From National Averages to Local Realities: A Subnational Vulnerability Index to Guide Sustainable Development in Low- and Middle-Income Countries

Abstract

To achieve a sustainable society that is resilient to climate change and natural disasters, it is essential that socioeconomic vulnerabilities are addressed. However, within poor countries, hotspots of vulnerabilities are difficult to identify, as for these countries existing vulnerability measures are only available at the national level. Here, we address this issue by presenting a subnational version of the Global Data Lab Vulnerability Index (GVI), a composite index designed to monitor and analyze the human dimensions of vulnerability to climate change, natural disasters, and other kinds of shocks. The Subnational GVI (or SGVI) is available for 1260 regions across 118 Low- and Middle-Income Countries (LMICs), providing an over 10 times more detailed picture of socioeconomic vulnerability than was possible before. Decomposition analyses reveal that this higher resolution increases the observed variation in vulnerability by more than 70 percent in the poorest regions. Since 2000, total vulnerability across LMICs has declined by about 24%, but at the same time the variation in vulnerability has increased, thus highlighting the importance of subnational data. By capturing localized disparities in coping capacity, adaptive capacity, and susceptibility, the SGVI provides vital new insights and data for climate adaptation and sustainable development planning.

1. Introduction

As awareness has grown that hazard-focused approaches are insufficient to address the complex challenges posed by climate change and related disasters, increased attention has turned to the structural and institutional conditions that shape how populations experience and respond to these threats [1,2]. In this context, vulnerability—understood as the susceptibility of people and systems to harm, and their capacity to respond and adapt [3,4]—has emerged as a central concept in climate risk analysis. In this context, risk is now widely defined as the potential for adverse consequences for human or ecological systems, arising not only from the direct impacts of climate change but also from the human responses to those impacts [4]. Increased understanding of the interactions between physical and social drivers of risk—often operating across spatial and temporal scales—has led to the development of integrated frameworks for climate change risk assessment [5].

The IPCC’s Sixth Assessment Report [6,7] frames climate risk as the outcome of three interacting components: climate hazards, exposure, and vulnerability. Hazards refer to physical climate conditions—including extremes and variability—that can affect human or ecological systems [8]. Exposure refers to the presence of people, assets, ecosystems, infrastructures and livelihoods in areas that could be adversely affected by hazards. Vulnerability captures the propensity of systems to be harmed and comprises both sensitivity and adaptive capacity: the degree to which systems are affected by climate-related hazards, and their ability to cope with or adjust to them.

In this framing, reducing vulnerability—especially through improvements in adaptive capacity—is central to building more resilient and sustainable societies. Over the past decade, the focus has shifted toward better understanding the multiple and intersecting drivers of vulnerability, including income, poverty, education, health, gender disparities, infrastructure, governance, demographic change, and sociocultural norms [3,4,9,10,11,12,13,14,15]. These human dimensions are key not only to understanding differential risks, but also to guiding sustainable development strategies and adaptation policies. As a result, an increasing number of social indicators and thematic indices have been developed to facilitate international comparisons on these dimensions [13,16,17,18].

In addition to these dimension-specific metrics, composite indices have been created to capture broader patterns of vulnerability, including the INFORM Index [19], the World Risk Index [20], the ND-GAIN Vulnerability Index [21], and the GDL Vulnerability Index [22]. These indices provide integrated measures of societal capacity to withstand shocks and adapt to a changing climate. Their strength lies in their ability to synthesize many indicators into a single score, allowing users to track trends, make cross-country comparisons, and help guide policies and resources toward the most vulnerable areas [23,24,25,26]. However, these composite indices are typically limited to the national level, which restricts insight into subnational variation.

This is a critical limitation, as substantial differences in socioeconomic conditions—and thus vulnerability—can exist within countries. Evidence shows that economic development, education, health, governance, and other key factors vary widely between subnational areas [27,28,29,30,31,32]. These differences are not only relevant for large countries such as China, Nigeria, or Brazil, but also for smaller nations like Nepal, Rwanda, or Costa Rica, where local conditions can also differ substantially. Addressing climate risks in a sustainable and targeted way therefore requires a subnational index.

While expert-based indices such as INFORM, ND-GAIN, and WRI draw on a wide set of indicators (ranging from 27 to 54) and aim to cover the full spectrum of exposure, sensitivity, and adaptive capacity, their complexity and reliance on expert weighting makes them less suited for disaggregated, localized analysis. In contrast, the GDL Vulnerability Index (GVI) offers a simple, transparent, and formula-based approach that is fully data-driven and specifically focuses on the human and societal dimensions of vulnerability. It brings together seven core dimensions—economy, education, health, gender equality, demographics, governance, and infrastructure—which are measured by eleven indicators [22].

Physical aspects (landscape, soils, natural vegetation, etc.) as well as exposure-related aspects are not included. Neither are human–environment interactions like agriculture and nature management. This focus on the human and societal aspects of vulnerability makes for a strong index, as the different aspects of human development are known to be highly correlated [33,34,35]. The GVI has demonstrated validity, with correlations between 0.77 and 0.94 with the susceptibility, coping capacity and adaptive capacity subcomponents of INFORM, WRI and ND-GAIN [22]. Because it uses a formula-based calculation method, it can be computed for any geographic unit and point in time for which the necessary input data are available. This makes it highly suitable for subnational applications and for longitudinal monitoring [36].

The central aim of this study is to use this flexibility to develop a subnational version of the GVI, called SGVI, for Low- and Middle-Income Countries (LMICs). We first build a database with information on the eleven underlying indicators for 1260 subnational regions within 118 LMICs and then use this database to construct the SGVI on a yearly basis for the period 2000–2023.

The SGVI constructed in this way is subsequently used to assess the spatial patterns and temporal trends in vulnerability across the developing world. We generate maps of the (subnational) variation in vulnerability for LMICs and demonstrate the extent to which national-level statistics obscure local realities. Using a decomposable inequality measure, we quantify how much the observed variation in vulnerability increases when moving from the national to the subnational level, for the whole developing world and for major global regions.

Our results show that by using subnational data, the observed variation in vulnerability increases by about 20 percent for the complete developing world, thus underlining the dominance of country differences on a global scale. However, at the level of major regions, the increase in observed inequality is (much) higher and runs to over 70 percent in Sub-Saharan Africa (SSA), which highlights the importance of disaggregated data for targeted policy action.

We find that, on average, vulnerability has declined over the 2000–2023 period, indicating gradual improvement in underlying social conditions. However, disparities remain wide and progress are uneven, especially in SSA. The SGVI provides a powerful tool to identify where sustainable development investments are most needed and to track whether such investments translate into improved resilience. By focusing on the human dimensions of climate vulnerability, the SGVI contributes directly to the evidence base for sustainable climate adaptation and disaster risk reduction.

2. Materials and Methods

2.1. Construction of the SGVI

The GVI as it was originally developed by Smits and Huisman [22] (S&H) covers seven dimensions of socioeconomic vulnerability that are measured with eleven indicators. These dimensions are the economy, education, health, gender, demographics, governance and infrastructure. An overview of the dimensions and indicators is presented in

Table 1. Dimensions and indicators on which the GVI is based ^a.

Dimension	Indicator(s)	Indicator Unit
Economy	Gross Domestic Product per capita	Constant 2017 International $ (PPP)
	Poverty headcount ratio at USD 3.65 a day	% of population below poverty line
Education	Mean years of schooling 25+ population	Number of years
Gender	Gender Development Index	Number
Health	Life expectancy at birth	Years
Infrastructure	Access to clean water	% of population
	Access to electricity	% of population
	Mobile cellular subscriptions per 100 people	Number
Governance	Worldwide Governance Indicators	Standardized (about −3 to +3)
Demography	Urbanization	% of population
	Dependency Ratio	% of population

^a Table derived from Smits and Huisman [22]. For indicator sources see Section 2.2.

.

To create the GVI, S&H built a database with information on the eleven indicators for 189 countries. The economic dimension was indicated by GDP per capita of the region and by the poverty level of the population, measured by the poverty headcount ratio at USD 3.65 a day. Both indicators were expressed in constant 2017 International Dollars PPP and were derived from the World Development Indicators (WDI) of the World Bank [37]. The educational level of the population was indicated by the mean years of schooling of the adult (25+) population derived from the Human Development Index Database (HDID) of the UNDP [38]. Health status of the population was indicated by the region’s life expectancy at birth derived from HDID. For gender inequality, the Gender Development Index (GDI), which indicates the difference in human development between men and women in a region [39], was derived from the HDID.

For the demographic dimension, two indicators were used. The age structure of the population measured by the dependency ratio, which compares the number of young and old individuals (the dependent population) with the number of individuals in the working age population [40]; and urbanization, measured as the percentage of a region’s population residing in urban areas according to each country’s national definition. Both demographic indicators were derived from WDI. The available infrastructure of a region was measured by the percentages of households with access to electricity and clean drinking water and the number of mobile cellular phone subscriptions per 100 people, also derived from the WDI. Governance was measured by the Worldwide Governance Indicators (WGI) of the World Bank (2024) [41], which described patterns in the perception of the quality of governance.

To construct the GVI, S&H performed Principal Component Analysis [42,43] on this database to estimate weight factors for the indicators. These weight factors were subsequently transformed into an additive formula with which the GVI can be computed as an indicator running from 0 to 100, with 0 meaning lowest and 100 for highest vulnerability. The indicator weights of this formula are presented in

Table 2. Indicator weights for computing GVI with the GVI formula ^a.

Indicators	Weights
GDP per capita (GDPc)	0.00009511
Poverty headcount at USD 3.65	−0.09402847
Years of schooling	0.71349195
Gender Development Index	25.64387153
Life expectancy at birth	0.32153768
Access to clean water	0.15911601
Access to electricity	0.09070003
Phone subscriptions	0.06927752
World Governance Index	2.36889662
Dependency Ratio	−0.13611513
Urbanization	0.08743449
Constant	−23.46384145

^a Weights derived from S&H [22].

.

2.2. Subnational Indicators

To estimate the SGVI with the GVI formula, subnational data are used derived from the Global Data Lab (GDL, www.globaldatalab.org) [44], which provides freely downloadable subnational development indicators for LMICs [45,46,47,48]. The GDL databases used for this purpose are the Subnational Human Development Database (SHDD) [45], the Subnational Gender Development Index (SGDI) Database [46], the Global Data Lab Area Database (GDL-AD) [47] and the Subnational Corruption Database (SCD) [48], all of which were accessed on 29 September 2025. These databases were constructed by GDL by aggregation from household datasets derived from the Demographic and Health Surveys [49], UNICEF Multiple Indicator Cluster Surveys [50], population censuses distributed by IPUMS International [51] and public opinion surveys, including Afrobarometer [52], LAPOP Surveys [53], Asian Barometer [54] and some stand-alone surveys.

Five of the eleven indicators needed for the SGVI were available in the GDL databases with the required definition, so no further transformations were required. These indicators are the mean years of schooling of the adult (25+) population (MYS), life expectancy at birth (LEB), the Gender Development Index (GDI), the dependency ratio (DEP) and the degree of urbanization (URB). For the other six indicators, the definitions in the GDL databases differed somewhat from those used by S&H and regression-based prediction models were developed to translate the subnational indicators into the correct form. Both linear and nonlinear models were tested and the models with the highest explained variance were chosen. For indicators that could not have values below zero, regressions through the origin were used. To develop those models the national versions of the GDL indicators were related to their national equivalents in the GVI Database of S&H [22]. For this purpose, version 3.0 of this database was used derived from the GDL.

For subnational GDP per capita (GDPc), we used the estimate of Gross National Income per capita (GNIc) available in the SHDD of GDL as starting point. We developed a regression-based prediction model that estimated the national level of GDPc in the GVI Database on the basis of the national GNIc values in the SHDD. This model has a very high explained variance (adj. R2) of 98.1%, hence the prediction is expected to be very good. The model is:

GDPc = 1.19227195 × GNIc

(1)

A similar model was developed to estimate the poverty headcount ratio at USD 3.65 a day (Pov365) on the basis of the IPov50 poverty measure available in GDL-AD. The IPov50 poverty measure is based on the International Wealth Index (IWI), an asset-based wealth index running from 0 (no assets) to 100 (all included assets) that indicates the wealth level of households on the basis of their ownership of durables, housing quality and access to basic services [55]. IPov50 reflects the percentage of households under an IWI value of 50. Our prediction model for estimating national Pov365 on the basis of national IPov50 has a very high explained variance (adj. R²) of 94.6%. The model is:

Pov365 = 0.76877955 × IPov50 + 0.00353646 × year

(2)

For governance, S&H [22] relied on the Worldwide Governance Index (WGI) of the World Bank [56], which combines six governance indicators—voice and accountability, political stability, government effectiveness, regulatory quality, rule of law, and control of corruption. To create a subnational version of this index, we used the Subnational Corruption Database (SCD) [48] as starting point, as this is the only database with a governance indicator at the subnational level for the complete developing world. Given that the perception of corruption is one of the major indicators of the WGI, we used national versions of the Subnational Corruption Index (SCI) derived from the SCD to create a regression-based prediction model to estimate the WGI. The prediction model has a highly explained variance (adj. R²) of 83.9%. The model is:

WGI = −2.04759085 + 0.10616570 × SCI + 1.90585697 × ln((SCI − min(SCI)) + 1)

(3)

The infrastructure in a country was indicated by S&H by the percentage of the population using safely managed drinking water services (CWAT) and the percentage of the population with access to electricity (Elec). In the GDL-AD, the percentage of households with access to piped water is available (Piped). We used the national values of this indicator in the prediction model. This prediction model has a high explained variance (adj. R²) of 90.5%. The model is:

CWAT = 0.72706093 × Piped + −0.00256948 × Piped² + 0.02552638 × year

(4)

Given that in the GDL-AD the percentage of households with electricity (Electr) was available, we used the national value of this indicator to develop a prediction model to estimate the national value of Elec, the percentage of the population with electricity, on the basis of Electr. Although both definitions seem similar, they are not identical, and we estimated a prediction model. As could be expected, this prediction model has a very high explained variance (adj. R²) of 99.9%. The model is:

Elec = 0.99814693 × Electr

(5)

The communication infrastructure available in a country was indicated by S&H [22] by the number of mobile cellular subscriptions per 100 people (MCS). Given that in the GDL-AD the percentage of households with a mobile phone (PHMP) was available, we used the national value of this indicator to develop a prediction model to estimate the national value of MCS on the basis of PHMP. This prediction model has a very high explained variance (adj. R²) of 95.0%. The model is:

MCS = 0.15092665 × PHMP + 0.01030008 × PHMP²

(6)

By applying these prediction models to the subnational data derived from the Global Data Lab, subnational estimates for these indicators were generated, which were subsequently brought together with the indicators for which no estimating models were needed in the Baseline SGVI Database.

2.3. Addressing Missing Data

Given that household surveys and censuses are not held every year, part of the indicators in the Baseline SGVI Database are only available for a restricted number of years, i.e., the years in which a survey was conducted. This was the case with Pov365, MCS, Elec, CWAT, DEPR and URB. To obtain the values of these indicators for the whole period 2000–2023, we had to estimate the missing information by using interpolation or extrapolation. This estimation process was facilitated by the fact that the latest (2000–2023) GVI Database made available by S&H [22] (at www.globaldatalab.org/gvi (accessed on 2 June 2025)) contains the national values for all indicators for each year in this period, which means that only the subnational variation had to be interpolated or extrapolated [45,46,48].

For countries for which data for several points in time was available, linear interpolation was used to fill in empty years between the years for which data were available. As no other information for the missing years was available, presuming that in each of the in-between years the indicator values changed in a similar way does seem a reasonable assumption, which has also been used in other data projects [45,46,48]. If the subnational indicator values were only available for an earlier or a later year, extrapolation had to be used. When extrapolating, no change in the composition of subnational values was assumed, as the closest available subnational distribution seems to offer the most likely prediction, given that subnational variation is a sticky phenomenon [45,48,57].

2.4. Standardization Around National Values

The interpolated and extrapolated indicator values for the period 2000–2023 were subsequently standardized around their national values in the GVI Database for this period. To do this, they were rescaled in such a way that their population weighted national means for a given year were exactly the same as their national values in the GVI Database. In this way, indicators were obtained that at the national level are exactly equal to the values used by S&H for constructing the GVI, while at the same time their subnational distribution within the countries was in line with the variation in the subnational databases. Such a rescaling approach has been effectively applied in earlier studies [45,46,48,58]. The population data used for creating population weighted means were derived from the SHDD.

Using the described procedures, the baseline version of the SGVI Database was transformed into a version that contains values for the eleven dimension indicators and population size for each year in the period 2000–2023. This information was available for 1260 regions across 118 LMICs. Because national values for one or more indicators for the first years of the period were occasionally missing in the GVI Database, the rescaling approach could not be applied to all region–year combinations. In total, the SGVI Database therefore includes data for 32,904 region–year combinations, of which 26,557 (81%) contain complete information for all indicators. Detailed information on data availability by region, country, and year is provided in the SGVI Database available in the online Supplementary Materials. An overview of the included countries and subnational regions can also be found in Figure 1.

2.5. Constructing the SGVI

To construct the SGVI on the basis of the indicators in the SGVI Database, the values of these indicators were entered into the GVI formula developed by S&H. The indicator weights of this formula are presented in Table 2. To compute the SGVI, the indicator values had to be multiplied with their weight and summed up, as shown in Equation (1).

SGVI′ = −23.46384145 + ∑ β_n × x_n

(7)

In this equation, SGVI′ is the estimated vulnerability score, β_n the indicator weight of the nth indicator and x_n the indicator value of the nth indicator. The SGVI′ scale constructed in this way was subsequently reversed to create a scale in which higher values mean more vulnerability, with values running potentially from 0 to 100:

$$\mathrm{S}\mathrm{G}\mathrm{V}\mathrm{I}=100-\mathrm{S}\mathrm{G}\mathrm{V}\mathrm{I}\mathrm{\prime }$$

(8)

Using this approach, SGVI values could be computed for the 24,406 region–year combinations (81%), for which all indicators were available. These SGVI values were subsequently standardized around their national values in the GVI Database of S&H to be sure that their population weighted means were exactly equal to those national values.

For the combinations with one or more indicators missing, alternative formulas were used developed by S&H for situations in which not all indicators are available. For 199 combinations (0.7%), poverty was missing; for 1586 combinations (5.3%), GDI was missing; for 597 combinations (2.0%), access to clean water was missing, for 1676 combinations (5.6%), governance was missing; for 147 combinations (0.5%), the dependency ratio was missing; for 353 combinations (1.2%), education and GDI were missing; for 76 combinations (0.3%), GDI and governance were missing; for 45 combinations (0.1%), GDI and the dependency ratio were missing; for 744 combinations (2.5%), water and electricity were missing; for 18 combinations (0.1%), water and governance were missing; for 120 combinations (0.4%), electricity and phone were missing; for 72 combinations (0.2%), GDI, water, electricity and the dependency ratio were missing and for 137 combinations (0.5%), water, electricity, phone and the dependency ratio were missing.

For the situations in which one or two indicators are missing, alternative formulas developed by S&H were used. For the two situations in which four indicators were missing, the approach of S&H was used to develop new formulas. By using these alternative formulas for the combinations with missing indicators, we were able to create SGVI values for all subnational regions of the 121 countries for which we had data. These SGVI values were subsequently standardized around their national values in the GVI Database in the same way as this was achieved with the underlying indicators (see Section 2.4). In this way, SGDI values were obtained that at the national level are equal to the values of the original GVI, while their distribution within the countries aligns with the variation in the subnational indicators.

By adding these SGVI values plus the formulas used to create the index to the database, the final version of the SGVI Database was constructed, which is made available in the Supplementary Materials. It can also be downloaded from the Global Data Lab (www.globaldatalab.org) where future updates will be made available.

3. Results

3.1. Contribution of Indicators

When applying the GVI at the subnational level, it is useful to verify whether the contribution of the eleven indicators to overall vulnerability remains consistent with the original country-level formulation. Subnational data differ from cross-national data in that within-country variation may not mirror between-country variation, while unobserved national factors may influence regions within a country.

To test whether each indicator has a significant independent association with vulnerability when applied subnationally, we have estimated regression models in which the SGVI values in our database were regressed on the eleven subnational indicators. This was done for all included countries together and separately for four global regions: Sub-Saharan Africa (SSA), Latin America and the Caribbean (LAC); the Middle East, North Africa and the former Soviet States (MENAS), and central and south Asia and the Pacific (CSAP). Because the subnational regions are clustered within countries, multilevel regression models with country-specific random intercepts were used [59,60].

The results in

Table 3. Coefficients of multilevel regression models with SGVI as dependent variable and the eleven indicators as independent variables ^a.

	ALL	SSA	LAC	MENAS	CSAP
Intercept	54.20	53.45	54.49	54.32	54.47
GDPc	−1.00	−1.49	−0.91	−0.85	−0.78
Poverty365	3.02	3.17	2.61	2.73	2.45
Education	−2.31	−2.83	−2.03	−2.04	−1.81
GDI	−1.45	−1.12	−2.16	−2.00	−2.13
Life expectancy	−2.26	−2.25	−2.21	−2.20	−2.89
Clean water	−2.83	−2.85	−2.88	−2.90	−2.87
Electricity	−2.61	−2.38	−2.95	−2.63	−3.73
Phones	−1.66	−1.54	−2.13	−2.04	−1.89
Governance	−2.89	−3.85	−3.43	−3.41	−2.99
Dependency ratio	2.87	2.85	2.58	2.62	2.44
Urbanization	−2.52	−2.31	−2.59	−2.61	−2.87

^a All coefficients are significant at p < 0.0001 level.

for all LMICs combined and for each of the four global regions show that all eleven GVI component indicators remain highly significant at the subnational level after controlling for country-specific random effects. All p-values are below p = 0.0001. The size and direction of coefficients are remarkably stable across global regions, with only SSA showing some deviations. This consistency indicates that the statistical relationships on which GVI was built nationally hold robustly when applied within countries, supporting the conceptual validity of the index for subnational applications.

3.2. SGVI Level and Changes

Table 4. Average SGVI and change therein for world regions and time periods.

	All	LAC	SSA	MENAS	CSAP
Mean SGVI
2000	73.1	54.6	89.8	60.1	72.2
2011	61.9	43.2	80.0	48.0	59.7
2023	54.1	39.3	70.0	44.9	48.7
Percentage change in mean SGVI a
2000–2011	15.3	20.8	10.8	20.1	17.4
2012–2023	12.7	9.0	12.5	6.4	18.4
2000–2023	26.0	28.0	22.0	25.2	32.6

^a Percentage change computed as 100 × (highest value − lowest value)/highest value.

presents the mean SGVI values for all included LMICs and for major global regions for the years 2000, 2011, 2023. The mean SGVI level across the developing world was 73 in 2000 and decreased via 62 in 2011 to 54 in 2023. There was a substantial reduction of 26 percent over this 23-year period.

Table 4 also shows that the mean SGVI and changes therein differ substantially between world regions. Of the four distinguished regions, SSA had the highest level of vulnerability in 2000 and showed the smallest percentage improvement in vulnerability over the period 2000–2023. In 2000, vulnerability in SSA was 65% higher than in LAC, the global region with the lowest level of subnational vulnerability in all three years. Of the other two regions, MENAS had a slightly higher level of SGVI than LAC over the whole period. The fourth region, CSAP, started at a relatively high SGVI of 72 in 2000. However, the decrease in vulnerability was with 33 percent highest in this region. As a result of the differences in speed of decrease, three of the four regions reached mean SGVI scores in the 39–49 range in 2023 and only SSA remained with a mean of 70 at a very high level.

Improvement in subnational vulnerability declined on average over time, but this was due to a slowdown in the improvement in LAC and MENAS. In SSA and CSAP, the decrease continued over time, probably because of the higher starting value in 2000.

In Figure 1, SGVI values for 2000, 2011 and 2023 are displayed on world maps. For countries without data in 2000, data for the first year it was available was used. As could be expected, socioeconomic vulnerability is lowest in subnational regions of the more developed countries and highest in regions of the poorest countries. Particularly high SGVI values, shown as dark orange and red on the maps, are found in subnational regions of SSA and south and southeast Asia. In 2023, the most recent year with data, the lowest SGVI values are found in Malaysia and in subnational regions of Latin America, e.g., in Chile, Cuba, and Uruguay. The highest SGVI values can be observed in subnational regions of the poorest and least stable countries, e.g., Somalia, Congo Democratic Republic, and the Central African Republic.

Figure 1. SGVI in 2000, 2011 and 2023.

Figure 1. SGVI in 2000, 2011 and 2023.

The maps clearly show that between 2000 and 2023 the level of socioeconomic vulnerability decreased in most of the subnational regions. Whereas in 2000, 42.1% of the subnational regions had a GVI value above 75, indicating very high levels of vulnerability, in 2011, this had dropped to 31.7%, and in 2023 to 20.1%. Improvements are most substantial in regions of China, Cambodia, Laos, and Timor Leste, with SGVI value decreases of more than 30 points. However, there are also some subnational regions, mainly in Venezuela and South Sudan, where vulnerability has increased over the study period.

3.3. Decomposing Variation in GVI

To study the variation in SGVI across the developing world, two inequality measures are used. The well-known GINI coefficient [61] is used to gain an idea of the size of the variation, as its values are comparable to estimates for other kinds of indices, like income or wealth. With GINI, we can also observe to what extent the variation in vulnerability has changed over time since 2000.

Besides GINI, we also use the Theil coefficient [62], because it is an additively decomposable inequality index. This means that it can be used to observe how much the variation in vulnerability increases, when besides the variation between countries also the variation between regions within countries are observed. In this way, we can gain insight into the importance of using subnational instead of national data for studying vulnerability in LMICs.

The GINI coefficient for the total variation in SGVI across the developing world and for major global regions is presented in

Table 5. GINI coefficient and increase in inequality when including subnational level for all LMICs and for major global regions.

	GINI			Increase in Observed Inequality When Including Subnational Level
Region	2000	2011	2023	2000	2011	2023
All	0.15	0.19	0.19	16.3	18.2	20.6
SSA	0.07	0.11	0.13	83.4	77.6	72.7
LAC	0.13	0.15	0.15	30.8	39.4	56.4
MENAS	0.10	0.11	0.11	31.9	30.0	24.0
CSAP	0.12	0.15	0.16	36.3	35.7	27.4

for the years 2000, 2011 and 2023. Over the period 2000–2011, total inequality among all included countries and subnational regions increased from 0.15 to 0.19, after which it remained stable at that value until 2023. This level of inequality is higher than in the four separate world regions, which show values between 0.07 and 0.16. This difference is not surprising as the countries within those global regions tend to resemble each other more than they resemble countries in other global regions. The higher level of variation, when all countries are compared, is also reflected in the increase in total observed inequality when subnational variation is added. At the level of all countries, this increase is with 16–21 percent rather low, thus revealing that at that level the between-country variation is dominant over the within-country variation.

However, if we look at the four global regions separately, the picture is different. Within these regions, and particularly in SSA, we observe lower levels of inequality than among all LMICs. Thus—not surprisingly—the countries within these regions resemble each other more than they resemble countries in other parts of the world. However, the increase in observed inequality within these regions when subnational variation is added is substantially higher than among all LMICs. In SSA, taking subnational variation into account increased the observed inequality in 2000 with as much as 83 percent. Between 2000 and 2023, this percentage decreased, but in 2023, it was still considerable (73 percent). Hence, in this poorest region of the world, within-country variation plays a role of importance.

To a certain extent, this is also true in other global regions. In LAC, inequality was relatively stable (0.13–0.15), but the increase in inequality when subnational variation was added rose from 31 percent in 2000 to 56 percent in 2023. This last number is also higher than what we see in the other two regions, which both start at roughly one-third but then decreased to 24 percent (MENA and former Soviet states) or to 27 percent (rest of Asia and Pacific) when adding subnational variation.

Overall, these findings indicate that while within-country variation is important in all world regions, it is most important in the poorest regions, which is also in line with earlier research on global inequality in human development [45].

4. Discussion

Recent advances in climate risk analysis underscore the limitations of hazard-centric approaches, highlighting the importance of structural and institutional conditions that shape how populations experience and respond to climate-related threats [1,2]. In this context, vulnerability has emerged as a central concept, emphasizing the social dimensions of risk alongside physical hazards and exposure. Vulnerability is conceptualized as a function of both sensitivity and adaptive capacity and reducing it—particularly by enhancing adaptive capacity—is now widely recognized as critical to building resilient and sustainable societies [3,4].

To operationalize vulnerability, various composite indices have been developed, including the INFORM Index [19], World Risk Index [20], ND-GAIN [21], and the GDL Vulnerability Index (GVI) [22]. While these tools offer valuable cross-country insights, their national-level focus limits their ability to capture within-country (subnational) variation, which can be substantial, particularly in LMICs.

In this study, we aim to contribute to the field by presenting a subnational version of the GDL Vulnerability Index (SGVI) [22], covering 1260 regions across 118 LMICs for the period 2000–2023. Unlike the other indices mentioned above, which incorporate a large number of indicators and depend on expert evaluations, the GVI offers a transparent, formula-based, and fully data-driven measure of socioeconomic vulnerability. The index aggregates eleven indicators across seven core dimensions: economy, education, health, gender equality, demographics, governance, and infrastructure [22].

Data for constructing the SGVI were sourced from the Global Data Lab (GDL; www.globaldatalab.org). Five indicators (mean years of schooling, life expectancy, the Gender Development Index, dependency ratio and urbanization) were directly available, while the remaining indicators were estimated from the GDL data using regression-based estimation procedures. Missing years were imputed through interpolation and extrapolation techniques, and indicators were normalized relative to national values in the GVI database of S&H.

The general GVI formula, developed by S&H through Principal Component Analysis (PCA), was applied to the subnational dataset. For cases with missing indicators, alternative formulas—developed specifically for partial data contexts—were used. The SGVI thus retains compatibility with the original GVI structure, while maximizing data coverage at the subnational level, enabling a more granular understanding of vulnerability as a barrier to sustainability.

To assess the validity of applying the GVI framework subnationally, multilevel regression models with country-level random intercepts were estimated. Results confirmed that all eleven indicators have strong associations with SGVI scores across all global regions (SSA, LAC, MENAS, CSAP), with high consistency in coefficient size and direction (all p < 0.0001). This consistency supports the robustness and conceptual validity of the GVI framework and underscores its potential as a tool for informing targeted, equity-focused sustainability policies.

Between 2000 and 2023, the average SGVI score across the developing world declined significantly, from 73 to 54, reflecting a 26% improvement. However, regional disparities remain substantial. Subnational regions in Sub-Saharan Africa (SSA) consistently exhibited the highest levels of vulnerability and the smallest reductions over the study period, while such areas in Latin America and the Caribbean (LAC) maintained the lowest SGVI scores throughout. The most significant declines in subnational vulnerability were observed in Asian countries, followed by the Latin American countries and MENA.

In 2000, 46% of subnational regions had SGVI values above 75 (indicating very high vulnerability), a proportion that fell to 19% by 2023. While this trend signals positive movement toward sustainability, setbacks in regions of countries like Venezuela and South Sudan illustrate the fragility of these gains in the face of conflict and institutional breakdown.

Analysis of inequality using the GINI coefficient revealed an increase in SGVI inequality at the global level from 0.15 in 2000 to 0.19 in 2011, remaining stable thereafter. This overall variation reflects both between- and within-country disparities. Notably, within-country variation contributes significantly to overall inequality in LMICs, especially in SSA and Latin America—underscoring the importance of place-based interventions to promote inclusive and sustainable development. These findings reaffirm that national averages often obscure substantial intra-country disparities, particularly in low-income regions. The results in this paper thus highlight the importance of spatially disaggregated data in designing effective vulnerability reduction policies aimed at improving sustainability.

When using the SGVI, it should be kept in mind that the underlying method has certain restrictions. First, interpolation and extrapolation were used to impute missing years for the indicators. Although national mean values are standardized around the national GVI values in the GVI Database, and subnational variation is known to be relatively stable over time [45,46,48,57], in the case of extrapolation over longer periods the subnational picture may still be distorted to some extent. Second, the GVI formula was created by S&H using PCA, with the first component—reflecting the common variation in the eleven indicators—selected as the vulnerability index. While this component explained a substantial share (about two-thirds) of the total variance of the indicators and all indicators contributed meaningfully, there remains the possibility that certain aspects of vulnerability are not fully captured. Some relevant indicators may be missing, or those included may only partially cover their dimension. We consider this risk limited, since S&H based their selection on an extensive literature review and the GVI correlates strongly with other indices based on a much larger number of indicators. Nevertheless, it remains important to approach the SGVI critically, as it is the first vulnerability index with subnational coverage across the developing world, and there are currently no other indices available at this scale to compare its performance with.

In conclusion, sustainability cannot be achieved without addressing vulnerability at the subnational level. Tools like the SGVI are critical for aligning climate resilience efforts with long-term environmental stewardship. By providing spatially disaggregated insights, the SGVI equips policymakers, researchers, and development practitioners with the evidence needed to design interventions that not only reduce risk but also advance sustainability.