PaperlayerView exact revision on arXiv

Source-linked AI summary

What Would it Cost to End Extreme Poverty?

Roshni Sahoo, Joshua Blumenstock, Paul Niehaus, Leo Selker, Stefan Wager

arXiv:2609.02013v1econ.GNstat.ML

TL;DR

The paper asks whether extreme poverty can be eliminated more rapidly through redistribution while avoiding inequitable allocations. It formulates feasible transfer targeting as statistical learning under real-world information constraints and estimates the cost of reducing poverty to 1% in sampled countries and globally.

  • Problem

    Extreme poverty remains widespread, while directly minimizing the poverty rate can produce inequitable allocations that favor households just below the poverty line over the poorest households.

  • Method

    The paper learns transfer policies mapping observable household characteristics to nonnegative transfers, using conditional quantile functions learned with deep learning and evaluating schedules in a transductive setting.

  • Results

    $211B nominal per year would reduce the poverty rate to 1% across sampled countries using a gap-minimizing policy; this is 19% of universal basic income cost, while global extrapolation costs $304B per year, or 0.28% of global GDP.

  • Takeaways & Limitations

    Gap minimization provides weakly equitable support favoring poorer households and formal guarantees on poverty-rate reduction, while observable targeting costs less than uniform policies.

  • Takeaways & Limitations

    The estimates require incremental financing rather than diversion from existing programs, and many countries would likely also require international transfers.

Abstract

from arXiv · show

We study poverty minimization via direct transfers, framing this as a statistical learning problem while retaining the information constraints faced by real-world programs. Using nationally representative household consumption surveys from 34 countries that together account for 76% of the world's poor, we estimate that reducing the poverty rate to 1% (from a baseline of 13%) would cost $211 B nominal per year. This is 4.0 times the corresponding reduction in the aggregate poverty gap, but only 19% of the cost of universal basic income. Extrapolated globally, the results imply a cost of 0.28% of global GDP to (approximately) end extreme poverty.

1 Introduction

The paper asks whether extreme poverty can be reduced through feasible, information-constrained transfers and formulates targeting as statistical learning. It finds that targeted policies can substantially lower costs, while simpler designs and broader objectives introduce trade-offs.

  • Motivation: Extreme poverty remains a policy question because hundreds of millions of people live below very low consumption thresholds, while direct identification of poor households is infeasible.Household surveys cover only 0.001% to 0.01% of a country’s population, and regular data collection would be prohibitively expensive.
  • Approach: The paper learns transfer policies mapping observable household characteristics to nonnegative amounts while minimizing expected loss subject to a budget constraint.The approach retains the information constraints faced by real-world programs and builds on statistical decision-rule learning.
  • Policy objective: Gap-minimizing policies direct weakly greater support to poorer households and formally guarantee reductions in worst-case subgroup poverty rates.The paper emphasizes gap minimization while reporting rate-minimizing results for comparison.
  • Policy simplicity: 39% higher costs result from restricting policies to binary transfers, while practical transfer amounts that are too small cannot raise more than a few households out of poverty.Simpler structures remain feasible only when transfers are large enough and incur meaningful added cost.
  • Welfare: Gap-minimizing policies increase mean log consumption by only 1% less than policies learned to maximize welfare while remaining effective for headcount poverty.This supports their use as an approximate welfare-maximizing compromise despite the arbitrariness of any poverty line.
  • Sensitivity: 64% higher costs follow from replacing the $2.15 2017 PPP line with the $3.00 2021 PPP line, while geographic-only targeting raises costs by 45%.Reducing training data by 10% increases costs by 1%.
  • Global extrapolation: Global extrapolation implies that reducing the global poverty rate to 1% would cost $304B per year, or 0.28% of global GDP.The estimate uses results from 34 countries to predict costs where data are unavailable.

2 Targeting Transfers via Statistical Learning

The paper formulates feasible cash-transfer targeting under imperfect information as statistical loss minimization. It shows that loss choice determines both equity and poverty-reduction guarantees.

  • Policy setup: Cash-transfer policies map observable household characteristics and budgets to nonnegative transfers subject to an expected-budget constraint.The observable characteristics are proxies for living standards rather than direct consumption measures.
  • Policy setup: Proxy means testing thresholds conditional mean consumption, but it has no formal link to minimizing a particular post-transfer welfare loss.Explicit loss minimization instead targets the selected poverty measure directly.
  • Policy setup: Statistical loss minimization learns unrestricted or binary transfer policies by minimizing expected post-transfer loss under a budget constraint.Binary policies require jointly learning the nonzero transfer size and the eligibility rule.
  • Equity and loss choice: Nonconvex losses can produce policies that are not weakly equitable, whereas broad convex poverty measures yield unique weakly equitable policies under stated conditions.The poverty rate is nonconvex; FGT losses with α ≥1 satisfy the positive result under the theorem’s budget conditions.
  • Equity and loss choice: Poverty-gap minimization is weakly equitable and equivalent to minimizing the worst-case conditional poverty rate across covariate-defined population subgroups.This provides a formal poverty-rate reduction guarantee while directing weakly greater support to poorer households.

3 Data Sources and Uses

The empirical analysis learns transfer policies from nationally representative household surveys while addressing prediction, implementation, and data-timing constraints. The resulting estimates are tied to survey-period conditions and the selected observable covariates.

  • Data sources and preparation: The study selects countries with substantial poverty or large shares of the world’s poor and recent, high-quality, nationally representative LSMS-style surveys.The surveys are generally those underlying recent World Bank Poverty and Inequality Platform estimates.
  • Scope and limitations: The cost estimates describe poverty reduction at the time of the underlying surveys because living-standard information and predictors were collected contemporaneously.Applying policies in later years could reduce accuracy as poverty fluctuates over time.
  • Data sources and preparation: The analysis computes household per-capita consumption in 2017 PPP USD from detailed expenditure and self-production information, treating household members as having equal consumption.Household analysis weights are also constructed from the survey data.
  • Data sources and preparation: Covariates are chosen from plausibly verifiable characteristics to limit strategic behavior and facilitate enforcement of eligibility criteria.The final predictors are restricted to variables from the prespecified covariate-selection rubric.
  • Empirical methods: The study compares empirical rate minimization, gap minimization, and proxy means testing using survey data, with procedures prescribed in advance in a Data Use Plan.Datasets are split into geographically stratified training and test samples, and test-set budgets are verified for deployment units.
  • Empirical methods: Unrestricted gap minimization is convex, while binary policy learning uses nested optimization over transfer size and eligibility rules; rate minimization is handled through a discrete approximation.Deep learning estimates conditional quantiles or household-level loss changes, depending on the policy problem.
  • Scope and limitations: Results from Malawi carry a caveat because its 2018–2019 survey was used as a sandbox for iterative experimentation before the Data Use Plan was finalized.The paper specifically cautions how estimates from this survey should be interpreted.

4 Empirical Results

Empirical results show that learned transfer policies can substantially reduce poverty at costs below universal benchmarks, while approaching zero poverty becomes increasingly expensive. Across Togo and the full sample, unrestricted gap minimization generally performs favorably, and its welfare is close to welfare maximization.

  • 4.1 Togo Case Study: 20% or less of funds go to non-poor households under the least ambitious Togo policies, but greater poverty reduction involves a tradeoff between efficiency and remaining poverty.The comparison uses the share of initially poor households remaining poor and the share of transferred dollars not closing the poverty gap.
  • 4.1 Togo Case Study: 1% post-transfer poverty in Togo is achieved with unrestricted policies costing less than binary policies, while rate minimization offers no systematic cost advantage over gap minimization.Unrestricted policies can assign smaller transfers to households closer to the poverty line.
  • 4.2 Multi-Country Analysis: $211B per year reduces the poverty rate to 1% in all sample countries, equal to 4.0 times oracle cost and 19% of $2.15 UBI cost.The estimate has a bootstrapped 95% confidence interval of $207B–$215B.
  • 4.2 Multi-Country Analysis: 39% more cost is required by binary gap minimization than unrestricted gap minimization, while learned PMT targeting can perform respectably when transfer size minimizes the poverty gap.Setting PMT transfers to increase poor recipients’ mean consumption by 20% produces transfers too small to raise many poor households above the poverty line.
  • 4.3.2 Poverty Minimization vs. Welfare Maximization: 1% poverty-rate policies achieve welfare only 1% below welfare maximization despite noticeably different transfer-size distributions.The welfare comparison uses the same country-specific budgets.

5 Discussion & Implications

The discussion evaluates the scale, financing, and broader implications of transfers designed to reduce poverty, while noting implementation and macroeconomic boundaries. It also outlines a global extrapolation based on country-level estimates.

  • Sources of funds: 16% of GDP is the average estimated cost of reducing national poverty rates to 1% in the sample.The authors describe this as implausible for many countries to finance entirely domestically.
  • Sources of funds: The learned policies cost a similarly large share of national income after accounting for targeting challenges and imperfect information.The comparison is made with transfers that would bring households to the poverty line under perfect targeting.
  • Macroeconomic effects: Foreign financing would raise macroeconomic questions involving currency conversion and recipients’ spending or saving of local currency.These effects are explicitly beyond the scope of the analysis.
  • Macroeconomic effects: 16% of GDP is the average sample cost, compared with 7% of GDP in 2023 official development assistance flows.The authors note that exchange-rate effects could therefore be large enough to matter.
  • Implementation costs: The analysis estimates transfer costs but abstracts from administrative expenses for identifying, enrolling, and paying eligible households.Household registration and payment delivery create additional implementation costs in real-world programs.
  • Global scale: 0.27% of global GDP is the estimated 2023 cost using World Poverty Clock poverty rates and gaps.This estimate is described as close to the paper’s other global-scale estimate.

A.1.1 Finite-Data Regime

The finite-data analysis characterizes poverty-rate-minimizing policies and converts the resulting optimization into a tractable computational procedure. The key structure is α-validity, followed by grid search and fractional knapsack optimization.

  • Policy characterization: Tα(x) consists of zero, the maximum transfer c, and interior transfers where the conditional density at c − t equals α.An α-valid policy selects a transfer from this set for every covariate value x.
  • Policy characterization: An optimal deterministic poverty-rate policy exists and must be α-valid for some α ≥ 0.α-validity restricts transfers to a structured set determined by conditional densities.
  • Optimization: The inner optimization is equivalent to a multiple-choice knapsack problem over α-valid transfers for each unit.Each unit forms an item class, with transfers associated with loss and cost.
  • Algorithm: The algorithm grids plausible α values, solves the fractional knapsack problem for each grid point, and returns the policy with the lowest poverty rate.The grid is bounded using the supremum of the conditional density.
  • Optimization: The fractional relaxation yields a stochastic policy with cost no greater than the optimal deterministic α-valid policy.This relaxation is computationally efficient even though the standard multiple-choice knapsack problem is NP-hard.
  • Algorithm: The algorithm’s time complexity is |X| · K log |X| · K.

A.2.2 Finite-Data Regime

In the finite-data regime, the paper estimates the distributions needed for policy learning from training data and applies plug-in policies to unlabeled test covariates. It also develops binary-transfer and welfare-maximization procedures.

  • Finite-data setup: The finite-data procedure uses an i.i.d. training sample with outcomes and an unlabeled test sample of covariates.The estimated policy is applied to the test covariates rather than to fully labeled observations.
  • Finite-data policy: The plug-in policy runs the population algorithm with estimated conditional densities and the empirical test-set covariate distribution.
  • Density estimation: Conditional densities are estimated with an extension of Lindsey’s method using a nonparametric carrier density and a flexible exponential-family model.The likelihood is concave in the parameters and is optimized by stochastic gradient descent.
  • Welfare maximization: Welfare maximization is treated analogously to loss minimization, using a discrete grid of transfer amounts and a budget-matching dual parameter.
  • Binary policies: Binary-policy learning jointly optimizes recipient selection and the common transfer size under any decreasing poverty measure.The inner problem for a fixed transfer size is a capacity-constrained classification problem.
  • Binary policies: The binary-policy procedure is optimal over binary-valued policies for decreasing poverty measures, without requiring convexity of the loss.The paper applies it to both the poverty rate and poverty gap index.

B.2 Proof of Theorem 1

The proof studies weak equity violations for nonconvex loss functions by constructing smoothed distributions and loss functions with kernel methods. The truncated Gaussian kernel provides the distributional construction, while the Gamma kernel handles nonsmooth losses.

  • Proof strategy: The proof considers nonconvex regions where the loss is either continuously differentiable or not continuously differentiable.
  • Proof strategy: The truncated Gaussian kernel constructs distributions under which weak equity violations occur.
  • Truncated Gaussian kernel: As h → 0, the truncated Gaussian kernel converges weakly to a point mass at m.The construction also preserves stochastic dominance when m1 < m2.
  • Gamma kernel: The Gamma kernel smooths loss functions that are not continuously differentiable.

B.2.1 Continuously Differentiable Case

For a nonconvex loss, the authors construct environments where the optimal transfer policy violates weak equity, including after smoothing the loss and conditional distributions.

  • Continuously Differentiable Case: The construction uses nonconvexity to choose points where the loss derivatives differ enough to favor transferring to the higher-consumption type.The authors select y1<y2 with L′(y1)>L′(y2), then establish the relevant derivative inequality over the feasible transfer range.
  • Continuously Differentiable Case: A two-type mixture with point-mass conditional distributions can make the optimal policy assign the entire budget to the higher-consumption type.The policy is t*(x)=0 for x1 and ∆ for x2, despite x1 being stochastically poorer than x2.
  • Continuously Differentiable Case: The resulting policy is non-monotone increasing in consumption and therefore violates weak equity.This follows because t*(x2)>t*(x1) while the conditional consumption distributions are ordered by first-order stochastic dominance.
  • Continuously Differentiable Case: For sufficiently small kernel bandwidth h, every minimizer under the smoothed conditional distributions violates weak equity.The minimizer set is nonempty, and all its elements satisfy t2>t1 for budget pair (B,B′)=(∆/2,0).
  • Continuously Differentiable Case: When nonconvexity occurs where L is not continuously differentiable, Gamma-kernel smoothing produces a continuously differentiable loss that preserves nonconvexity.The smoothed loss converges to L at continuity points as the bandwidth approaches zero.

B.3.3 Weak Equity

Under convex losses, the optimal transfer policy is monotone decreasing in consumption, and incremental transfers combine consistently with earlier cumulative transfers.

  • Weak Equity: Incremental transfers are monotone decreasing in post-transfer consumption.This is the first component of the weak-equity result.
  • Weak Equity: Together, monotone incremental transfers and cumulative consistency imply the stated weak-equity property.The proof verifies the property by considering whether the initial transfer is positive or zero.
  • Weak Equity: The optimal transfer is monotone decreasing in consumption for every nonnegative multiplier λ.The policy has the form tλ(x), and stochastic dominance of conditional consumption distributions determines the ordering of transfers.
  • Weak Equity: The optimal policy under budget B equals the sum of the optimal policy under a smaller budget B′ and the incremental policy for the remaining budget B−B′.The corresponding multiplier satisfies λ(B)≤λ(B′) when B>B′.

B.4 Proof of Lemma 3

The proof reduces worst-case conditional poverty-rate minimization to an optimization with an auxiliary upper-bound parameter and shows equivalence with aggregate poverty-gap targeting.

  • Proof of Lemma 3: An auxiliary parameter λ upper-bounds the conditional poverty rate for every observable household type.The objective is the smallest λ satisfying the conditional poverty-rate constraints across all x.
  • Proof of Lemma 3: The conditional poverty constraints are separable across x, and nonnegativity restricts each feasible transfer accordingly.This separability permits considering policies that meet the constraint with the smallest transfer at each x.
  • Proof of Lemma 3: It is sufficient to consider policies tλ(x)=(c−F_Y|X=x^-1(λ))+ when optimizing the upper bound λ.Any larger feasible transfer has the same objective value and can be replaced by this policy.
  • Proof of Lemma 3: For 0<B<c, an optimal gap-minimizing policy also minimizes the worst-case conditional poverty rate because both policy forms coincide and use the full budget.The equivalence follows when the budget constraint binds.

B.5 Proof of Lemma 4

For the convex loss L(z)=(c−z)+, the proof derives the optimal transfer from stationarity and boundary feasibility, yielding the stated policy form.

  • Proof of Lemma 4: L(z)=(c−z)+ is convex, so the optimal policy can be characterized through a constrained convex optimization problem.The paper specializes the general policy problem to the poverty-gap loss.
  • Proof of Lemma 4: The first-order stationarity condition balances the conditional expected loss derivative against the multiplier λ.For a given x, the condition is d/dt E_F[L(t+Y)|X=x]+λ=0.
  • Proof of Lemma 4: The optimum either satisfies the KKT condition or lies at the boundary t=0.The nonnegativity constraint requires transfers to remain feasible.

B.6 Proof of Theorem 5

The proof uses KKT conditions to characterize feasible transfer policies and shows that every optimal deterministic policy is α-valid for some α ≥ 0.

  • A compact feasible set and continuous objective ensure a minimizer exists when transfers lie between 0 and c.This follows under Assumption 2 and the specified loss function.
  • Every minimizer satisfies the KKT conditions because the optimization problem meets linear constraint qualification.The proof introduces the Lagrangian and its derivative before listing the KKT conditions.
  • KKT policies satisfy primal feasibility, so each transfer lies in [0, c].The argument explicitly identifies E_F[t(X)] ≤ B as primal feasibility.
  • Boundary transfers are α-valid automatically, while interior transfers satisfy f_Y|X=x(c − t(x)) = λ and are α-valid with α = λ.The proof separates cases t(x) ∈ {0, c} and 0 < t(x) < c.
  • An optimal deterministic policy is therefore α-valid for some α, allowing optimization by computing t*_α for each α and selecting the best objective value.The resulting search minimizes P_F[t*_α(X) + Y < c] over α ≥ 0.

B.9 Proof of Lemma 8

The proof establishes weak rank preservation from continuity in the budget and weak equity: once two post-transfer outcomes meet, their ordering cannot reverse at higher budgets.

  • Assuming rank reversal between budgets B′ and B implies a crossing budget B0 where the two post-transfer outcomes are equal.Continuity of t(y; B) and the Intermediate Value Theorem produce B0 ∈ [B′, B].
  • Weak equity at the crossing forces the post-transfer difference at every higher budget to equal the original income difference.The proof writes the resulting difference as y1 − y2 + t(y1; B) − t(y2; B).
  • This contradicts the assumed reversal, so continuous weakly equitable policies weakly preserve ranks.The conclusion applies to policies continuous in B and satisfying weak equity.

B.10 Proof of Lemma 9

The proof characterizes the optimal policy within monotone transfer policies, then verifies that this policy satisfies weak equity and is optimal in the original class.

  • The analysis restricts attention to policies with decreasing transfers and increasing post-transfer consumption, then shows this class contains the original policy class.The optimization minimizes the poverty probability subject to the expected-budget constraint.
  • Only units below c need transfers, and an optimal policy gives t(y) = c − y to the interval of incomes lifted exactly to the poverty line.Monotonicity implies the targeted set has the form (z, c).
  • The resulting policy gives c − λ to units with y ≤ λ, c − y to units with λ < y ≤ c, and zero to units above c.This piecewise form preserves increasing post-transfer consumption.
  • The threshold λ(B) decreases as budget B increases because policy cost is strictly decreasing in λ.The derivative argument uses that F is a cumulative density function positive when λ > 0.
  • The policy family satisfies weak equity and is therefore optimal for the constrained problem.The proof checks the budget cases and concludes optimality over the original class.
  • The auxiliary objective is differentiable under Leibniz-rule conditions and strictly concave when the loss is strictly convex.For the FGT indices, the proof establishes convexity and strict concavity on [0, c].

C.3 Proof of Lemma 12

The proof handles nonconvex losses through continuity-point approximations, while the accompanying data construction defines the survey-based consumption measures and final country sample.

  • Proof of Lemma 12: A nonconvex loss violates convexity at some points, and dense continuity points allow sequences that preserve the strict violation in the limit.The construction uses xn ↓ x, yn ↓ y, zn ↑ z with zn = λn xn + (1 − λn)yn.
  • Proof of Lemma 12: For sufficiently large n, the approximating points satisfy L(zn) − λnL(xn) − (1 − λn)L(yn) > 0, proving the desired nonconvexity claim.All approximating points are continuity points and λn ∈ (0, 1).
  • Proof of Lemma 12: Differentiation and integration can be interchanged for the smoothed loss using Leibniz rule and dominated convergence.The proof supplies integrability, derivative-existence, and domination conditions before concluding continuity of the derivative.
  • Data construction: Country inclusion requires substantial poverty relevance, a recent nationally representative survey, enough poor households, and publicly available consumption data.The criteria are designed to support reliable statistical learning while covering countries influential for aggregate costs.
  • Data construction: The final sample contains 34 countries covering 76% of the world’s extreme poor after currency-conversion exclusions.Afghanistan, Somalia, and South Sudan are excluded because required conversion factors were unavailable or misaligned.
  • Data construction: Household per-capita consumption is computed from detailed expenditure and self-production aggregates, converted to 2017 PPP USD and weighted using household size.The approach treats household members as having equal consumption.

D.4 Secondary Data

This section documents the secondary data used to construct, adjust, and contextualize the poverty-transfer cost estimates, including poverty, population, currency, inflation, GDP, revenue, and aid data.

  • Secondary sources provide country-level poverty headcount rates, poverty gap indexes, population estimates, GDP, government revenue, and official development assistance.
  • Poverty headcount rates and poverty gap indexes are taken for survey years and globally for 2023, with interpolation used when necessary.
  • Currency conversion uses PPP factors and effective exchange rates, with sources including the World Bank, IMF, and private correspondence for Rwanda.
  • Venezuela’s missing 2017 conversion factor is estimated from the nearest year with complete data, assuming a stable ratio between the nominal exchange rate and PPP conversion factor.
  • Country-level Consumer Price Index data are used to adjust for inflation.
  • The section also records the PPP indicator definition as domestic currency per international dollar in PPP terms for ICP benchmarks from 2017–2021.

E Additional Exhibits

The additional exhibits document survey characteristics, country-level cost accounting and comparisons, exchange-rate sensitivity, and country-by-country comparisons across transfer-policy designs.

  • Survey data and characteristics: The survey-characteristics table reports household counts, covariate counts, and estimated versus World Bank poverty rates.
  • Cost accounting: The Togo cost-accounting table compares year-long programs targeted using survey data versus satellite imagery.
  • Scale comparisons: The country-GDP and government-revenue exhibit reports the ratio of 1% poverty-rate policy costs to GDP and government revenue in the survey year.
  • Exchange-rate sensitivity: Official-versus-effective exchange-rate exhibits show how country-level costs change when official exchange-rate estimates are used, restricting inclusion to usable cases with changes of at least 1%.
  • Policy comparisons: Country-by-country comparisons include gap minimization, rate minimization, lasso-based PMTs, universal basic income, and Universal Supplemental Income.
  • Policy comparisons: Gap minimization with restricted features performs almost identically in practice and remains more cost-effective at reducing the poverty rate, indicating algorithmic performance is the main driver.
Loading 2609.02013v1…