The Value of Prediction in Identifying the Worst-Off

Christoph Kern; Juan Carlos Perdomo; Unai Fischer-Abaigar

REVIEW 4 major objections 6 minor 37 references

The Value of Prediction in Identifying the Worst-Off

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Expanding screening access usually beats improving prediction.

desk verdict Solid theoretical extension of PAR to worst-off targeting; the empirical claim that capacity dominates prediction rests on a single improvement pathway and should be treated as provisional. read the letter →

arxiv 2501.19334 v3 pith:HLFCQD4L submitted 2025-01-31 cs.CY cs.LGstat.ML

Unai Fischer-Abaigar , Christoph Kern , Juan Carlos Perdomo This is my paper · ORCID

classification cs.CYcs.LGstat.ML

keywords prediction-accessratioworst-offidentificationscreeningcapacityalgorithmictargetingwelfareallocationlong-termunemploymentriskpredictionbureaucratic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks how a government agency that wants to identify the worst-off should invest: in better risk prediction or in screening more people. Its central claim is that, for moderate prediction quality and typical screening budgets, expanding access produces at least as much improvement in identifying the truly needy as improving the predictive model. The claim is captured by the prediction-access ratio (PAR), the ratio of the policy-value gain from expanding capacity to the gain from better prediction. In the Gaussian model the paper proves PAR is at least 1 for many realistic regimes, and the German unemployment case study ($R^2=0.15$) observes the same. The upshot is that prediction matters most at the extremes: when it explains almost nothing, or almost everything.

What carries the argument

The central object is the prediction-access ratio (PAR), defined as $\frac{V(\alpha+\Delta\alpha,\beta,R^2)-V(\alpha,\beta,R^2)}{V(\alpha,\beta,R^2+\Delta R^2)-V(\alpha,\beta,R^2)}$, where $V$ is the fraction of the truly at-risk population that is screened. In the theoretical model $Y$ and $\hat{Y}$ are bivariate normal with squared correlation $R^2$, so $V$ is $\Phi_2(\Phi^{-1}(\alpha),\Phi^{-1}(\beta);\rho)/\beta$ with $\rho=\sqrt{R^2}$. The argument differentiates this bivariate normal CDF with respect to $\alpha$ and $R^2$, bounds the ratio of the derivatives using the normal hazard ratio $\Phi(z)/\varphi(z)$, and compares the bounds in the regimes of interest. This yields the proofs that PAR is at least 1 for moderate $R^2$ with $\alpha\le\beta$, and that PAR grows rapidly when capacity is scarce.

What would settle it

Compute the prediction-access ratio on the German data after genuinely retraining the model with additional features or more data, rather than artificially scaling residuals; if the observed PAR falls below 1 for $\alpha\le\beta$ at $R^2=0.15$, the paper's case-study claim in that regime is contradicted.

Watch

Extended reading notes

Core claim

In the screening problem defined here, an agency observes features $X$, builds a predictor $\hat{Y}$ of a welfare outcome $Y$, and screens the fraction $\alpha$ of people whose predicted outcome is lowest, hoping to capture the $\beta$ fraction who are truly worst-off. The paper's central result is that the prediction-access ratio — the marginal gain in the fraction of worst-off people identified from expanding $\alpha$ divided by the marginal gain from increasing $R^2$ — is at least 1 whenever $R^2$ is moderate, $\alpha$ does not exceed $\beta$, and $\beta$ is not tiny. Proposition 3 states this for $0.15\le R^2\le 0.85$ with $\alpha\le\beta$; Theorem 3.1 shows capacity dominates dramatically when capacity is very scarce; and Theorem 3.2 shows prediction's marginal value peaks only as $R^2\to 0$ or $R^2\to 1$. The empirical study of German unemployment durations, with $R^2=0.15$, reproduces the theoretical PAR pattern, and a simple decision tree needs a modest extra screening capacity to match the policy value of a much more complex model.

Load-bearing premise

The empirical conclusions assume that actual prediction improvements act like uniformly shrinking every prediction error at once, preserving the ranking of who is hardest to predict; if real model improvements are uneven across groups, PAR could look different.

Editorial extensions

If this is right

For agencies already achieving modest predictive power ($R^2$ around 0.15–0.5), a small increase in the share of the population screened yields a policy-value gain at least as large as an equal-sized improvement in prediction, before costs are considered.
When screening capacity is very small relative to the target group ($\alpha\ll\beta$), expanding capacity is overwhelmingly more valuable than improving prediction.
Prediction is a first- and last-mile effort: marginal improvements in $R^2$ have their highest relative impact when $R^2$ is near 0 or near 1 (with $\alpha=\beta$ in the latter case).
Because the decision rule depends only on PAR and the cost ratio, a planner should expand access whenever the cost of expanding access divided by the cost of improving prediction is below PAR.
In the German unemployment application, reaching 75% of long-term unemployed jobseekers required roughly 25 percentage points of additional screening capacity beyond $\alpha=\beta$, and a simple 4-depth decision tree could match the complex model's policy value given a small extra capacity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The uniform residual-scaling model of prediction improvement may understate the value of targeted interventions like adding new features or cleaning measurement error in high-risk subgroups; if real improvements concentrate where errors are worst, PAR could fall below 1 earlier than the paper's regime analysis suggests.
The paper's separation of PAR from costs implicitly gives agencies an operational order: measure PAR first, then invest in prediction only where the cost ratio beats it; the authors do not spell out this workflow, but it follows directly from their formulas.
For programs with recurring costs (staff, data collection) and fixed costs (model development, infrastructure), amortization over time could change the effective cost ratio even though PAR itself is unchanged; the paper mentions this as future work.
A natural benchmark for the residual-scaling assumption is to estimate PAR under actual retraining with additional features or samples at several $R^2$ levels; a systematic gap between the two estimates would indicate when the paper's practical guidance needs a richer improvement model.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

This is a genuine extension of Perdomo's PAR idea to continuous welfare outcomes with a worst-off identification objective. The theory is the strong half: a clean closed-form policy value for the Gaussian case, a small-capacity lower bound, the first-and-last-mile characterization, and a moderate-R2 regime where the local PAR is at least 1. The appendix proofs look algebraically consistent, and the bounds are more than a restatement of Perdomo 2024 – the objective is different and the lemmas are re-derived.

The German case study gives the paper its practical face, and the "screening gap" comparison between CatBoost and a depth-4 tree is a nice, directly usable piece of analysis. The empirical PAR computation is transparent.

The soft spots are real but not disqualifying. The strongest is the residual-scaling counterfactual: improved predictions are simulated by shrinking every residual by a common factor, which preserves who is predicted well and who isn't. If real model improvements concentrate on the hardest-to-predict worst-off cases, the value gain can be larger for a given delta R2, and the PAR could drop below 1 in the very R2=0.15 regime the paper highlights. The paper acknowledges this in Section 4 and the conclusion, but the abstract and introduction push the "capacity usually beats prediction" line without that caveat. So the empirical conclusion is conditional on the improvement pathway, not robust to it.

A second, minor issue: the claim that the Gaussian model's insights "exactly match up" in the case study is too strong. The figures show broad qualitative agreement, not exact match. And the PAR estimates have no uncertainty quantification; point estimates without intervals are thin for a policy-facing quantity.

The connection to Perdomo [2024] is handled honestly: the PAR definition comes from there, and the new results are derived for a distinct setting. Not a circularity problem.

Who this is for: people building or auditing algorithmic profiling systems, and researchers who want a compact way to reason about targeting the worst-off. The paper deserves a serious referee. I'd send it out, and in review I'd push on the empirical counterfactual: test a couple of alternative improvement pathways (feature additions, subgroup-scaled residuals) or, failing that, substantially soften the policy claim and add intervals around the PAR. The theory alone is a solid contribution.

Referee Report

4 major / 6 minor

Summary. The paper studies the welfare value of machine-learned risk prediction when the policy goal is to identify the worst-off segment of a population, measured by the fraction of truly at-risk individuals who are screened. It formalizes a screening problem, derives a closed-form Gaussian policy value V(α,β,R²) as a bivariate normal probability, defines the Prediction-Access Ratio (PAR) comparing marginal gains from expanding screening capacity α versus improving prediction R², and proves several theoretical results: PAR grows without bound in the scarce-capacity regime (Theorem 3.1), local prediction improvements dominate only near R²=0 or R²=1 with α=β (Theorem 3.2), and PAR is at least 1 for moderate R² with α≤β (Proposition 3). The empirical part trains a CatBoost model on German administrative labor-market data to predict unemployment duration (test R²=0.15), computes PAR nonparametrically under a uniform residual-scaling counterfactual for prediction improvements, and finds that capacity expansion tends to dominate prediction improvement in the operating regime. The paper also compares a shallow decision tree with CatBoost and expresses the value gap in terms of additional screening capacity.

Significance. If correct, the theoretical results provide a simple, falsifiable characterization of when prediction improvements matter for worst-off targeting, and the case study gives a transferable methodology for policy evaluation. The paper's strengths include explicit distributional assumptions, algebraically detailed proofs in the appendix, a clearly stated policy metric rather than generic accuracy, and a real administrative-data demonstration that engages with the practical regime of social screening systems. The theoretical Proposition 3 is a crisp, non-obvious claim, and the paper's framing of prediction as a first- and last-mile effort is a useful counterweight to prediction-centric policy discussions. However, the empirical conclusion is tied to a specific counterfactual pathway for prediction improvements and lacks uncertainty quantification, so the applied claim is not yet as robust as the theoretical core.

major comments (4)

[Section 4 and Appendix B.3] The empirical PAR counterfactual assumes that prediction improvements are equivalent to uniformly scaling residuals, Ŷ⁺ = Ŷ + δ(Y−Ŷ). This compresses all errors proportionally and preserves the ranking of who is well-predicted. Real prediction improvements—through new features, changed model architectures, or subgroup-targeted data collection—can be non-uniform, potentially concentrating gains on the hardest-to-predict worst-off individuals. For a fixed ΔR², such targeted gains can increase V more than the uniform-scaling path, lowering the PAR and possibly pushing it below 1 in the very regime the paper highlights (R²=0.15, α≈β). The validation in Figure 13 only varies training sample size, which preserves residual shape; it does not cover feature additions or subgroup-specific error reductions. Since the abstract and Section 5.1 state the conclusion without this caveat, the manuscript should either test alternative improvement pathways (e.g., subgroup-specific residual scaling or reweighting improvements toward the tail of Y) or explicitly restrict the empirical conclusion to the uniform-improvement pathway.
[Section 5.1 and Figure 6] The PAR values are reported as point estimates without any uncertainty quantification. The test set is large (274,515 jobseekers), but PAR is a ratio of differences of recall-like quantities; near PAR=1, sampling variability and the choice of tie-breaking at thresholds could plausibly change the qualitative conclusion. For the applied claim that 'capacity improvements stay relatively more effective' in the operating regime, the paper should provide confidence intervals (e.g., bootstrap over unemployment spells) or a sensitivity analysis over the α values where PAR crosses 1.
[Section 2 (Proposition 1) versus Section 5.1] The empirical policy ranks individuals by the raw predicted duration Ŷ and screens the bottom α. Proposition 1, however, shows that the optimal policy given a predictor ranks by the conditional probability s(Ŷ) = Pr[Y ≤ F_Y⁻¹(β) | Ŷ]. The manuscript does not verify that these two rankings coincide in the case study. If the conditional distribution of Y given Ŷ is not stochastically increasing in Ŷ, the computed policy values may understate what the model could achieve, which could bias the estimated PAR (and the comparison between capacity and prediction improvements). A concrete test would be to compare the empirical value of thresholding on Ŷ with thresholding on a nonparametric estimate of s(Ŷ), or to check monotonicity of estimated s(Ŷ) against Ŷ.
[Section 3.1 (Discussion) and Figure 2] The cost-ratio adjustment appears inconsistent. The decision rule stated in Section 1.1 is: expand access whenever C_Access/C_Pred < PAR. If the cost ratio C_Access/C_Pred equals 1/4, then access is cheaper and should be favored whenever PAR > 0.25, so the region where investing in better prediction is more efficient should shrink relative to the cost-free PAR=1 boundary. However, the text says 'the regions where investing in R² is more efficient expand,' and the figure caption displays '1/4×PAR.' If the plotted quantity is PAR/4, the decision boundary in the plot corresponds to PAR=4, which is C_Access/C_Pred = 4, not 1/4. The manuscript should clarify which ratio is meant (C_Access/C_Pred versus C_Pred/C_Access) and correct either the caption or the verbal description so that the direction of the cost effect is consistent with the stated decision rule.

minor comments (6)

[Section 1.1] The claim that the Gaussian model yields 'surprisingly precise numerical insights that exactly match up in our real-world case study' overstates the correspondence; Section 5.1 says the empirical observations 'broadly match' the theory. Please rephrase to avoid implying exact quantitative agreement, especially since the case-study data are non-Gaussian and the empirical PAR uses non-local increments.
[Equation 3] The notation for PAR is introduced in Eq. (3) without arguments, while later text uses PAR(α,R²,Δα,ΔR²) and PAR(α,β,Δ). Please define the full argument list at first use and keep it consistent throughout.
[Section 4] The subsection heading 'What is the impact of improving screening capacity versus prediction errors?' should read 'prediction improvements' rather than 'prediction errors' to match the content.
[Proof of Proposition 3 (Appendix D.5)] The numerical lower bound for T2 at R²=0.15 and β=0.03 is stated as 0.59, but direct computation gives Φ(−1.25)/φ(−1.25) ≈ 0.578. Please check the arithmetic and either correct the value or note that it is approximate.
[Figure 13] The caption states that residual distributions are shown for models trained with varying sample sizes; please add a sentence noting that this validates the uniform-scaling assumption only for the sample-size pathway and does not cover other ways of increasing R².
[Theorem 3.1] The statement 'PAR is at least [expression] + o(1)' is slightly ambiguous because o(1) appears to be part of the lower bound rather than an error term in the definition. Consider writing the bound as 'PAR ≥ f(α)(1 + o(1))' or making explicit that the o(1) is relative to the leading term as α→0.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PAR framework and Gaussian lemmas are cited from prior work by one co-author but are used as transparent tools, and the empirical PAR is measured under an explicitly stated counterfactual rather than fitted to the conclusions.

full rationale

The derivation chain is self-contained. Equation (3) defines PAR as a ratio of policy-value increments, and the paper does not define capacity or prediction improvements in terms of the target conclusion. Proposition 2 and the derivative lemmas are proved in Appendices D.2–D.6 from the bivariate normal CDF. The only imports from Perdomo [2024] are the PAR concept itself and two standard Gaussian tail bounds (Lemmas B.5 and A.6 in the proof of Theorem 3.1); these are transparent references to prior theoretical work by one co-author, not an unverified uniqueness theorem, and they are not the load-bearing source of the main regime analysis, whose proofs are supplied here. The case-study R2 of 0.15 is measured from a CatBoost model on administrative data, and the residual-scaling construction in Section 4 and Appendix B.3 is explicitly labeled as a simulation of 'similar but slightly better' models, with the paper itself noting non-uniform residual adjustments and retraining as alternative pathways (Section 4, Conclusion). Thus the empirical PAR values are not fitted inputs renamed as predictions; they are evaluated under a stated counterfactual. The residual-scaling assumption is a limitation and a robustness concern, but it is not circular: it does not presuppose the conclusion that capacity dominates prediction. The self-citations are real prior-work references and do not by themselves make the derivation circular.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The theoretical core uses a Gaussian model with no fitted constants; the only hand-set quantity is the equal increment size Δ used to define PAR. The axioms are the Gaussian error structure, the recall-based objective, the local-improvement framing, standard asymptotic lemmas, and separable costs. No new entities are postulated.

free parameters (1)

increment size Δ for PAR = Δα = ΔR2 = 0.1 (sensitivity 0.01)
PAR is defined for equal increments in α and R2; the choice of increment size is a hand-set analysis parameter, not derived from data, and can affect whether PAR is above or below 1.

assumptions (5)

domain assumption Welfare outcomes Y follow a normal distribution and prediction errors ε = Y − Y^ are independent zero-mean Gaussians, so R2 fully characterizes prediction quality.
Used in Proposition 2 and all theorems; the real case study does not satisfy this, so the theoretical PAR bounds are exact only under this model.
domain assumption The planner's objective is to maximize the fraction of the worst-off identified, with no cost or harm from false positives.
This justifies the recall-style value function V and excludes settings where over-screening has negative consequences.
domain assumption Policy improvements can be modeled as local changes (infinitesimal Δ) for the theoretical bounds; non-local behavior may differ.
Theorems 3.1, 3.2 and Proposition 3 concern local derivatives; the empirical study uses Δ=0.1, and the paper notes the theory is conservative.
standard math Asymptotic Gaussian bounds from Perdomo [2024] (Lemma B.5, Lemma A.6) are correct and apply in this setting.
Used in the proof of Theorem 3.1 to control the Gaussian pdf near large quantiles.
domain assumption The cost of expanding access and improving prediction can be summarized by a single constant cost ratio C_access/C_pred.
The decision rule 'expand access if cost ratio < PAR' assumes separable, constant marginal costs; the paper acknowledges more complex cost structures as future work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Value of Prediction in Identifying the Worst-Off." pith.science (2026). https://pith.science/paper/HLFCQD4L

@misc{pith2026250119334,
  author       = {Pith},
  title        = {Pith review of: The Value of Prediction in Identifying the Worst-Off},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HLFCQD4L}},
  note         = {Machine review of arXiv:2501.19334}
}

read the original abstract

Machine learning is increasingly used in government programs to identify and support the most vulnerable individuals, prioritizing assistance for those at greatest risk over optimizing aggregate outcomes. This paper examines the welfare impacts of prediction in equity-driven contexts, and how they compare to other policy levers, such as expanding bureaucratic capacity. Through mathematical models and a real-world case study on long-term unemployment amongst German residents, we develop a comprehensive understanding of the relative effectiveness of prediction in surfacing the worst-off. Our findings provide clear analytical frameworks and practical, data-driven tools that empower policymakers to make principled decisions when designing these systems.

Figures

Figures reproduced from arXiv: 2501.19334 by the authors.

**Figure 1.** Screening Policy in Gaussian Setting. (Left) Probability of being screened for an individual with a specific welfare outcome Y , given R 2 = 0.25, α = 0.2, and β = 0.2. The dashed line represents the unconstrained oracle policy, which perfectly screens those in need. (Middle) Policy with expanded screening capacity, where α increases by ∆α = 0.2. (Right) Policy under an improved prediction model with R 2 + ∆R2 , whe… view at source ↗

**Figure 2.** Numerical Simulation of the Prediction-Access Ratio (PAR), Equation 3, for ∆R2 =∆α =0.01 and β = 0.2. (Left) The PAR values. (Right) 1/4 × PAR, representing a cost ratio of 1/4. Each point represents a screening capacity α (x-axis) and R 2 value (y-axis), with the color bar showing the PAR clipped to the range [0.5,2.0]. Dotted black lines represent PAR=1, where improvements in α and R 2 are equally effective. The p… view at source ↗

**Figure 3.** Unemployment duration The red line marks the 12 month threshold used to classify a jobseeking episode as long-term unemployment (LTU) in Germany. We secured access to a dataset1 on German jobseekers derived from German administrative labor market records that cover a large portion of the German labor force. It covers a period from 1975 to 2017 and merges multiple administrative data sources, containing a wide spectr… view at source ↗

Figures from the paper (13 more)

**Figure 4.** Figure 4: Stacked timeline diagram illustrating training (2010–2013), validation (2012–2014), and test [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: Policy value across different screening capacities (α) and worst-off fractions β evaluated on the test set using the CatBoost regression model. A β value of 0.15 corresponds to the 12-month cutoff used to define long-term unemployment in Germany. additional capacity ∆α…

**Figure 6.** Figure 6: Prediction-Access Ratio for ∆R2 = ∆α = 0.1 across three regimes. As expected from our theoretical intuition, the PAR is large for small α and for the trained model (b), which represents the typical regime for allocation systems. R 2 , gains in policy value diminish so …

**Figure 7.** Figure 7: The rate of local improvements in V (R 2 ) with respect to small changes in R 2 . In both regimes, the local improvements diverge as ∆R2 approaches zero. Note that these are on a logarithmic scale. planner might instead increase screening capacity. Formally, we define …

**Figure 8.** Figure 8: (a) The difference in policy value between a 4-depth decision tree and CatBoost model. (b) The minimum additional screening capacity that would need to be invested for the decision tree to achieve a policy value comparable to that of the CatBoost model. 6 Conclusion Th…

**Figure 9.** Figure 9: Normal welfare distribution, with vertical lines marking the quantile cuto [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

**Figure 10.** Figure 10: Numerical Simulation of the Prediction-Access Ratio (PAR), Equation 3, for ∆R2 = ∆α = 0.01 and β = 0.05. Each point represents a screening capacity α (x-axis) and R 2 value (y-axis), with the color bar showing the PAR clipped to the range [0.5,2.0]. The vertical black…

**Figure 11.** Figure 11: Residual Distribution Before and After Adjustment Figure (a) shows the residual distribution for the original predictions (∆R2 = 0), while Figure (b) shows the residual distribution after increasing the R 2 -value (∆R2 = 0.1) for the CatBoost model. The adjustment pre…

**Figure 12.** Figure 12: The R 2 value on the test set for varying training set size (CatBoost Regression). −20 −15 −10 −5 0 5 10 15 20 Residuals (Y − Yˆ ) 0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 Density σ 2 = 37.19 (a) 500 Training Samples −20 −15 −10 −5 0 5 10 15 20 Residuals (Y − Yˆ ) 0.00…

**Figure 13.** Figure 13: Residual distributions on the test set for models trained with varying training set sizes. [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]

**Figure 14.** Figure 14: Prediction-Access Ratio for R 2 = 0 and ∆R2 = ∆α = 0.01. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]

**Figure 15.** Figure 15: V (R 2 +∆R2 ) − V (R 2 ) for CatBoost model and ∆R2 = 0.05. B.5 Binary Classification Instead of predicting the exact duration of unemployment, the problem can be reframed as a binary classification task. For a fixed β, we can define a binary outcome: Y = 1{Y ⩾ F −1 Y…

**Figure 16.** Figure 16: Policy Value and True Positive Count on Test Set (Classification). [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 23 canonical work pages

[1]

Aiken, T

E. Aiken, T. Ohlenburg, and J. Blumenstock. M oving targets: W hen does a poverty prediction model need to be updated? In Proceedings of the 6th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, COMPASS '23, page 117, New York, NY, USA, 2023. Association for Computing Machinery. ISBN 9798400701498. doi:10.1145/3588001.3609369. URL https...

work page arXiv 2023
[2]

Antoni, A

M. Antoni, A. Ganzer, and P. vom Berge. Factually anonymous version of the Sample of Integrated Labour Market Biographies ( SIAB-Regionalfile ) -- Version 7517 v1. Research Data Centre of the Federal Employment Agency (BA) at the Institute for Employment Research (IAB), 2019 a . 10.5164/IAB.SIAB-R7517.de.en.v1

work page doi:10.5164/iab.siab-r7517.de.en.v1 2019
[3]

Antoni, A

M. Antoni, A. Ganzer, and P. vom Berge. Sample of Integrated Labour Market Biographies Regional File (SIAB-R) 1975 - 2027. FDZ-Datenreport 04/2019 (en), Research Data Centre of the Federal Employment Agency (BA) at the Institute for Employment Research (IAB), Nürnberg, 2019 b . 10.5164/IAB.FDZD.1904.en.v1

work page doi:10.5164/iab.fdzd.1904.en.v1 1975
[4]

Athey and S

S. Athey and S. Wager. P olicy L earning with O bservational D ata. Econometrica, 89 0 (1): 0 133--161, 2021

work page 2021
[5]

R. L. Bach, C. Kern, H. Mautner, and F. Kreuter. The impact of modeling decisions in statistical profiling. Data & Policy, 5: 0 e32, 2023. doi:10.1017/dap.2023.29

work page doi:10.1017/dap.2023.29 2023
[6]

Barocas, M

S. Barocas, M. Hardt, and A. Narayanan. Fairness and Machine Learning: Limitations and Opportunities. MIT Press, 2023

2023
[7]

J. E. Blumenstock. Fighting P overty with D ata. Science, 2016

work page 2016
[8]

Boehmer, Y

N. Boehmer, Y. Nair, S. Shah, L. Janson, A. Taneja, and M. Tambe. E valuating the E ffectiveness of I ndex- B ased T reatment A llocation. arXiv preprint arXiv:2402.11771, 2024

arXiv 2024

Show all 37 references

[9]

C. W. Chan, V. F. Farias, N. Bambos, and G. J. Escobar. O ptimizing I ntensive C are U nit D ischarge D ecisions with P atient R eadmissions. Operations Research, 60 0 (6): 0 1323--1341, 2012. doi:10.1287/opre.1120.1105. URL https://doi.org/10.1287/opre.1120.1105

2012
[10]

Chouldechova, D

A. Chouldechova, D. Benavides-Prado, O. Fialko, and R. Vaithianathan. A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions. In Conference on Fairness, Accountability and Transparency, pages 134--148. PMLR, 2018

2018
[11]

Clementi and M

F. Clementi and M. Gallegati. Pareto’s law of income distribution: E vidence for G ermany, the U nited K ingdom, and the U nited S tates. Econophysics of wealth distributions: Econophys-Kolkata I, pages 3--14, 2005

2005
[12]

Coston, A

A. Coston, A. Kawakami, H. Zhu, K. Holstein, and H. Heidari. A V alidity P erspective on E valuating the J ustified U se of D ata-driven D ecision-making A lgorithms. In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 690--704, 2023. doi:10.1109/...

2023
[13]

Desiere and L

S. Desiere and L. Struyven. Using Artificial Intelligence to classify Jobseekers : The Accuracy-Equity Trade-off . Journal of Social Policy, 50 0 (2): 0 367--385, Apr. 2021. ISSN 0047-2794, 1469-7823. doi:10.1017/S0047279420000203

2021 doi
[14]

Desiere, K

S. Desiere, K. Langenbucher, and L. Struyven. Statistical P rofiling in P ublic E mployment S ervices: An I nternational C omparison. Technical Report 224, OECD Publishing, 2019. URL https://doi.org/10.1787/b5e5f16e-en

2019 doi
[15]

Drezner and G

Z. Drezner and G. O. Wesolowsky. On the C omputation of the B ivariate N ormal I ntegral. Journal of Statistical Computation and Simulation, 1990

1990
[16]

predict, then optimize

A. N. Elmachtoub and P. Grigas. Smart “predict, then optimize”. Management Science, 68 0 (1): 0 9--26, 2022

2022
[17]

Fern\' a ndez-Lor\' a and F

C. Fern\' a ndez-Lor\' a and F. Provost. C ausal D ecision M aking and C ausal E ffect E stimation A re N ot the S ame…and W hy I t M atters. INFORMS Journal on Data Science, 1 0 (1): 0 4--16, 2022. doi:10.1287/ijds.2021.0006. URL https://doi.org/10.1287/ijds.2021.0006

2022
[18]

Fischer-Abaigar, C

U. Fischer-Abaigar, C. Kern, N. Barda, and F. Kreuter. Bridging the G ap: Towards an E xpanded T oolkit for A i-driven D ecision-making in the P ublic S ector. Government Information Quarterly, 41 0 (4): 0 101976, 2024. ISSN 0740-624X. doi:https://doi.org/10.1016/j.giq.2024.10...

2024
[19]

Guerdan, A

L. Guerdan, A. Coston, K. Holstein, and Z. S. Wu. C ounterfactual P rediction U nder O utcome M easurement E rror. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT '23, page 1584–1598, New York, NY, USA, 2023. Association for Compu...

2023
[20]

S. Jain, K. Creel, and A. C. Wilson. P osition: S carce R esource A llocations T hat R ely O n M achine L earning S hould B e R andomized. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=44qxX6Ty6F

2024
[21]

R. A. Johnson and S. Zhang. What is the B ureaucratic C ounterfactual? C ategorical versus A lgorithmic P rioritization in U . S . S ocial P olicy. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, page 1671–1682, New York, NY,...

2022
[22]

N. Kallus. More E fficient P olicy L earning via O ptimal R etargeting. Journal of the American Statistical Association, 116 0 (534): 0 646--658, 2021. doi:10.1080/01621459.2020.1788948. URL https://doi.org/10.1080/01621459.2020.1788948

2021
[23]

C. Kern, R. Bach, H. Mautner, and F. Kreuter. When S mall D ecisions H ave B ig I mpact: F airness I mplications of A lgorithmic P rofiling S chemes. ACM Journal on Responsible Computing, 1 0 (4), Nov. 2024. doi:10.1145/3689485. URL https://doi.org/10.1145/3689485

2024 doi
[24]

Kitagawa and A

T. Kitagawa and A. Tetenov. W ho S hould B e T reated? E mpirical W elfare M aximization M ethods for T reatment C hoice. Econometrica, 86 0 (2): 0 591--616, 2018. doi:https://doi.org/10.3982/ECTA13288. URL https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA13288

2018 doi
[25]

Kleinberg, J

J. Kleinberg, J. Ludwig, S. Mullainathan, and Z. Obermeyer. Prediction P olicy P roblems. American Economic Review, 105 0 (5): 0 491--495, 2015

2015
[26]

K \"o rtner and G

J. K \"o rtner and G. Bonoli. Predictive A lgorithms in the D elivery of P ublic E mployment S ervices. In Handbook of Labour Market Policy in Advanced Democracies, pages 387--398. Edward Elgar Publishing, 2023

2023
[27]

Loxha and M

A. Loxha and M. Morgandi. Profiling the unemployed: a review of OECD experiences and implications for emerging economics. Social protection discussion papers and notes, 0 (91051), 2014

2014
[28]

C. F. Manski. S tatistical T reatment R ules for H eterogeneous P opulations. Econometrica, 72 0 (4): 0 1221--1246, 2004. doi:https://doi.org/10.1111/j.1468-0262.2004.00530.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1468-0262.2004.00530.x

2004
[29]

J. C. Perdomo. The R elative V alue of P rediction in A lgorithmic D ecision M aking. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024

2024
[30]

J. C. Perdomo, T. Britton, M. Hardt, and R. Abebe. Difficult L essons on S ocial P rediction from W isconsin P ublic S chools. arXiv preprint arXiv:2304.06205, 2023

2023 arXiv
[31]

Potash, J

E. Potash, J. Brew, A. Loewi, S. Majumdar, A. Reece, J. Walsh, E. Rozier, E. Jorgenson, R. Mansour, and R. Ghani. P redictive M odeling for P ublic H ealth: P reventing C hildhood L ead P oisoning. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Dis...

2015
[32]

M. J. Salganik, I. Lundberg, A. T. Kindel, C. E. Ahearn, K. Al-Ghoneim, A. Almaatouq, D. M. Altschul, J. E. Brand, N. B. Carnegie, R. J. Compton, D. Datta, T. Davidson, A. Filippova, C. Gilroy, B. J. Goode, E. Jahani, R. Kashyap, A. Kirchner, S. McKay, A. C. Morgan, A. Pentlan...

2020
[33]

Shirali, R

A. Shirali, R. Abebe, and M. Hardt. A llocation R equires P rediction O nly if I nequality I s L ow. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net, 2024. URL https://openreview.net/forum?id=WUicA0hOF9

2024
[34]

T. Q. Sun and R. Medaglia. Mapping the challenges of A rtificial I ntelligence in the public sector: E vidence from public healthcare. Government Information Quarterly, 36 0 (2): 0 368--383, 2019

2019
[35]

A. Wang, S. Kapoor, S. Barocas, and A. Narayanan. A gainst P redictive O ptimization: O n the L egitimacy of D ecision-making A lgorithms T hat O ptimize P redictive A ccuracy. ACM J. Responsib. Comput., 1 0 (1), Mar. 2024. doi:10.1145/3636509. URL https://doi.org/10.1145/3636509

2024 doi
[36]

Wilder and P

B. Wilder and P. Welle. Learning treatment effects while treating those in need. arXiv preprint arXiv:2407.07596, 2024

2024 arXiv
[37]

B. W. Wirtz, J. C. Weyerer, and C. Geyer. Artificial I ntelligence and the P ublic S ector— A pplications and C hallenges. International Journal of Public Administration, 42 0 (7): 0 596--615, 2019

2019

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Aiken, T

E. Aiken, T. Ohlenburg, and J. Blumenstock. M oving targets: W hen does a poverty prediction model need to be updated? In Proceedings of the 6th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, COMPASS '23, page 117, New York, NY, USA, 2023. Association for Computing Machinery. ISBN 9798400701498. doi:10.1145/3588001.3609369. URL https...

work page arXiv 2023

[2] [2]

Antoni, A

M. Antoni, A. Ganzer, and P. vom Berge. Factually anonymous version of the Sample of Integrated Labour Market Biographies ( SIAB-Regionalfile ) -- Version 7517 v1. Research Data Centre of the Federal Employment Agency (BA) at the Institute for Employment Research (IAB), 2019 a . 10.5164/IAB.SIAB-R7517.de.en.v1

work page doi:10.5164/iab.siab-r7517.de.en.v1 2019

[3] [3]

Antoni, A

M. Antoni, A. Ganzer, and P. vom Berge. Sample of Integrated Labour Market Biographies Regional File (SIAB-R) 1975 - 2027. FDZ-Datenreport 04/2019 (en), Research Data Centre of the Federal Employment Agency (BA) at the Institute for Employment Research (IAB), Nürnberg, 2019 b . 10.5164/IAB.FDZD.1904.en.v1

work page doi:10.5164/iab.fdzd.1904.en.v1 1975

[4] [4]

Athey and S

S. Athey and S. Wager. P olicy L earning with O bservational D ata. Econometrica, 89 0 (1): 0 133--161, 2021

work page 2021

[5] [5]

R. L. Bach, C. Kern, H. Mautner, and F. Kreuter. The impact of modeling decisions in statistical profiling. Data & Policy, 5: 0 e32, 2023. doi:10.1017/dap.2023.29

work page doi:10.1017/dap.2023.29 2023

[6] [6]

Barocas, M

S. Barocas, M. Hardt, and A. Narayanan. Fairness and Machine Learning: Limitations and Opportunities. MIT Press, 2023

2023

[7] [7]

J. E. Blumenstock. Fighting P overty with D ata. Science, 2016

work page 2016

[8] [8]

Boehmer, Y

N. Boehmer, Y. Nair, S. Shah, L. Janson, A. Taneja, and M. Tambe. E valuating the E ffectiveness of I ndex- B ased T reatment A llocation. arXiv preprint arXiv:2402.11771, 2024

arXiv 2024