How many children, men, and women are living in poverty globally?

An approach combining structural and machine-learning models

Dean Jolliffe, Christoph Lakner, Daniel Mahler, Samuel Tetteh-Baah

World Bank

2026-04-24

Disclaimer and Acknowledgements

The findings, interpretations, and conclusions expressed in this presentation are entirely mine. They do not necessarily represent the views of the World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. I gratefully acknowledge financial support from the UK government through the Data and Evidence for Tackling Extreme Poverty (DEEP) Research Programme.

Introduction

Motivation

Standard per-capita measures of global poverty assume:

Equal distribution of household resources
All consumption is private
Identical needs across individuals

Motivation

Standard per-capita measures of global poverty assume:

Equal distribution of household resources
All consumption is private
Identical needs across individuals

But the evidence shows:

Inequality within households (men vs. women vs. children)
Sharing of resources (e.g., housing) thus economies of scale
Different needs (adults vs. children)

This is the first global study to integrate all three facts.

Identifies poor individuals (not just poor households)
Improves poverty monitoring and policy targeting

Key findings

Of the 802 million extreme poor in 2024:
- Women: 39%
- Men: 30%
- Children: 31%
- Children in Sub-Saharan Africa: 26%
- Women in Sub-Saharan Africa: 21%

Children and women especially in Sub-Saharan Africa are disproportionately poor.

A quarter of a million children are poor.

74 million more women than men are poor.

Millions living in extreme poverty, 2024

Presentation outline

How do we incorporate the three facts into poverty measurement?
- Structural models, Machine learning, Derive global distribution (inequality)
- Derive a scaled, adult-equivalized poverty line (sharing, different needs)

Global and regional poverty estimates

Conclusion

Structural models of collective household decision-making

Structural model: core idea

Household resources are not equally shared.

Each individual type $i \in \{m, f, c\}$ receives a share $\eta_i$ of total private resources $C$:

\[ x_i = \eta_i C \tag{1} \]

Individual consumption $x_i$ is not observed.

Instead, use assignable goods (e.g., clothing).

The individual’s budget share $w_g$ on an assignable good $g$ is:

\[ w_g = a_g + b_g \ln x_i \tag{2} \]

Engel curve above links spending to resources
$b_g$: sensitivity to income (i.e., preferences)

Combining (1) and (2), household budget share

\[ W_g = \eta_i \left[a_g + b_g \ln(\eta_i C)\right] \tag{3} \]

Differentiating (3) with respect to $\ln C$ gives:

\[ \frac{\partial W_g}{\partial \ln C} = b_g \eta_i \tag{4} \]

Slope = preferences × resource shares

Under Similarity Across People (SAP), resource shares are proportional to Engel curve slopes:

\[ \eta_i = \frac{g_i}{g_m + g_f + g_c} \tag{5} \]

Higher spending response ⇒ larger share

Empirical strategy of this paper

Combine:

structural estimates of intra-household resource shares for 45 countries

with

machine-learning predictions for the remaining countries

to construct an individualized global consumption distribution and re-estimate poverty.

Country group	Women /Men	Children /Adult	Obs	Pop. (%)
East Asia & Pacific	0.70	0.14	2	0.2
Europe & Central Asia	0.90	0.24	5	3
Latin America & Caribbean	0.88	0.45	11	77
Middle East & North Africa	0.72	0.20	4	42
South Asia	0.86	0.38	2	97
Sub-Saharan Africa	0.79	0.26	21	70
Low-income countries	0.81	0.27	15	61
Lower-middle-income	0.83	0.33	15	75
Upper-middle-income	0.84	0.40	13	21
High-income countries	1.03	1.03	2	2
All	0.83	0.34	45	42

Source: Aminjonov et al. (2025)

Ratio of women-to-men resource shares from Aminjonov et al.

Machine-learning models

Outcome variables from Aminjonov et al.

Prediction variables from different data sources

Source	Description	Countries	Period	Variables
World Bank	Gender Data Portal	217	1960–2025	1392
World Bank	Poverty & Inequality Platform	218	1977–2026	190
Gethin & Saez (2025)	Hours worked by gender/age	159	1900–2023	189
Gallup Polls	Cultural values and beliefs	165	2006–2024	135
United Nations	Gender Development Index	195	1990–2023	15
World Values Survey	Cultural values and beliefs	65	2017–2023	13
Pew Research Center	Population shares of religions	195	1990–2024	12
World Bank	World Governance Indicators	206	1996–2023	6
World Economic Forum	Global Gender Gap Index	157	2004–2021	5
World Inequality Lab	Female labor income share	211	1990–2023	1
Jolliffe et al. (2025)	Food share of consumption	167	2022–2022	1
All data		218	1900–2026	1959

XGBoost is the preferred ML model

Linear models: LASSO, Ridge, Elastic Net
Non-linear models: Random Forest, XGBoost

Why XGBoost?

Model builds trees sequentially
Each tree corrects previous errors

\[ \hat{y}_i^{(t)} = \hat{y}_i^{(t-1)} + f_t(x_i) \tag{6} \]

Captures interactions
Handles missing data
Controls risk of overfitting (p=1935 ≫ n=45)

Regularization addresses overfitting

The regularization term is given as: \[ \Omega(f_t) = \gamma T + \frac{\lambda}{2} \sum_{j=1}^{T} w_j^2 + \alpha \sum_{j=1}^{T} |w_j| \tag{7} \]

$\gamma$ → limits number of leaves (simpler trees)
$\lambda$ → shrinks predictions (stability)
$\alpha$ → keeps only strong signals (sparsity)

Feature trimming, hyperparameter tuning, model selection

Grid search (128 specifications):
- top_k ∈ {10,15,20,25}
- learning rate, η ∈ {0.03, 0.05}
- maximum tree depth ∈ {2,3}
- λ ∈ {1,5}, α ∈ {0,0.5}, γ ∈ {0,0.5}
Model selection:
- Leave-One-Out CV (LOOCV)
- Optimize out-of-sample R²
Key results:
- Best model: 15 predictors
- Inequality by gender: R² ≈ 0.52
- Inequality by age: R² ≈ 0.70

Top predictors of women-to-men resource share
Variable	Gain
Female GNI per capita (2021 PPP$)	0.200
Women’s employment rate	0.106
Female vocational enrollment (% of secondary)	0.101
Share of men who believe religion is important	0.083

Top predictors of children-to-adults resource share
Variable	Gain
GNI per capita, Atlas method (current US$)	0.323
Female road traffic injury mortality (per 100,000)	0.126
Life expectancy at birth, female (years)	0.109
Government effectiveness: estimate	0.033

Note: Gain reflects each variable’s contribution to reducing prediction error.

Individual-level global distribution and scaled poverty line

From Ratios to Resource Shares

Step 1: Define inequality ratios

\[ \frac{w_{shr}}{m_{shr}} = wm_{gap} \qquad\qquad \frac{c_{shr}}{0.5(m_{shr} + w_{shr})} = ca_{gap} \tag{8–9} \]

Step 2: Resource constraint, total private household expenditure/resources:

\[ T_{prv} = M \cdot m_{shr} + W \cdot w_{shr} + C \cdot c_{shr} \tag{10} \]

Step 3: Solve system (3 equations, 3 unknowns)

\[ m_{shr} = \frac{T_{prv}} {M + wm_{gap}\cdot W + 0.5\cdot ca_{gap}\cdot(1+wm_{gap})\cdot C} \tag{11} \]

\[ w_{shr} = wm_{gap}\cdot m_{shr} \qquad c_{shr} = 0.5\cdot ca_{gap}\cdot(1+wm_{gap})\cdot m_{shr} \tag{12-13} \]

Welfare Aggregation and Equivalent Poverty Line

Total household welfare is decomposed as:

\[ T = T_{pub} + T_{prv} \tag{13} \]

\[ T_{pub} = welf_{pc}\,(1 - pvt)\,N \tag{14} \]

\[ T_{prv} = Nwelf_{pc}\,pvt\,N/A \tag{15} \]

Note:
- Household size, $N = W + M + C$
- Adult-equivalent size, $A = W + M + 0.6C$
- $welf_{pc}$: welfare per capita (2021 PPP$/day)
- $pvt$: private share of household welfare

Allocation:
- Public consumption enjoyed by everyone
- Private consumption allocated via Eqs. (8-10)

Adjustments account for:
- Economies of scale
- Different needs of children and adults

Scaled, adult-equivalized poverty line $z$ solves:

\[ F(z) = \int_0^z f(y(x))\,dx = P^0 \tag{16} \]

where:
- $f(y(\cdot))$: transformed global distribution
- $z$: equivalent poverty line
- $P^0$: global per-capita poverty rate (11%)

Poverty lines used:
- International poverty line: $3.00 per capita
- Equivalent international poverty line: $7.50 per adult [60% for children (i.e. under 14)]

Global and regional poverty estimates

Extreme poverty rate by gender, age, region - per capita approach

Global poverty rates, 2024
- Women: 9%
- Men: 8%
- Children: 21%
- All: 11%
Men and women have similar poverty rates.
Children are much poorer than adults.

Extreme poverty rate by gender, age, region - alternative approach

Global poverty rates, 2024
- Women: 12%
- Men: 9%
- Children: 15%
- All: 11%
Child poverty falls by 6pp.
Still, children are poorer than adults (Africa effect).
Women face higher poverty risk than men.

Millions of extreme poor by gender, age, region - alternative approach

Of the 802 million extreme poor in 2024:
- Women: 39%
- Men: 30%
- Children: 31%
- Children in Sub-Saharan Africa: 26%
- Women in Sub-Saharan Africa: 21%

Children and women especially in Sub-Saharan Africa are disproportionately poor.

A quarter of a million children are poor.

74 million more women than men are poor.

Conclusion

Global poverty looks different once we move from poor households to poor individuals.

The paper’s central conclusion is:

women are substantially more likely than men to live in extreme poverty
child poverty is the highest, even after accounting for sharing, inequality, and needs
the demographic and regional composition of global poverty changes materially once we account for sharing, inequality, and needs

Thank you

Comments welcome

Samuel Tetteh-Baah
World Bank

stettehbaah@worldbank.org

Important limitations

structural estimates are available for only 45 countries
machine-learning prediction is informative, but imperfect
country-level inequality parameters are applied uniformly within countries
the current version focuses on 2024, not trends

Still

Even with these constraints, the paper provides a meaningful first global picture of poverty by age and gender under more realistic household assumptions.

How many children, men, and women are living in poverty globally?

Disclaimer and Acknowledgements

Introduction

Motivation

Motivation

Key findings

Presentation outline

Structural models of collective household decision-making

Structural model: core idea

Empirical strategy of this paper

Ratio of women-to-men resource shares from Aminjonov et al.

Machine-learning models

Outcome variables from Aminjonov et al.

Prediction variables from different data sources

XGBoost is the preferred ML model

Why XGBoost?

Regularization addresses overfitting

Feature trimming, hyperparameter tuning, model selection

Ratio of women’s share to men’s share - structural model

Ratio of women’s share to men’s share - structural model + ML model

Ratio of children’s share to adults’ share - structural model + ML model

Individual-level global distribution and scaled poverty line

From Ratios to Resource Shares

Welfare Aggregation and Equivalent Poverty Line

Global and regional poverty estimates

Extreme poverty rate by gender, age, region - per capita approach

Extreme poverty rate by gender, age, region - alternative approach

Millions of extreme poor by gender, age, region - alternative approach

Conclusion

Thank you

Comments welcome

Important limitations

Still