Using night light emissions for the prediction of local wealth
| Author | Nils B Weidmann,Sebastian Schutte |
| Published date | 01 March 2017 |
| Date | 01 March 2017 |
| DOI | 10.1177/0022343316630359 |
| Subject Matter | Research Articles |
Research Article
Using night light emissions for
the prediction of local wealth
Nils B Weidmann
Department of Politics and Public Administration, University of Konstanz
Sebastian Schutte
Zukunftskolleg & Department of Politics and Public Administration, University of Konstanz
Abstract
Nighttime illumination can serve as a proxy for economic variables in particular in developing countries, where data
are often not available or of poor quality. Existing research has demonstrated this for coarse levels of analytical
resolution, such as countries, administrative units or large grid cells. In this article, we conduct the first fine-grained
analysis of night lights and wealth in developing countries. The use of large-scale, geo-referenced data from the
Demographic and Health Surveys allows us to cover 39 less developed, mostly non-democratic countries with a total
sample of more than 34,000 observations at the level of villages or neighborhoods. We show that light emissions are
highly accurate predictors of economic wealth estimates even with simple statistical models, both when predicting
new locations in a known country and when generating predictions for previously unobserved countries.
Keywords
economic data, night lights, spatial prediction
Introduction
In many developing countries, official economic statis-
tics are either not available or of poor quality (Jerven,
2013). This may create problems for cross-national
research on political violence that has established strong
links between economic conditions and civil war (Hegre
& Sambanis, 2006). These limitations, however, may be
much more serious in disaggregated analyses at the sub-
national level, since data requirements are considerably
higher. Are wealthier regions predominantly affected by
conflict? What is the economic damage of violence across
the regions of a country, or the local impact of develop-
ment aid on economic recovery? To reach solid conclu-
sions on these important questions, it is necessary that
empirical researchers base their analyses on reliable
economic datasets with high temporal and spatial resolu-
tion. These statistics are available for many developed
countries – in Europe, for example, via the GEOSTAT
project (EUROSTAT, 2015) – but typically not for those
countries that conflict researchers are most interested in.
In order to overcome this problem, researchers have
resorted to alternative data sources. When it comes to
economic data, nighttime illumination observed from
satellites has been proposed as an alternative measure-
ment approach. So far, however, these attempts have
mostly been limited to coarse resolutions; we know, for
example, that night light patterns track economic growth
at the national level. How far can we increase the reso-
lution of this approach? Is it possible to predict economic
wealth at the village or the neighborhood level solely
from the level of nighttime illumination? In this article,
our aim is to show that it is. Our approach is different in
several ways from the majority of works on conflict pre-
diction, and therefore also from most of the other con-
tributions to this special issue. First, rather than
predicting violence or war, we predict a covariate that
is frequently required in conflict studies: economic
Corresponding author:
nils.weidmann@uni-konstanz.de
Journal of Peace Research
2017, Vol. 54(2) 125–140
ªThe Author(s) 2016
Reprints and permission:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0022343316630359
journals.sagepub.com/home/jpr
wealth. Therefore, we see prediction as an improvement
of the data collection process for conflict studies, before
any analyses of the actual outcome of interest are con-
ducted. Second, and more importantly, our innovation
as regards prediction is the use of an alternative data
source, not a new prediction model or methodology.
The latter typically constitutes the focus of current
research in prediction (see for example Beger, Dorff &
Ward, 2014, or other contributions to this special issue).
In contrast, we show that progress can also be made by
using new data in conjunction with simple models.
In our analysis below, we find night lights to be good
predictors of wealth at the local level. Across the countries
we analyzed, the correlation between night light emission
and wealth is on average 0.73, and can be as high as 0.87.
In order to test the accuracy of our predictions out-of-
sample, we set up two prediction tasks. The first is
within-country: given a training set with data on wealth
and night lights for a number of locations in a country,
can we predict the wealth of unseen locations based only
on their night light illumination? Predictive performance
is very high, indicating that night lights data have great
potential for subnational analyses. The second, cross-
national task is to predict subnational wealth for new
countries, training the models on data for other coun-
tries. This task is more difficult, because absolute levels of
night light emissions can vary considerably across coun-
tries (Kyba et al., 2014). However, we show that with a
simple normalization procedure, we can also reach a
reasonable predictive performance in this second task.
Our article starts by briefly reviewing the literature
on the use of night lights in the social sciences, before
describing our data and methodology. We then present
our results, starting with simple descriptives and moving
on to our two prediction tasks, within-country and
across-country. Last, we summarize our findings and
outline ways in which they can be used in actual research.
Existing work
The use of night lights as a proxy for economic variables
typically assumes that nighttime illumination corre-
sponds to wealth through one of at least three channels.
First, access to the power grid (or a power generator)
requires financial investment, which is likely to be made
by people with the necessary resources. Second, night
lights indicate economic activity, which can lead to
higher levels of wealth for the people involved (Hender-
son, Storeygard & Weil, 2011). Third, nighttime illu-
mination (street lamps) can be a result of preferential
treatment by the state for certain societal groups (Hodler
& Raschky, 2014). Whatever mechanism we assume,
high light emissions should be correlated with high
levels of wealth. This assumption, however, is not unpro-
blematic. For example, economic activity may not ben-
efit the people living at the location where it occurs –
bright commercial centers in cities may be inhabited by
poor people. Also, the amount of light emitted by eco-
nomic activity may not scale directly with the benefits it
generates for the local population: oil refineries, for
example, are typically illuminated at night, but require
few staff and do not coincide with residential areas.
Therefore, the question of whether night lights corre-
late with wealth is an empirical one. Existing analyses
have tried to assess the use of these data at different levels
of analysis. Early work conducted at the country level
reveals a clear correlation between the area illuminated at
night and economic output (Elvidge et al., 1997). This
correlation alone, however, does not tell us whether
night lights track economic activity, since the correlation
could simply be due to country size – on average, larger
countries have larger economies, but also emit more light
at night. This is why subsequent work has examined this
relationship further. Henderson, Storeygard & Weil
(2011) show that changes in night lights track economic
growth, which provides strong support for the night
lights–wealth relationship.
Similar results hold for economic output of subna-
tional units of analysis (states or provinces), which have
been approximated using night light patterns. An article
by Sutton, Elvidge & Ghosh (2007) conducts such an
analysis for four countries (China, India, Turkey, and
the USA), showing that night lights track economic out-
put also at this finer level of resolution. More recently,
Chen & Nordhaus (2011) present a global study that
compares night light emissions to economic output mea-
sured at the level of 1-degree (approx. 100 km by
100 km) grid cells. One of the most detailed analyses
so far was carried out by Mellander et al. (2013) for
Sweden using fine-grained official socio-economic data
on businesses and individuals. This study clearly shows
the limitations of night lights-based analyses, in particular
in developed countries. In these countries, a key problem
of night light measurement – top-coding (see below) –
makes these datamuch less useful as a predictor of wealth.
Because of the number of promi sing results, night
lights data have seen some adoption in the social sciences
and conflict research in order to approximate economic
variables. Shortland, Christopoulou & Makatsoris
(2013) use fine-grained data on night light emission to
estimate the economic impacts of violence in Somalia. A
number of works on inequality rely on night lights data,
126 journal of PEACE RESEARCH 54(2)
To continue reading
Request your trial