The health of populations is inextricably linked to the depth of their poverty 12. Breaking the vicious cycle of poverty and ill-health has formed the basis of the international community's Millennium Development Goals (MDGs) 3. At national levels, targeting resources to those most in need is a guiding principle of poverty reduction strategies and health policies 4. However, obtaining accurate metrics on the depth and spatial disparities in poverty poses several problems. Measures of poverty at household level are often computed from complex survey data on income, consumption or expenditure 5. These data are difficult to reliably collect at regular intervals nationally; are subject to significant reporting bias; show large fluctuations over time; or are seen as indicative only of the short term economic status of the sampled households 67. A default metric that is used more frequently, and is easier to collect during household surveys, is based on assets variables 68. In sub-Saharan Africa (SSA), most national household surveys now have a standardized welfare module that routinely collects information on household assets and are used to report the socio-economic patterns in health outcomes 910. Several of these common asset variables have also been shown to be associated with income and consumption 1112 and this relationship is now the basis of poverty mapping using small-area estimation methods 1314.

Asset-based wealth indicators, although easier to collect, suffer from limitations similar to those of income- and consumption-based indicators often resulting in metrics that are not comparable across countries, or even within countries, especially where the relationship of input variables to well-being varies across different social and geographical settings 615. Therefore, where the aim is to relate poverty to other metrics such as health across multiple geographic entities, these poverty measures become deficient. Furthermore, regardless of which survey-based measure of poverty is used, the process of collecting the relevant data to allow the examination of detailed sub-national differences in poverty and resource need is expensive. Alternative measures are therefore required that are easier to interpret, comparable temporally and spatially across national and sub-national boundaries and for which data are less expensive to obtain.

The spatial distribution and intensity of satellite-derived night time lights (NTL) has been shown in several studies to correlate with per capita gross domestic product (GDP) and other national level socio-economic indicators 1516171819. It has also been shown to be a good proxy for population distribution 20. This simple source of information is derived from satellite imagery at high spatial resolutions and is readily available in the public domain 18. However, until now analysis using NTL as a proxy for poverty has only considered its relationship with consumption-based measures in high-income countries 1519 where such data exist. Here we seek to examine the correlation between NTL and wealth asset indicators of poverty at sub-national spatial resolutions in Africa.

Comparable household assets data were available for 338 Administrative 1 units in 37 out of 56 African countries (Figure 1 & Table 1). Assets data for 15 out of 37 countries were obtained from surveys done in the year 2000 (corresponding to the production year of the NTL): seven countries in 1999 or 2001; two countries in 1998; and the rest of the countries in 2003 – 2006 (Table 1). The first component from which the asset index was derived explained 43.3% of the variation in the asset data. Overall the wealth index based on the five asset variables ranged from a mean of -1.67 in Somalia to 3.45 in Egypt. The mean (standard deviation) of brightness of light pixels ranged from 0.0061 (0.3419) digital numbers in Chad to 1.9321 (8.1347) in Egypt. Chad and Somalia ranked as the countries with lowest mean brightness of NTL each with a value of 0.0097. Overall, 2.2% of the total area of the 37 countries was covered by NTL, ranging from 0.07% in Chad to 17.28% in Swaziland while Egypt had 12.18% of area covered by NTL. The mean distance to nearest NTL pixel was highest for Central African Republic (163.71 km) and lowest for Comoros Islands (7.07 km). Overall, 26 out of 338 Administrative 1 units did not have any NTL pixels.

According to the asset index 18 out of 37 countries did not have a single Administrative 1 unit in the least poor quintile, with 97 out of 165 units in these countries ranked in the poorest and second poorest quintiles (Table 2& Figure 3). Among those 18 countries which did not have Administrative 1 units in the least poor quintile, Somalia, Chad, Central African Republic, Niger and Angola had 50% or more of their units in the most poor quintile. In contrast, all of 7 and 26 Administrative 1 units in Morocco and Egypt respectively were in the least poor quintile. When the quintile rankings based on the mean brightness and distance to, and proportion of area covered by, NTL were considered, the countries that dominated the bottom and top quintiles generally remained the same with Chad, Somalia and the Central African Republic consistently ranked as the 'poorest' while Egypt, Morocco, Swaziland and South Africa the 'richest' (Table 2).

Scatter and box plots of the continuous and ordinal (quintile) relationships between the asset-based wealth index and the three NTL measures at Administrative 1 level are shown in Figure 2. While mean brightness of, and proportion of area covered by, NTL exhibited positive correlation with the assets-based index, the mean distance to nearest NTL pixel, as anticipated, showed a negative correlation. All the NTL indicators distinguished unambiguously between the most and least poor quintiles based on the assets index. Their strength, however, in separating the middle quintiles was generally weak (Figure 2). The Pearson and Spearman correlation coefficients of the asset index versus the three NTL measures are presented in Table 3. In the continuous form, the mean brightness of NTL exhibited the strongest correlation with asset-based wealth index of all three NTL indicators (Pearson correlation = 0.64, p < 0.01) (Table 3). When the quintiles based on the assets-based wealth index were compared to those based on the three NTL measures, the quintile rankings of the mean of NTL brightness had the highest Spearman's rank correlations of 0.79 while those of the mean distance to nearest NTL pixel had the lowest correlation (-0.62) with the asset-based index. The corresponding Kappa statistic was 0.64 and 0.58 showing substantial and moderate agreement with assets index respectively (Table 3).

Currently international development milestones such as the MDGs, which comprise a set of eight internationally agreed goals that cover areas such as poverty reduction, education, infrastructure and health, use asset-based wealth quintiles as a way of monitoring changes in socio-economic inequity 34. NTL, whilst representing a narrower dimension of human development compared to the combined asset variables of wealth, provide the benefit of being easily available and comparable spatially and temporally at a high spatial resolution. In this study we have shown that the mean brightness of the NTL human settlement product had a reasonably high linear correlation with asset-based indices at the Administrative 1 unit level in Africa (Table 3 & Figure 2) as both a continuous (Pearson's correlation coefficient = 0.64) and ordinal (Spearman's correlation coefficient = 0.79; Kappa = 0.64) variable. The ordinal forms of all the NTL metrics clearly separated the most and least poor quintiles with the median asset-based index of these quintiles not overlapping (Figure 2). While we have examined solely the use of 2000 NTL data here, the forthcoming production of more contemporary human settlement products 31, the constant acquisition of new NTL imagery 35 and even the possibility of finer resolution NTL imagery 36 mean that the potential to track changes in poverty levels over large scales exists, and this will be a focus of future research.

The main attraction of presenting poverty or socio-economic data on an ordinal scale, such as quintiles, is the ease with which results can be interpreted by policy makers and planners. This is especially the case when such a scale is used to define heterogeneity in specific population indicators such as fertility, mortality or access to public services. The problem with ordinal scales, however, is that information in intermediate classes, (2nd, 3^rd and 4^th in the case of quintiles), is rarely distinct and difficult to interpret. Consequently, most studies and programmes focus mainly on the difference between the top (least poor) and bottom (most poor) quintiles. In this regard, the significant positive correlation between asset indices and the mean brightness of NTL, particularly in the ordinal form, provides an opportunity for using the latter as an alternative poverty metric to asset-based indices, with the additional benefit of preserving independence and comparability across geographic settings, particularly in most of Africa where the use of electric lighting remains generally low with significant between and within country variation 37. Arguably, as more recent national survey data that record household level variables become available, the need for such NTL metrics will decrease for within country evaluations. In addition, it is possible the NTL metric is a weak proxy of poverty at cluster level given that its distribution at such small area level is likely to be homogenous. The strength of NTL data, however, is in their ease of extraction, their comparability across space and their repeated measurements.

Our findings on the relationship of asset indices and NTL in Africa are comparable with previous studies where various NTL metrics were shown to be useful indicators of economic activity and correlated with GDP 17 and income per-capita 15 in Europe and the USA. In combination with gridded global population maps, NTL brightness was also shown to be a relatively accurate metric for computing populations below national and international poverty lines 19. However, there are issues of scale dependence 1738 whereby different results can be observed from the same data aggregated at different geographic scales which can lead to erroneous imputations from observations at a smaller geographic unit to a larger one or vice versa 17. In this analysis it is not clear whether the fidelity of our observations will remain when aggregated to resolutions finer than the Administrative 1 level in Africa. In addition, the NTL data used here suffer from a 'blooming' effect – the tendency to over-estimate the extent of large, well-electrified urban areas 1839, a problem which the new generation of NTL products in production attempt to resolve 40. It is possible, therefore, that the strength of the relationship between asset-based indices and NTL metrics observed at Administrative 1 level for Africa may not hold at lower resolution and caution should be exercised when extrapolating the findings of these results.

The study shows that in Africa mean brightness of NTL is highly correlated with asset-based indices at the Administrative 1 level. The observations made here are plausible given that where there are more investments in infrastructural development, particularly in urban settings, people are on the whole wealthier 41. The rate of urban development and electrification in Africa is discussed elsewhere 37. What this study shows, however, is that public domain, spatially continuous and temporally dynamic data on NTL can be used to track changes in poverty levels and that these relate to current standards of poverty measurement.

DMSP: Defense Meteorological Satellite Program; GDP: Gross Domestic Product; IMF: International Monetary Fund; MDGs: Millennium Development Goals; NOAA-NGDC: National Oceanic and Atmospheric Administration-National Geophysical Data Centre; NTL: Night-time lights; OLS: Operational Linescan System; PCA: Principal Component Analysis; SSA: sub-Saharan Africa; UN: United Nations; WHO: World Health Organization;

The authors declare that they have no competing interests.

AMN was responsible for study design, data cleaning, analysis, interpretation and production of the final manuscript. VAA contributed to data cleaning and analysis and production of final manuscript. PWG contributed to analysis, interpretation and production of the final manuscript. AJT contributed to analysis, interpretation and production of the final manuscript. RWS was responsible for overall scientific management, analysis, interpretation and preparation of the final manuscript.

AMN is supported by the Wellcome Trust as a Research Training Fellow (#081829). RWS is supported by the Wellcome Trust as Principal Research Fellow (#079081). Both AMN and RWS acknowledge the support of the Kenyan Medical Research Institute. This work forms part of the output of the Malaria Atlas Project (MAP: http://www.map.ox.ac.uk), principally funded by the Wellcome Trust.