Poverty Targeting with Imperfect Information

Juan C. Yamin

arxiv: 2506.18188 · v2 · submitted 2025-06-22 · 💰 econ.EM

Poverty Targeting with Imperfect Information

Juan C. Yamin This is my paper

Pith reviewed 2026-05-19 07:47 UTC · model grok-4.3

classification 💰 econ.EM

keywords poverty targetingempirical Bayestargeting errorsantipoverty transfersimperfect informationstatistical decision problemdeveloping countriestransfer allocation

0 comments

The pith

Treating noisy income estimates as exact when allocating antipoverty transfers is inadmissible, and a nonparametric empirical Bayes rule using posterior poverty gap distributions performs better.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper frames the allocation of fixed-budget antipoverty transfers as a statistical decision problem when only estimated incomes are available. It shows that the standard plug-in rule, which inserts point estimates directly into the allocation, is inadmissible because it ignores estimation uncertainty. The author develops a nonparametric empirical Bayes targeting rule that assigns transfers according to posterior distributions of true poverty gaps. Simulations using household survey data from nine African countries demonstrate that this rule reaches substantially more poor households and improves poverty reduction compared to plug-in OLS and machine-learning approaches. A sympathetic reader would care because small improvements in targeting can stretch scarce public resources to reduce more poverty within the same budget and no-taxation rules.

Core claim

The paper claims that the standard plug-in rule is inadmissible and develops a nonparametric empirical Bayes targeting rule that assigns transfers using posterior distributions of true poverty gaps. Although the budget and no-taxation constraints make the targeting rule nonsmooth, Bayes regret is governed by the accuracy of the posterior functionals that determine the oracle allocation. In simulations, the empirical Bayes rule reaches substantially more poor households and systematically improves poverty reduction relative to plug-in OLS and machine-learning benchmarks.

What carries the argument

Nonparametric empirical Bayes targeting rule that assigns transfers using posterior distributions of true poverty gaps to minimize a poverty-targeting loss subject to budget and no-taxation constraints.

If this is right

The empirical Bayes rule reaches substantially more poor households than plug-in methods within the same budget.
It systematically improves poverty reduction relative to OLS and machine-learning plug-in benchmarks.
Bayes regret depends primarily on the accuracy of the posterior functionals rather than the nonsmoothness of the constraints.
Policymakers can translate noisy income estimates into feasible transfers without direct taxation by incorporating posterior uncertainty.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could extend to other proxy-based allocation problems such as health or education subsidies where estimates carry similar error.
National-scale implementation would require reliable survey sampling to estimate the necessary posteriors for each region.
The decision-theoretic framing suggests that pure prediction accuracy is not the right objective for policy rules under constraints.
If the method generalizes, it could inform the design of targeting systems in additional low-income settings beyond the nine countries studied.

Load-bearing premise

The accuracy of the estimated posterior distributions of poverty gaps is sufficient to approximate the oracle allocation even when budget and no-taxation constraints make the rule nonsmooth.

What would settle it

A test on held-out household survey data from the same nine African countries in which the empirical Bayes rule fails to reach more poor households or reduce poverty more than the plug-in OLS or machine-learning rules under identical budgets.

read the original abstract

A key challenge for targeted antipoverty programs in developing countries is that policymakers must rely on estimated rather than observed income, which leads to substantial targeting errors. The policy problem is not only to predict income, but to decide how noisy income estimates should be translated into feasible transfers. I formulate this as a statistical decision problem in which a policymaker chooses transfers to minimize a poverty-targeting loss subject to a fixed budget and a no-taxation constraint. I show that the standard plug-in rule, which treats estimated incomes as true, is inadmissible. I develop a nonparametric empirical Bayes targeting rule that assigns transfers using posterior distributions of true poverty gaps. Although the budget and no-taxation constraints make the targeting rule nonsmooth, Bayes regret is governed by the accuracy of the posterior functionals that determine the oracle allocation. In simulations using household survey data from nine African countries, the empirical Bayes rule reaches substantially more poor households and systematically improves poverty reduction relative to plug-in OLS and machine-learning benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows plug-in targeting rules are inadmissible under budget and no-taxation constraints and that a nonparametric empirical Bayes rule using posterior poverty gap distributions improves outcomes in simulations from nine African countries.

read the letter

The main point is that treating estimated incomes as exact produces an inadmissible targeting rule, while the nonparametric empirical Bayes approach that works with posterior distributions of poverty gaps reduces Bayes regret and reaches more poor households under the stated constraints. The simulations back this up across nine countries with real survey data, showing systematic gains over OLS and machine-learning plug-ins. What is new is the explicit framing of the problem as a constrained statistical decision problem where the loss is poverty targeting and the constraints are fixed budget plus no taxation. That setup makes the optimal rule nonsmooth, but the paper claims regret is still driven by accuracy of the posterior functionals rather than the nonsmoothness itself. The inadmissibility result for the plug-in follows from standard decision theory once the loss and constraint set are fixed, and the stress-test note confirms no internal contradictions appear in the argument. The cross-country evidence is the strongest part here because it uses actual household surveys rather than toy data. A minor soft spot is that the abstract leaves the exact nonparametric estimator and any sensitivity checks implicit, so referees will want to see the implementation details and whether the gains hold under different bandwidths or sample sizes. The work is aimed at people who design or evaluate targeted transfers in developing countries and who already think in terms of statistical decision theory. A reader focused on improving antipoverty spending with noisy income data will get concrete simulation results and a clean theoretical reason to move beyond plug-in methods. I would send this to peer review. The combination of the inadmissibility claim and the multi-country evidence is solid enough to merit referee time even if revisions are needed on the estimator details.

Referee Report

2 major / 2 minor

Summary. The paper formulates poverty targeting under imperfect income information as a statistical decision problem: a policymaker selects transfers to minimize a poverty-targeting loss subject to a fixed budget and no-taxation constraint. It proves that the standard plug-in rule (treating estimated incomes as true) is inadmissible and develops a nonparametric empirical Bayes rule that assigns transfers based on posterior distributions of true poverty gaps. Simulations on household survey data from nine African countries show that the empirical Bayes rule reaches more poor households and yields better poverty reduction than plug-in OLS or machine-learning benchmarks.

Significance. If the inadmissibility result and the claim that Bayes regret remains controlled by posterior functional accuracy (despite nonsmoothness induced by the constraints) hold, the paper supplies a decision-theoretic justification for moving beyond plug-in methods in development policy. The multi-country simulation evidence provides concrete, falsifiable support for improved targeting performance under realistic constraints.

major comments (2)

[§3] §3 (regret decomposition): the argument that Bayes regret is governed solely by the accuracy of the posterior functionals determining the oracle allocation, even though the budget and no-taxation constraints render the targeting rule nonsmooth, requires an explicit bound or decomposition showing that the nonsmoothness does not introduce additional terms that depend on the estimation error distribution.
[§4] §4 (inadmissibility proof): the demonstration that the plug-in rule is inadmissible is central to the contribution, yet the text provides only a high-level decision-theoretic argument; a self-contained derivation that explicitly uses the loss function, constraint set, and posterior structure would strengthen the claim.

minor comments (2)

[Table 2, Figure 3] Table 2 and Figure 3: the reported gains in households reached and poverty reduction should include standard errors or bootstrap intervals so readers can assess whether the improvements over OLS and ML benchmarks are statistically distinguishable.
[§3] Notation: the definition of the posterior functional that enters the oracle allocation is introduced without a numbered equation; adding an explicit display equation would improve traceability when the regret bound is later invoked.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation and for identifying points where greater rigor in the exposition would strengthen the paper. We respond to each major comment below and will revise the manuscript accordingly.

read point-by-point responses

Referee: §3 (regret decomposition): the argument that Bayes regret is governed solely by the accuracy of the posterior functionals determining the oracle allocation, even though the budget and no-taxation constraints render the targeting rule nonsmooth, requires an explicit bound or decomposition showing that the nonsmoothness does not introduce additional terms that depend on the estimation error distribution.

Authors: We agree that an explicit decomposition would make the argument more transparent. The manuscript currently notes that the oracle allocation is a functional of the posterior distributions and that the regret is therefore controlled by the accuracy of those functionals, with the constraints handled through a projection that does not introduce leading error terms dependent on the estimation distribution. In the revision we will add a formal regret bound in §3 that decomposes the excess risk and verifies that any contribution from the nonsmoothness is of strictly lower order under the maintained conditions on posterior convergence. revision: yes
Referee: §4 (inadmissibility proof): the demonstration that the plug-in rule is inadmissible is central to the contribution, yet the text provides only a high-level decision-theoretic argument; a self-contained derivation that explicitly uses the loss function, constraint set, and posterior structure would strengthen the claim.

Authors: The referee correctly observes that the current presentation is concise. The inadmissibility result follows because the plug-in rule corresponds to a degenerate posterior that ignores uncertainty and therefore cannot achieve the Bayes risk for the given loss and feasible set. In the revised version we will expand §4 to include a self-contained derivation that begins from the definition of the poverty-targeting loss, incorporates the budget and no-taxation constraints explicitly, and shows that the empirical Bayes rule strictly improves upon the plug-in rule for a positive-measure set of posterior distributions. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The derivation begins from a standard statistical decision problem with a poverty-targeting loss, fixed budget, and no-taxation constraint. The inadmissibility of the plug-in rule and the Bayes regret bound for the nonparametric empirical Bayes rule follow directly from decision-theoretic arguments once the loss and constraint set are specified; the nonsmoothness induced by the constraints is acknowledged but does not alter the claim that regret is governed by posterior functional accuracy. No equations or steps in the provided text reduce a claimed prediction or uniqueness result to a fitted parameter or self-citation by construction. The approach relies on established nonparametric estimation and simulation evidence across nine countries rather than any load-bearing self-referential step.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the method rests on standard decision-theoretic loss minimization and nonparametric posterior estimation under budget and no-taxation constraints.

pith-pipeline@v0.9.0 · 5685 in / 1049 out tokens · 35867 ms · 2026-05-19T07:47:52.891282+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the plug-in rule... is inadmissible... nonparametric empirical Bayes targeting rule... Bayes regret... governed by the accuracy of the posterior functionals
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

projection onto simplex... oracle Bayes action... max{0, z - E[μi|Ŷi] - λ/2}

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.