Is newer better? Evaluating the suitability of nighttime luminosity in proxying poverty in Africa

Abstract

Purpose

The purpose of the study is to evaluate whether nighttime luminosity sourced from the Defense Meteorological Satellite Program-Operational Linescan System satellite sensors is a suitable proxy for measuring poverty in Africa.

Design/methodology/approach

Our study performs wavelet coherence analysis to investigate the time-frequency synchronization between the nightlight data and “income-to-wealth” ratio for 39 African countries between 1992 and 2012.

Findings

All-in-all, the authors find that approximately a third of African countries produce positive synchronizations between nighttime data and “income-to-wealth” ratio and hence conclude that most African countries are not at liberty to use nighttime data to proxy conventional poverty statistics.

Originality/value

In differing from previous studies, the authors examine the suitability of nightlight intensity as a proxy of poverty for individual African countries using much more rigorous analysis.

Citation

Hamman, N. and Phiri, A. (2023), "Is newer better? Evaluating the suitability of nighttime luminosity in proxying poverty in Africa", African Journal of Economic and Management Studies, Vol. 14 No. 1, pp. 150-167. https://doi.org/10.1108/AJEMS-02-2022-0042

1. Introduction

The first objective of the United Nations Sustainable Development Goals is to eradicate global poverty by 2030 and whilst most industrialized economies are making significant progress in reaching these objectives, African economies lag far behind as the continent is confronted by serious development challenges, such as underweight children, child mortality, primary school completion and access to safe water and electricity, all which hinder the continent’s progress in escaping its poverty trap (Moyer and Hedden, 2020). Further compounding the continent’s misfortunes is the lack of reliable poverty data which could assist policymakers to make interventions in the most needed areas, evaluate effectiveness of implemented programmes and navigate poverty alleviating strategies in an increasingly digitized and environmentally degrading world (Canudas-Romo, 2018; Moatsos and Lazopoulos, 2021). For instance, Ghosh et al. (2010) and Nordhaus and Chen (2012, 2015) highlight that at least four developing countries, including Somalia and Liberia, do not have functional national statistical offices and poverty indices are scarcely reported for African countries on the World Bank’s development indicators. Moreover, Engstorm et al. (2017) observe that between 2011 and 2017, as many as 57 developing countries had conducted no household surveys to collect poverty data whereas Yeh et al. (2020) caution that given the current frequency and coverage of surveys conducted in African countries, households in these nations will appear in a survey once every 1,000 years. These realizations have motivated researchers to explore alternative measures of poverty using remote sensing data derived from Defense Meteorological Satellite Program-Operational Linescan System (DMSP-OLS) satellite sensors and numerous studies have used nighttime data to map out economic development and poverty trends in developing countries (see Donaldson and Storeygard, 2016; Gibson et al., 2020; Levin et al., 2020 for in-depth reviews).

An ongoing consensus amongst academics is that regions/areas with more (less) nightlight intensity have higher (lower) development levels and lower (higher) poverty rates and by using/establishing empirical relationships between nightlight data and development statistics, synthetic measures of economic development and poverty can be developed as proxy measures at national (Elvidge et al., 2009), provincial (Wang et al., 2012; Yin et al., 2021), district (Engstorm et al., 2017; Sangkasem and Puttanapong, 2020; Cecchini et al., 2021; Gibson et al., 2021) and city (Yu, 2015; Xiang and Xiao, 2021; Yong et al., 2022) levels where this data cannot be easily collected using traditional survey-based techniques. The extensive geographical coverage of nightlight intensity is the primary advantage of nighttime data and contributes to big data analytics which is increasingly being recognized as the key to the fourth industrial evolution. Further considering that nightlight data is freely available, presents an additional advantage over survey-based collection methods which are costly and require well-developed statistical institutions and trained staff to ensure regular collection of quality statistics. For instance, Cecchini et al. (2021) highlight the difficulties faced by statistical offices worldwide in conducting and collecting survey household statistics during the COVID-19 pandemic due to the lockdowns and social distancing policies, and insinuate that these problems could be circumvented using remote-sensing data.

For nightlight luminosity to be considered a suitable proxy for poverty, a basic requirement is the validation of a significant empirical relationship between nighttime intensity and poverty and whilst we acknowledge the existence of some empirical studies which provide correlation estimates between the variables, our study proposes a re-examination of the empirical evidence for African countries based on two observations. Firstly, there exists very little empirical evidence for African countries with the works of Noor et al. (2008) and Mveyange (2015) being sole exceptions. We note that these studies are panel-based and hence ignore any heterogeneity effects amongst the individual countries. Secondly, previous studies base their analysis on traditional estimation techniques which tend to ignore asymmetries and time variation existing within the data. These asymmetries could be due to satellite sensor noise and capturing of light activities which are unrelated to different levels of poverty (Bickenbach et al., 2016).

Considering the increasing use of DMSP-OLS nighttime data as a proxy for modelling and forecasting poverty (Jean et al., 2016; Engstorm et al., 2017; Galimberti, 2020), the research question which our paper addresses is whether the observed correlation between nightlight intensity and conventional measures of poverty in Africa (i.e. wealth-to-income ratio) can hold under rigorous empirical scrutiny. To this end, our study investigates the co-movement between DMSP-OLS nighttime data and wealth-to-income ratio for 39 African countries using the complex wavelet coherence tools which allows us to examine the amplitude and phase dynamics in the synchronization between variables within a time-frequency domain. Wavelets are essentially mathematical tools which decompose a time series into a time-frequency space and consequentially yield localized time-frequency information on the series. This is unlike conventional econometric tools which can either analyze data localized in time (i.e. OLS, vector autoregressive (VAR), vector error correction model (VECM)) or in frequency (i.e. Fourier analysis) but not in both time and frequency.

All-in-all, our study presents a unified analytical framework which enables us to investigate the nightlights–inequality nexus over four dimensions, namely, time, frequency, magnitude (strength) and phase (negative or positive correlations as well as direction of causality). The use of wavelet tools allows us to reveal new stylized facts on the co-movements between the nightlight intensity and inequality for African countries which will allow us to inform policymakers and other stakeholders on which individual African countries are at liberty to use DMSP-OLS nightlight as a proxy for poverty based on time-varying and cyclical correlations observed between the variables.

The rest of the paper is presented as follows. The next section presents a brief review of the associated literature. The third section outlines the methodology whilst the fourth section presents our empirical findings. The study is concluded in the fifth section in the form of policy implications and venues for future research.

2. Literature review

Following the official release of the DMSP_OLS archive by the National Oceanic and Atmospheric Administration (NOAA) in mid-1992, there have been extensive empirical applications of nightlight data by the scientific community as proxies for human activity and development (Donaldson and Storeygard, 2016; Gibson et al., 2020; Levin et al., 2020). Even though the original purpose of nightlight data was to monitor changes in weather patterns for military intelligence, Croft (1978) made an interesting observation that nighttime data could be used to map out different kinds of geographical boundaries and human activity across the globe and since then, many researchers have used observed correlations between the nighttime and formal statistical data to create synthetic measures of economic activity.

To keep the review of the literature tunnel focused, we strictly provide a discussion of empirical studies which have used DMSP_OLS data to estimate regressions between nightlight intensity and poverty to create synthetic measures of poverty. Starting with the earlier study of Noor et al. (2008) who use principal component analysis (PCA) to create asset-based poverty indices from national survey household data for 338 administrative units in 37 African countries and find positive correlation coefficients from 0.64 to 0.79. Elvidge et al. (2009) creates a map of global poverty for 233 countries using a poverty index which is obtained by dividing the population by the nightlight intensity. The authors establish that the poverty index has a strong positive relationship with population living below the poverty line of $2 per day, producing a coefficient of determination (R²) of 0.7217.

Wang et al. (2012) use PCA to create integrated poverty index (IPI) and average light index (ALI) using data for 31 Chinese provinces and municipalities between 2007 and 2009 and find a positive relationship between the indices producing an R² value of 0.854. Mveyange (2015) estimates positive and significant correlation coefficients between nightlight intensity and the Gini index for 36 South African districts (ρ = 0.34) and 26 Brazil states (ρ = 0.45) for the years 2000 and 2012 and uses these estimates to create synthetic poverty time series measures for 622 regions in 42 African between 2000 and 2012. Yu (2015) uses PCA to create IPI and ALI for 38 counties in Chongqing city and 2,856 Chinese counties and finds a very weak relationship between the series. Sangkasem and Puttanapong (2020) also use PCA to create IPI and ALI indices for Thailand using time series data spanning from 1994 to 2005 and find significant correlation coefficients ranging from 0.7255 to 0.7588.

More recent works include Andreano et al. (2021) who estimate the relationship between nightlight intensity and Gini index using quasi-maximum likelihood estimates applied to time series data spanning between 1992 and 2013 and find a positive relationship between the variables with an R² = 0.78. Cecchini et al. (2021) estimate the relationship between nightlight intensity and poverty survey data from 139 Chilean municipalities for the year 2015 using Fractional Multinomial Logit model and produce the expected negative and statically significant relationship between the variables at different levels of poverty. Gibson et al. (2021) estimate the relationship between nightlight intensity and poverty indices (Gini and Theil index) for 497 Indonesian spatial units, 288 Chinese prefectures and 234 South African municipalities using 2011–2012 data and find a positive and significant relationship between the variables for all three countries in both rural and urban areas.

From the review of this literature, we point out at least three empirical gaps in the nightlight–poverty nexus which our study attempts to bridge. Firstly, we expand on the literature for African economies of which the works of Noor et al. (2008) and Mveyange (2015) are currently the only studies available in the literature which provide regression estimates of nightlight intensity and poverty for African countries. Secondly, most previous studies use panel estimates of nighttime and poverty data which aggregates the findings for different countries with different levels of development across a singular year. Only the studies of Sangkasem and Puttanapong (2020), Andreano et al. (2021) and Yong et al. (2022) use a time series approach for samples which do not include African data. Lastly, previous studies do not adequately account for possibly asymmetries in the data caused by satellite sensor noise which causes economic development to be reflected by nighttime lights in a non-monotonic manner and also differences in the sectoral composition of a country’s GDP, where some industries may generate more lights than others (Bickenbach et al., 2016; Galimberti, 2020).

3. Data and methods

3.1 Data

This study uses two sets of time series variables, nightlight intensity and the “wealth-to-income” ratio, collected for 41 sub-Saharan African countries on annual frequency between 1992 and 2012. On one hand, the DMSP_OLS nighttime light (NTL) time series dataset is retrieved from the US Air Force subdivision Earth Observation Group’s (OEOG_NOAA) online website [https://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html]. On the other hand, the “wealth-to-income” ratio (W/I) time series is sourced from the World Inequality Database [https://wid.world/] website.

3.2 Conventional estimators

Following conventional literature, we specify the following regression:

(1)W/Ii,t=αi,t+βi,t NTLi,t+εi,t

where i = 1, 2, … N; t = 1, 2, …, T; α is the intercept; β is the regression coefficient which is expected to be positive such that well (poor) lit areas have higher (lower) wealth-to-income ratios; ε_i,t is a well-behaved error term. Traditionally, Equation (1) has been estimated using traditional panel ordinary least squares (POLS) estimators. In our study, we employ three additional cointegration estimators to estimator regression (1). Firstly, we make use of the fully modified OLS (FMOLS) estimator which is constructed by making corrections for endogeneity and autocorrelation to the β coefficient from regression (1) i.e.

(2)β^=[∑i=1NL^22i−2∑i=1T(NTLit−NTL^i)2]−1[∑i=1NL^11i−2L^22i−2(∑i=1T(NTLit−NTL^i)WI^it∗−Tγ^)]

where L_i is the lower triangular matrix of the asymptotic covariance matrix and γ^ is a kernel estimate. Secondly, we make use of the dynamic OLS (DOLS) estimator which uses past and future values of the differences of the independent variable as additional regressors i.e.

(3)W/Ii,t = αi,t + βi,t NTLi,t + ∑j=−qqcijΔNTLit+j+u∗it

where u*_i, = u_it + ∑j>qqcijeit+j. Notably, FMOLS and DOLS estimators do not account for short-run and error correction equilibrium dynamics between the time series. Therefore, we employ the pooled mean group (PMG) estimator of Pesaran et al. (1999) as our last estimator. To this end, we respecify Equation (1) as the following panel autoregressive distributive lag (p, q,q…….q) model:

(4)W/Iit=∑j=1pλijW/Ii,t−j+∑j=0qδijNTLi,t−j+αi+εit

where t=1,2, …,T, i=1,2, …,N, αi is the fixed effect, λij and δij are vectors of parameters. The error correction representation of Equation (1) is as follows:

(5)ΔW/I=ϕiW/Ii, t−1+NTLitβi+∑j=1p−1λij∗ΔW/Ii,t−1+∑j=0q−1ΔNTLi,t−jδij∗+μi+εit

where εit' are serially not correlated across i and t, have zero means, variance σi2>0,  and finite fourth-order moment conditions, and

(6)ϕi = −1 (1−∑j=1pλij) and βi=∑j=0qδij 

The long-run relationship can compactly be denoted as follows:

(7)W/Iit=θiNTLit+ηit

where θi= −βi′ϕi are the long run-run coefficients and ηit is a stationary process. The long-run coefficients defined by θi are constrained to be the same for all cross-sectional units and can be expressed as follows:

(8)ΔW/Ii=ϕiξi(θ)+Wiki+εi, i=1,2,…, N

With

(9)Wi=(ΔW/Ii,−1,…,ΔW/Ii,−P+1,ΔNTLi,NTLΔi,−1,…,ΔNTLi,−q+1,ι),

(10) ki=(λi1∗, …,λi,p−i∗,δi0∗′ ,δi1∗′,…,δi,−q−1∗′, μi)′,

And the error correction term is computed as follows:

(11)ξi(θ)=W/Ii,−1−NTLiθ,             i=1,2,…N

And this measures the speed of “correction” back to steady-state equilibrium following a shock to the system of time series variables.

3.3 Wavelet coherence analysis

Wavelets are small waves that stretch and compress in a limited time period and are used to decompose a signal or time series across a time-frequency plane and these transforms can either be discrete (returns data vector of the same length as the input signal) or continuous (returns an output vector which is one dimension higher than the input). In our study, we focus on continuous wavelet transforms (CWT) of the GDP and NTL data:

(12)WW/I(s,τ)=∫−∞∞x(t)1sψ∗(t−τs)dt

(13)WNTL(s,τ)=∫−∞∞x(t)1sψ∗(t−τs)dt

where (τ) and (s) are time and scale parameters responsible for dilation and translation of the wavelet in time-frequency space. To explore the instantaneous phase information in the time-scale plane, a complex mother wavelet is required and in this study we use complex Morlet wavelets which is a complex sinusoid modulated by a Gaussian envelope:

(14)ψ(t)=π−14⁡exp(iωct)exp(−12t2),

• where ω_c = 2πf_c is the central frequency of the wavelet and determines the number of oscillations of the complex sinusoid inside the Gaussian. To ensure Equation (14) is admissible as a wavelet, with a zero-mean function, we set ω_c = 6. The term π−14 ensures the wavelet has unit energy. Since the wavelet function is complex, the wavelet transform is also complex and can be divided into real and imaginary parts and the wavelet power spectrum (WPS) for a discrete series measures the variance of the W/I and NTL series across a time-scale dimension i.e.

(15)WPSW/I(τ, s)=|WW/I(τ, s)|2

(16)WPSNIL(τ, s)=|WNIL(τ, s)|2

And using (4) and (5) we can compute the Cross-Wavelet Power Spectrum (CWPS) between NTL and W/I which is analogous to the covariance between the variables in time-frequency domain.

(17)(CWPS)W/I, NTL=WW/I, NTL=|WW/I, NTL|

The wavelet coherency is referred to as the ratio of the cross-spectrum to the product of each series spectrum and can be thought of as the local correlation between the pair of time series in time-frequency space i.e.

(18)Rn(s)=|S(WW/I,NTL)|[(S|WNTL|2)(S|WW/I|2)]12

where 0 ≤ Rn(s) ≤ 1 and S is a smoothing operator in both time and scale. To further distinguish between negative and positive correlations between a pair of time series as well as identify lead–lag causal relationships between the variables, we explore phase difference dynamics through a complex number which is parametrized in radians:

(19)ϕNTL,W/I=tan−1(Ι{WNTL}ℜ{WNTL}).

where ϕ_NTL,W/I is bound between π and −π which encompasses all possible lead–lag synchronizations between the time series in a time-frequency plane.

4. Empirical findings

4.1 Descriptive statistics and correlation analysis

Table 1 presents the descriptive statistics for the data from which we observe some stylized facts on the distribution of nightlight and poverty time series for African countries. For instance, the W/I series has an average of 0.464 with a low standard deviation reflecting the fact that the concentration of “wealth-gap” in Africa is around its mean whilst the NTL data produces an average of 75 with a very high standard deviation of 234.42 showing the high dispersion of the individual light data points from its mean. Further note that both NTL and W/I are positively skewed as their averages exceed their medians implying that the data points are biased towards their minimum values and this is more extreme for nightlight data. Lastly, note that the positive values on the kurtosis estimate exceed 3 for both series implying that the series have leptokurtic distributions with fat tails whilst the J-B statistics provides overall evidence of the variables being non-normally and asymmetrically distributed. In this case, traditional linear estimators may be or may not be sufficient to capture the co-movements underlying the asymmetrically distributed data.

4.2 Preliminary analysis

Prior to the main analysis, we present the findings from our preliminary analysis in which we perform panel regression analysis between NTL and W/I using five different estimators namely, panel ordinary least squares (POLS) models, fixed effects POLS (FE-POLS), random effects POLS (RE-POLS), fully modified POLS (FM-POLS), dynamic POLS (D-POLS) and pooled mean group (PMG) estimators.

The results reported in Table 2 show that whilst the POLS estimator produces negative and significant coefficients, the remaining estimators (i.e. FE-POLS, RE-POLS, FM-POLS, D-OLS and PMG) produce their expected positive and significant estimates whereas the PMG estimators produce a negative and significant estimate. Notably, the positive estimates obtained from the majority of the estimators are in sync with those previously obtained by Noor et al. (2008) and Mveyange (2015) for similar African data albeit these former works rely on cross-sectional estimates.

Note that the estimators used in this preliminary analysis are panel-based and therefore do not account for heterogeneities existing amongst the different African countries. Moreover, there may exist asymmetries in the nightlight–poverty relationship which may be reflected in the form of time-and-cyclical variations in the country-specific data which cannot be captured by these traditional estimators. In the following section, we present the findings from the wavelet coherence analysis which addresses these empirical shortcomings.

4.3 Wavelet coherence results

We now discuss the findings from the wavelet coherence analysis for the individual African countries which describes the dynamic correlations between NTL and W/I across a time-frequency space. The results of wavelet coherence analysis are presented in “heat maps” which describe these correlations across four dimensions. Firstly, the time dimension is measured along the horizontal axis of the wavelet plot. Secondly, the frequency or cyclical dimension is measured along the vertical axis. Thirdly, the colour contours within each heat map describe the strength of the correlation between NTL and W/I with cooler (warmer) colours indicating weaker (strong) synchronizations across time-frequency plane. Lastly, the phase dynamics, indicated by the arrows within the colour contours describe the lead–lag synchronizations between the series which gives information on the sign of the relationship (negative or positive) and causal dynamics (is NTL leading or lagging W/I).

From the wavelet coherence plots, there are four possible outcomes which can occur within the time-frequency space:

1. The time series are in-phase or positively correlated with W/I leading NTL if the arrow orientation are ↑, ↗ and →

2. The time series in-phase or positively correlated with W/I lagging nightlight intensity if the arrow orientation is ↘

3. The series are anti-phase or negatively correlated with W/I leading NTL if the arrow orientation are ↓, ↙ and ←

4. The series are considered anti-phase or negatively correlated with W/I lagging NTL if the arrow orientation is ↖

The 5% significance level is represented by the faint white lines surrounding the colour contours whilst the inverted cone shape is the “cone of influence” which accounts for edge effects.

It is important to note that we are interested in identifying countries which have in-phase dynamics since this is consistent with the expectation of a positive correlation between NTL and W/I. We, therefore, group the obtained wavelet plots into two categories to facilitate our discussion of the findings.

Firstly, we find 14 countries which produce in-phase (positive) correlations throughout the entire time domain (Burkina Faso, Burundi, Cote D'Ivoire, Ethiopia, Guinea, Malawi, Mauritania, Mauritius, Rwanda, Sierra Leone, South Africa, Seychelles, Togo and Zambia). The most dominant frequencies are at 8–16 year cycles. The wavelet plots for these countries are reported in Appendix 1.

Secondly, we find 25 countries which produce anti-phase (negative) correlations through some or a majority of the time period (Angola, Benin, Botswana, Chad, Cameroon, Cape Verde, Central African Republic, Comoros, Gabon, Ghana, Congo, DRC, Guinea-Bissau, Kenya, Lesotho, Madagascar, Mali, Namibia, Niger, Nigeria, Senegal, Sudan, Tanzania, Uganda and Zimbabwe). The wavelet plots for these countries are reported in Appendix 2.

Altogether, we find that approximately a third of African countries produce positive and significant correlations across time-frequency domain and notably most of these countries have faced civil wars during the period of observations (i.e. Burundi civil war (1993–2005), CAR bush war (2004–2007), Ivorian civil war (2002–2007; 2010–2011), DRC war (1996–2003), Ethiopia (various), Nigeria (Boko Haram conflicts since 2009), Sierra Leone civil war (1991–2002)) and these periods of conflict are dominated by high-frequency oscillations in the wavelet plots. As noted by Witmer and O'Loughlin (2011), nightlights are very visible during wars and periods of conflict especially where there are large fires and large refugee movements. It is, therefore, easier to trace poverty patterns in war/conflict-prone African countries using nightlight data as lit areas tend to decline (recover) during periods of conflict (peace) which correspond to declining (recovering) economic performance (Li and Li, 2014).

5. Conclusions

This study sought to evaluate the suitability of nighttime luminosity as a proxy for measuring poverty in Africa by examining the empirical relationship between DMSP-OLS nightlight intensity and the wealth-to-income ratio for 39 African countries between 1992 and 2012 using rigorous empirical analysis. In our preliminary analysis, we estimate nightlight–poverty regressions using six conventional estimators (i.e. POLS, FE-POLS, RE-POLS, FM-POLS, D-POLS, PMG) which produce the expected positive and statistically significant relationship between the variables which concurs with the findings of previous literature. However, re-examining the relationship for individual African countries using wavelet coherence analysis which depicts the synchronization between the series in time-frequency space, our findings are not as optimistic as we observe that only 13 out of the 39 African countries produce their hypothesized positive correlation across time-frequency domain.

Altogether, we conclude that less than a third of our sample data are at liberty to use regression analysis to create reliable synthetic poverty-based time series using DMSP-OLS nightlight intensity with most of the countries being concentrated in Central African region close to the equator. Despite our findings suggesting that nighttime intensity does not have good correlation properties with wealth-to-income ratio, we do not dispose of the idea of using remote sensing methods to measure economic development. To this end, we propose two avenues for future research endeavours. Firstly, future studies can focus on other satellite sensor light measures such as the Visible Infrared Imaging Radiometer Suite day–night band database which has longer time series data in higher frequency. Secondly, researchers could consider creating hybrid synthetic measures of poverty between nightlight intensity and other remote sensing images which can incorporate biodiversity, land cover and vegetation change features in measuring poverty in Africa.

Figures

Figure A1

Wavelet coherence plots dominated by in-phase co-movements

Figure A1

Figure A1

Figure A2

Wavelet coherence plots dominated by anti-phase co-movements

Figure A2

Figure A2

Figure A2

Figure A2

Table 1

Summary statistics and correlation coefficient

	NTL	W/I
Mean	79.87134	4.64905
Median	25.57545	4.521738
Maximum	1685.208	10.44018
Minimum	0.190523	0.373158
Std. Dev.	234.4203	1.136683
Skewness	5.412865	0.156177
Kurtosis	32.43515	5.948398
Jarque-Bera	34426.85	307.6715
Probability	0.00	0.00
Sum	67091.93	3821.202
Sum Sq. Dev.	46105471	1084.028
Observations	840	840

Table 2

Long-run estimates

Estimator	Coefficient	t-statistic	p-value	R2
POLS	−24.89	−3.52	0.00***	0.02
FE-POLS	9.77	7.76	0.00***	0.99
RE-POLS	9.72	7.77	0.00***	0.07
FM-POLS	10.10	6.03	0.00***	0.99
D-POLS	18.38	7.82	0.00***	0.99
PMG	0.23	2.97	0.00***	0.99

Note(s): ^*, ^** and ^*** denote 1%, 5% and 10% significance levels, respectively