What do we know about poverty in North Korea?

The paper addresses the challenge of measuring income and poverty in North Korea, where official economic data have not been published since 1990 and available figures are unreliable. Leveraging advances in satellite-based night-time light (NTL) data, the study aims to estimate subnational GDP per capita and, from that, absolute poverty rates, in a context where traditional data sources are missing or unusable. The authors propose a method that combines monthly NPP/VIIRS night-time light intensity over built-up areas with inferred income distributions based on similarity to other countries. The study suggests that earlier attempts may have underestimated extreme poverty and documents a sharp decline in inferred income around 2015–2016 followed by recovery from 2016 to 2018, with regional divergence early in the period and convergence later.

Data and preprocessing: The study uses NPP/VIIRS Day/Night Band (DNB) monthly products (2012–2018) from NOAA/NCEI to reconstruct night-time luminosity for North Korea. Extraction is restricted to built-up areas identified by the World Settlement Footprint 2015 (DLR). Median radiance per pixel is computed to reduce outlier effects prevalent in monthly VIIRS products. A sensor-related level shift between 2016 and 2017–2018 is corrected by projecting South Korea’s 2012–2016 luminosity trend to 2017–2018 and adjusting observed 2017–2018 values; the difference (9.43%) is used as the correction factor. Visualization and regional aggregation cover the provinces Chagang-do, Hamgyong-bukto, Hamgyong-namdo, Hwanghae-bukto, Hwanghae-namdo, Kangwon-do, Pyongan-bukto, Pyongan-namdo, Pyongyang-si, and Yanggang-do.

Luminosity-to-GDP conversion: Subnational luminosity is converted to GDP (2011 PPP USD) using the empirical elasticity-based relationship between NPP/VIIRS light intensity and GDP established for Chinese prefectures (Li et al., 2013). The range of North Korean NTL values falls in the lower tail of the Chinese distribution, but the linkage is considered robust enough for estimation. Population denominators use the 2008 North Korean census by region, extrapolated to 2018 using the national population growth rate from UN WPP 2017, assuming uniform growth across regions and no internal migration.

Income distribution and poverty estimation: To estimate the share of people living under $1.90/day (2011 PPP), the authors approximate regional income distributions using Beta-Lorenz curves. Comparator economies are selected by measuring similarity to North Korea (and its regions) via Euclidean distances between normalized vectors of characteristics: (i) employment shares (agriculture, manufacturing, services), (ii) educational attainment shares (primary, secondary, tertiary), (iii) age structure, and (iv) GDP per capita (in levels or logs). Across all countries/years with available data, the k most similar observations are identified; cross-validation over 34 countries and 40 country/year observations evaluates variants with k=3…10 and weighted schemes using inverse-distance weights (powers 1–4), assessing root-mean-squared errors.

Model selection and anchoring: Validation favors the median of the k nearest neighbors; k=7 yields the best predictive performance with low tail errors. The seven most similar economies to North Korea are Romania (1999), Albania (2002), Madagascar (2002), Georgia (1996), Bangladesh (2016), Vietnam (2010), and Armenia (1996). For each comparator, mean expenditure-to-GDP per capita ratios are compiled (PovcalNet, UNU-WIDER, Poverty and Equity Database). These ratios anchor North Korean regional mean expenditure, which is then combined with the respective Lorenz curves to compute poverty rates. The final regional estimate is the median across comparators. Robustness checks include hierarchical clustering (complete linkage) on each variable vector, which performs worse than k=7 in validation.

Validation and sensitivity: RMSEs across 24 model variants are reported; fixed-k median models (especially k=7) outperform distance-weighted schemes. Sensitivity bounds on national poverty counts are provided by using different comparator distributions (e.g., Romania 1999 vs. Armenia 1996) to bracket estimates.

• The estimated national extreme poverty rate (≤ $1.90/day, 2011 PPP) in North Korea is around 60% in 2018, higher than prior macro-based estimates (~40%). This corresponds to approximately 14.9–17.0 million people depending on the assumed comparator income distribution (Romania 1999 vs. Armenia 1996).
• Night-time light-based GDP per capita indicates substantial volatility over 2012–2018: a significant decline from 2012 to 2015, with recovery starting in 2016. The luminosity-based GDP per capita for 2015 is approximately $790 (2011 PPP).
• Regional dynamics show divergence in 2012–2016 (poorer regions experiencing larger declines) consistent with the Williamson hypothesis for low-income contexts, followed by convergence in 2016–2018 as luminosity growth is higher in initially low-light regions.
• NPP/VIIRS monthly radiance required a 9.43% correction for 2017–2018 to account for a product-wide level shift, using South Korea’s 2012–2016 trend as reference.
• The temporal pattern of estimated poverty (increase to 2016, decrease thereafter) mirrors the inferred GDP per capita dynamics; variance in poverty across regions increases to 2016 and declines by 2018.
• Geographical patterning suggests higher poverty in southern regions; relatively low rates in Yanggang-do may reflect assumptions (e.g., zero inter-regional population growth differentials) rather than true conditions.
• The estimated poverty series is negatively correlated with Bank of Korea GDP growth figures for 2013–2016, lending credibility to the inferred dynamics despite methodological differences.

The study addresses the absence of reliable North Korean economic data by exploiting the observed relationship between night-time lights and economic activity. By translating monthly VIIRS luminosity over built-up areas into subnational GDP estimates and combining these with empirically derived income distributions from similar economies, the authors produce the first set of regional poverty estimates for North Korea. The findings suggest more widespread extreme poverty than previously inferred from macro regressions, and the temporal pattern—decline through 2015/16 with recovery thereafter—aligns with independent GDP growth assessments from the Bank of Korea. Regional divergence during the downturn and subsequent convergence during recovery are consistent with development theory (Williamson hypothesis). The approach demonstrates that leveraging remote sensing with principled distributional assumptions can yield informative poverty and inequality diagnostics in data-poor environments, informing humanitarian and policy assessments where traditional statistics are unavailable or unreliable.

The paper introduces a methodology to estimate subnational GDP per capita and extreme poverty rates in North Korea by integrating monthly VIIRS night-time lights over built-up areas with income distribution proxies from the most similar economies. Validation against a global sample supports using the median poverty rate from the seven closest comparators, yielding a national extreme poverty estimate of roughly 60% in 2018 and revealing pronounced volatility and shifting regional dynamics (divergence to 2016, convergence by 2018). These results indicate that extreme poverty in North Korea is likely higher than previous macro-based assessments. The framework provides a replicable approach for poverty measurement in severely data-constrained settings and can be extended with improved remote sensing products, ancillary socioeconomic data, and refined distributional modeling. Future research should focus on better anchoring luminosity-to-GDP mappings for very low-light contexts, integrating mobility data to relax uniform population growth assumptions, and enriching comparator selection with additional structural indicators.

• Validation constraint: Lack of credible, independent GDP per capita or poverty benchmarks for North Korea prevents rigorous external validation of levels; assessment relies on cross-country validation and coherence with Bank of Korea growth signals.
• Population dynamics: Regional population projections assume uniform growth and no internal migration from the 2008 census baseline, potentially biasing regional GDP per capita and convergence assessments.
• Luminosity data issues: Monthly VIIRS products can contain contamination; although medians and a 9.43% level-shift correction (2017–2018) were applied, residual measurement error may persist. The luminosity–GDP elasticity is borrowed from Chinese prefectures and may not perfectly transfer to North Korea’s structural context.
• Income distribution mapping: Poverty estimates depend on comparator selection (k=7), Beta-Lorenz curve specification, and anchoring via mean expenditure-to-GDP ratios; alternative choices shift levels (illustrated by 14.9–17.0 million range in 2018).
• Regional anomalies: Some regional results (e.g., low poverty in Yanggang-do) may reflect data/assumption artifacts (e.g., constant regional population growth, census-based inputs) rather than true conditions.
• Limited auxiliary data: Scarcity of reliable subnational socioeconomic data restricts inclusion of additional covariates that could improve comparator matching and distributional inference.