annual values from masked monthly composites, where the purpose of masking was to rule out ephemeral lights [20]. We did not use near-nadir Black Marble data, as in Section 4.3.2, because 65% of the DHS clusters appear unlit in the near-nadir data, yet monthly all-angles data show that less than one-quarter of these clusters are continuously unlit. Hence, we tradeoff a loss in spatial precision (by including all-angles images) for greater temporal coverage.

While not mentioned in the original study, 57% of the children in the DHS lived in clusters where DMSP data show zero lights; most of these were ‘false zeros’ as VIIRS records luminosity in these places—a known problem with DMSP in dimly lit parts of Africa [40], yet the original study non-parametrically plotted anthropometric indicators against DN values without noting that this exercise could only cover the upper 43% of the distribution (as 57% live in areas with DN = 0). The main parametric regression used in the original study was a fourth-order polynomial, with linear, quadratic, cubic and quartic terms for the DN values on the right-hand side, along with the set of control variables listed above. We use the same specification for HAZ, WHZ, and WAZ, contrasting the results using DMSP data with those using VIIRS data. For compactness, we summarize the results in Figure 9 using marginal (partial) effects; these are interpreted as percentage changes in the (conditional) anthropometric indicators with respect to changes in NTL data.

To derive the values in Figure 9, the polynomial specification is used to calculate the marginal effects for each observation (child) in the sample (keeping constant the control variables), and these effects are then averaged over the whole sample. The negative (partial) marginal effects mean that, keeping parental education, household wealth and other factors constant, children in more luminous areas had slightly worse anthropometric indicators. Of course, if education, wealth and other controls were not held constant, there would be a strongly positive relationship because children in urban areas have more educated parents, wealthier households, etc., all of which help improve child nutritional status. However, our purpose is to use exactly the same set of control variables as the original study, and to then see how replacing the spatially mean-reverting pixel-level DMSP data with the more precise VIIRS data affects the regression results.

The marginal effects of luminosity (which are meant to be acting as a proxy for the level of urbanization here) appear far larger when using the DMSP data, which are subject to mean-reverting measurement errors, than when using the more accurate and precise VIIRS data. On average, the marginal effects with DMSP data are seven times as large as those with VIIRS data, for models where everything else (the outcome variables, the control variables and the sample) stays the same. While this exaggeration of the apparent effect of urbanization is contrary to the so-called ‘iron law of econometrics’ [68], it is consistent with the bias shown in Equation (4); if there is sufficiently strong mean reversion in a right-hand-side variable, the OLS regression coefficient is exaggerated rather than attenuated. In other words, a researcher using low-resolution, coarse, pixel-level DMSP data that have spatially mean-reverting errors is likely to exaggerate the effect of luminosity (or of whatever the NTL data are proxying for) on the outcomes of interest.

This exaggeration of effects when the DMSP data are used is not restricted to regression models based on parametric specifications. The original paper also reported nonparametric results, using an Epanechnikov kernel function employing a rule-of-thumb (ROT) bandwidth estimator, where the effects estimates are averages of derivatives [57]. The set of control variables mentioned in the notes to Figure 9 were also included.

Table 4. Non-parametric averages of derivatives for the effects of NTL data on the child anthropometric indicators, Nigeria, 2013.

	DMSP (1)	VIIRS (2)	Ratio of (1) to (2)
Height-for-age z-score (HAZ)	0.003	0.000	8.5
	(0.01)	(0.01)
Weight-for-height z-score (WHZ)	−0.097 ***	−0.064 ***	1.5
	(0.01)	(0.01)
Weight-for-age z-score (WAZ)	−0.066 ***	−0.043 ***	1.5
	(0.01)	(0.01)

Note: Each row of the table reports the (average derivative) marginal effect of log NTL data on the anthropometric indicator coming from a separate non-parametric regression that also includes the control variables listed in the notes to Figure 9. Bootstrap standard errors in ( ), *** p < 0.01, ** p < 0.05, * p < 0.10. N = 23,012.

shows the results using the same procedure, on the 2013 sample of ca. 23,000 children in the DHS. The average derivatives when the DMSP pixel-level data are used are about four times as large, on average, as are the average derivatives when the VIIRS data are used (and all other aspects of the specification stay the same). In other words, even with non-parametric models, there is the same exaggeration of effects if the spatially mean-reverting pixel-level DMSP data are used as the right-hand-side variables.

Finally, it is worth noting that the paper we are critiquing did not need to use DMSP data. A similar paper, also published in 2020, used DHS surveys for ten countries in east Africa to see how child anthropometric indicators vary with urbanization [71]. A key difference is that this other paper proxied for urbanization by using average radiances from VIIRS Night Lights (VNL) v.1 annual composites for a 2 km buffer around reported coordinates of DHS clusters in urban areas and a 10 km buffer for clusters in rural areas (the DHS program supplies these VIIRS estimates because reported coordinates are perturbed to preserve the confidentiality of surveyed communities—a point the authors of [57] ignored when they manually linked reported coordinates of DHS clusters to DMSP pixels). While child anthropometric indicators were generally better in urban areas, these improvements either plateau or reverse in the most luminous places, which was attributed to a worsening of feeding practices in bright city regions. This other paper, which provides this nuanced VIIRS-based evidence on child anthropometrics, is one of the few in economics to cite Elvidge et al. (2013) [23], which the authors did in the context of noting that the DMSP pixel footprint is far larger than for VIIRS. Potentially, this understanding helped the authors to choose the more appropriate NTL data source for their research study.

Discouragingly, the paper we critique [57], the results of which are affected by measurement error bias, has over 40% more citations than the related paper that used the more accurate (and appropriate) NTL data [71]. Until there is a greater penalty for economics articles that use NTL data in inappropriate ways, we can expect a continuation of poorly informed studies that misuse the NTL data and that have the potential to mislead policy-makers. It is especially unfortunate that many of these studies are for developing countries (where a dearth of other statistics makes NTL data popular), where there are currently too few researchers to allow a culture of critical replication studies to develop.

5. Discussion and Conclusions

This review provides evidence of a growing distance in the literature, between economics researchers using night-time lights data, and remote sensing scientists whose work is making those data more easily and widely available. This matters for at least two reasons. First, some of most highly cited studies using night-time lights data are now written by economists; in contrast, economists showed little interest in NTL data in the first two decades of their digital availability (roughly, during the 1990s and 2000s) even when remote sensing experts published NTL analyses in journals such as Ecological Economics. But from the 2010s, articles using NTL data started coming out in the highest ranked economics journals (journals which strongly influence research choices of the next generation of economists [13]), and were authored by economists (whose training typically has little to no coverage of remote sensing or GIS). Our bibliometric analysis of referencing practices shows that many economics articles have little engagement with the remote sensing literature, a gap that has grown over time, especially for highly cited economics articles.

This situation can be shown schematically, with a 2 × 2 setup where rows refer to time (earlier, seminal, articles versus recent ones) and columns to disciplines (economics versus remote sensing). Ideally, recent economics articles using NTL data are guided not only by seminal articles in economics using NTL data (e.g., [1]) but also by seminal articles on NTL data by remote sensing specialists (e.g., [5,29]) and by more recent remote sensing articles with updated guidance on new data products (e.g., [9,23,43,44,45]) and on a new understanding of problems with established NTL data (e.g., [25]). The left-hand side of Figure 10 shows this ideal transmission of ideas (where arrows run from cited articles(s) to articles that cite them and the box represents recent NTL-using economics articles).

Instead, what often happens is seen in the right-hand side of Figure 10. Many recent articles in economics using NTL data cite seminal economics articles that used NTL data, such as [1]. The citations may be for an intellectual debt, but also for signaling or strategic reasons; economists judge a list with articles mostly in higher ranked journals as superior to lists including lower ranked journals [72] (relatedly, strategic behavior occurs in the acknowledgements sections of economics articles [73]). Those seminal economics articles may have cited some remote sensing studies (as seen by the dashed horizontal arrow in Figure 10, where the dashed arrow represents weaker citation links than the solid arrows; as seen, for example, with [10] not citing any remote sensing studies). However, many recent economics articles using NTL data cite neither seminal articles on NTL data written by remote sensing experts (as shown in Table 2) nor contemporary remote sensing articles dealing with NTL data. Thus, authors of these recent economics articles may only understand the NTL data in a derivative way, based on the explanations in the earlier seminal economics articles, potentially, freezing understanding of NTL data at the point where it was previously (ca. 2010). For example, recent economics studies using NTL data show little awareness of the unstable orbits of DMSP satellites that observe the Earth earlier as they age [45]; this hampers temporal consistency as a proxy for economic activity because the type of illuminated activity varies with the hour of the evening. Likewise, there is little awareness of the inherent blurring of the DMSP images that occurs irrespective of atmospheric conditions, water, or ice [25]. With these referencing patterns in economics, there is incomplete transmission of ideas, with key developments from the remote sensing literature lost in translation between the disciplines.

For completeness, one could consider a version of Figure 10 showing influence flows from economics to remote sensing. It is unlikely that such a figure would reveal a reciprocal bias, where studies in the remote sensing literature that consider economics topics are ignoring various developments in the economics literature. There are two reasons for this claim. First, the economics data that may be used in such studies have a far longer history of use than NTL data, so there is broader familiarity with them and less chance of their properties being misinterpreted. Moreover, these types of data are normally produced by branches of government, like statistics offices, with a wider constituency than just economists and so clear guidelines on the construction and appropriate use of these data are typically made available (including opportunities for dissemination and feedback through multi-disciplinary end user engagement activities). Second, the recent influence of economics on other disciplines is especially through empirical work [74], a trend related to the so-called ‘credibility revolution’ that focuses on improved research designs that are expected to yield more plausible empirical estimates of causal effects [75]. This is a long-standing concern in economics, as witnessed by Nobel prizes in 2000 (Heckman, for analyzing selective samples) and 2021 (Angrist and Imbens, for methodological contributions to the analysis of causal relationships). While these developments have some relevance to remote sensing, especially because satellite-detected data can yield layers of information that help to econometrically model causal impacts of policies [76], there is a far wider set of motives for remote sensing analyses, such as monitoring and planning, and so even if remote sensing scientists are not attentive to these recent developments in the economics literature, the lack of updated discussion of research designs may affect only a small subset of remote sensing analyses.

The second reason why economists’ lack of awareness of the remote sensing literature matters is that the type of research performed with NTL data in economics is evolving. The evidence presented in Section 3 suggests that recent analyses are being conducted at a more spatially disaggregated level. Flaws in DMSP data are more apparent at the spatially disaggregated level (the effects of spatially mean-reverting measurement errors are mitigated by spatial aggregation) and so there is a greater payoff to using newer, more accurate and precise, NTL data, such as from VIIRS, when analyses are for lower-level and smaller spatial units, such as villages, microgrid cells or pixels, compared to when analyses were at national or regional level. Yet even with this move towards analyses focused on a more local scale, there is delayed recognition of VIIRS NTL data in economics, as shown in Section 3. A straightforward proxy for whether recent economics studies are informed on these issues is to check if they cite Elvidge et al., 2013 [23], which had very clear comparisons between DMSP and VIIRS in terms of spatial resolution. Our review shows that most of the recent economics studies using NTL data ignore this key reference.

The trend towards using NTL data for more finely scaled economic analyses, when combined with a lack of attention to contributions by NTL experts in the remote sensing literature, is likely to produce misleading economic analyses. For example, pixel-level regressions with DMSP data seem misguided if based on a mistaken belief of DMSP sensors separately imaging each 30 arc-second (ca. 1 km²) pixel (an interpretation suggested by our textural analysis in Section 4.1). To provide some evidence for this claim, we gave two examples of studies with pixel-level regressions that ignore spatially mean-reverting errors in DMSP data. The first example, based upon [61], shows that urban economic activity appears far more concentrated in low-lying and flood-prone areas when accurate and spatially precise Black Marble NTL data are used as outcome measures in regressions, compared to using spatially mean-reverting DMSP data. The attenuation of response coefficients corroborates what theories and examples [20,21,22] show when mean-reverting errors affect left-hand-side variables in a regression. Our second example (based on [57]) showed that effects of luminosity (as an urbanization proxy) on child nutritional outcomes were greatly exaggerated when pixel-level DMSP data were used as a right-side variable in regressions (the exaggeration showed up in both parametric and non-parametric models). This exaggeration is also consistent with what theory shows a strong enough mean-reverting error in the right-hand-side variable. Our two examples cover key issues facing developing countries—vulnerability to natural disasters like flooding and suboptimal growth of children, which can result in stunted, less productive, lives. Policy-makers in these countries can reasonably expect published economic analyses to provide reliable evidence, but our examples suggest that this is not so, partly due to the recent economic studies not being based on a clear understanding of the strengths and weaknesses of NTL data, where this enhanced understanding could have been informed by the remote sensing literature.

Beyond these two examples, the emerging literature shows how superior NTL data can enhance economic research by providing more precise and accurate indicators of economic activity. These studies have especially focused on spatial inequality [30,31] which is a topic where the spatial mean-reverting errors in the DMSP data are problematic because they create a consistent downward bias in inequality estimates, so there is a payoff to using more accurate VIIRS data for not only accurately measuring the level of, but also the time trends in, spatial inequality (which is a measurement task that is highly relevant to SDG Goal 10 to reduce inequality within and between countries). The spatial precision of the VIIRS data has also been shown to provide more plausible estimates of ‘treatment effects’ (the impact of some intervention, in this case the shutting down of an industrial zone as part of a package of sanctions in response to ongoing nuclear tests) in settings such as North Korea, where there are no other trustworthy data on local economic activity, so satellite-detected data have to be relied upon [22]. In general, any analyses where the impacts on economic activity are likely to vary locally (as with various natural disasters such as floods, hurricanes and earthquakes and for various interventions implemented at sub-national scale ) should benefit from the greater precision and accuracy of modern NTL data compared to the coarse and imprecise DMSP data.

The evidence in this review suggests that the full value of NTL data for research may not be realized, in part due to weak connections between economists using NTL data and remote sensing scientists who specialize in developing and analyzing those data. An obvious question concerns actions that might help to remedy this situation. Hopefully, economists would give more recognition to remote sensing studies, although that may require a reduced sense of ‘superiority’ (as the term is used by [50]). However, that obvious change may not gain much traction. An earlier review article published in an economics journal showed that some economics articles using NTL data did not cite the remote sensing literature [49], and the pattern has strengthened since then (Table 2 showed a higher rate of economics articles not citing remote sensing studies than what the earlier study found, for a sample that has newer articles than those studied in [49], and the coefficients on the time trends reported in Table 2 also show that the non-citation of remote sensing studies and studies by NTL experts is rising over time for economics articles). Evidently, bringing this issue to the attention of economists did not result in an observed change in their referencing patterns.

Future research could survey economists, to ask them about their referencing practices and their lack of recognition of the literature that covers developments in other disciplines, such as remote sensing. Such a survey could also ask more specifically about reasons for continuing to use DMSP data at the expense of newer, more accurate, VIIRS data. Notably, a previous survey in the United States, with samples of 100 professors from each of six social science disciplines, found that economists were very distinctive in not valuing knowledge from outside their own discipline [50]. Specifically, when asked whether they agreed with the following statement: “in general, interdisciplinary knowledge is better than knowledge obtained by a single discipline”. All of the other disciplines had a majority of professors who agreed with the statement, while 57% of the economists disagreed or strongly disagreed (in other disciplines, just 21% of respondents disagreed or strongly disagreed with the statement). Hence, the lack of reliance on the remote sensing literature by economists that is documented here may be part of a broader issue, but the paradox is that economists appear to be increasingly happy to use the night-time lights data (and other data) coming from remote sensing without also welcoming the scholarly insights that can obtained from the remote sensing literature.

There are three other possible actions; each putting more of an onus on the remote sensing researchers. One that may be ruled out immediately would be to limit access to the NTL data so as to enforce collaborations between outside researchers who want to use the data and the remote sensing scientists whose work has helped to produce the data. This is sometimes called ‘hostage authorship’ [77], and it requires a situation where outside researchers working on the article cannot proceed with their work unless conditions that the added author(s) raise are met. This might be the case if the added author(s) are needed to provide access to the data. Putting aside ethical issues, the NTL experts providing the data have been so successful that there are multiple pathways to the data, including directly from portals such as the Earth Observation Group (https://eogdata.mines.edu/products/vnl/, accessed on 16 March 2025) or LAADS DAAC (https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 16 March 2025) and also from Google Earth Engine or the World Bank’s Light Every Night (https://registry.opendata.aws/wb-light-every-night/, accessed on 16 March 2025). Thus, there does not seem to be a monopoly power that could be exploited. Moreover, a coercive approach may undermine the organic development of fruitful collaborations between remote sensing specialists and economists, such as [78,79,80,81,82,83].

The second action would be to describe how the NTL data are produced, and their consequent properties, in ways that even an economist could not misunderstand. For example, the very clear distinction made by [54] between the spatial resolution of the Ground Instantaneous Field of View and the resolution of the generated global grids is not always made in remote sensing articles describing NTL data, with typically just the spatial resolution of the generated grid reported. Perhaps this is because remote sensing practitioners and researchers have always understood the distinction, so it did not need to be spelled out so clearly. However, evidently, economists do not understand this distinction, as seen both in the textural evidence in Section 4.1 with regard to how they described DMSP spatial resolution and considering the way they use DMSP pixel-level data in regression analyses. By showing that NTL data are becoming widely used in economics, with some of these studies being amongst the most highly cited applications of NTL data, this review might help make this component of the potential audience for articles by NTL data producers more apparent. Of course, it still requires the economists to pay attention to these descriptions, and given how often the very clear discussion in Elvidge et al. (2013) [23] has been ignored, we should not be too hopeful.

The final course of action would be to embark upon more critical replications, along the lines of the two examples in Section 4.3. Showing cases where analyses may have gone astray because the economist authors had not engaged with the remote sensing literature that provides guidance on NTL data may have some cautionary effect on future authors. However, critical replications in economics are barely cited and do not seem to affect citations of the original paper, even if the original paper had serious flaws [84]. Thus, even this action may not help to close the gap in the literature that we have discussed here.