Pith. sign in

REVIEW 3 major objections 6 minor 90 references

Improving Access to Essential Medicines via Decision-Aware Machine Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A decision-aware machine-learning allocation tool raised consumption of essential medicines by an estimated 19% in Sierra Leone's five pilot districts.

desk verdict A genuine field deployment of decision-aware ML in Sierra Leone with a real but testable supply-confounding question. read the letter →

arxiv 2607.20542 v1 pith:CGURWG7U submitted 2026-07-10 cs.LG cs.AIcs.CY

Angel Tsai-Hsuan Chung , Jatu Abdulai , Patrick Bayoh , Lawrence Sandi , Francis Smart , Hamsa Bastani , Osbert Bastani This is my paper · ORCID
classification cs.LGcs.AIcs.CY
keywords essentialmedicinesdecision-awaremachinelearningresourceallocationsupplychainoptimizationSierraLeonemulti-taskcatalyticpriorsdifference-in-differences
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 argues that a machine-learning system that predicts demand and then optimizes how scarce medicines are distributed can substantially improve access to essential medicines even when historical data are sparse and noisy. Its evidence comes from a randomized pilot in five Sierra Leone districts: using a synthetic difference-in-differences analysis, consumption of allocated products rose an estimated 19% in treated districts relative to a constructed counterfactual. The tool was then adopted nationwide, covering roughly two million women and children under five, at a running cost of about $30 per month in server fees. The significance for global health is that data-efficient machine learning can improve health-system performance in low- and middle-income countries without new physical supply or extra staff.

What carries the argument

The load-bearing component is a decision-aware learning rule: instead of minimizing mean squared forecast error, the model minimizes a weighted absolute error whose weights come from a Taylor expansion of the expected-unmet-demand loss, up-weighting observations for facilities likely to be under-stocked. Demand is predicted by a multi-task random forest (one model per product category, sharing data across facilities) regularized toward a population-based catalytic prior built from catchment-population estimates, which protects data-poor facilities. Given predicted demand distributions, allocations are computed by a linear-programming sample average approximation that minimizes expected unmet

What would settle it

Compare actual central-stock deliveries (quantities and dates per facility from warehouse records) between treated and control districts in 2023 Q2. If treated districts received systematically more units, or received them earlier, and the 19% consumption gap shrinks or disappears after controlling for quantity received and delivery timing, the allocation-efficiency claim would be falsified. A cleaner test would use a placebo allocation: run the tool on a product whose supply was fixed by the government and check for a consumption effect.

Watch

Extended reading notes

Core claim

The central discovery is that training the demand-forecasting model with the downstream allocation objective in mind—rather than for raw prediction accuracy—yields allocations that measurably raise medicine consumption. In 2023 Q2, five random districts used the system for facility-level allocation while eleven continued with the prior procedure; treated districts showed a statistically significant 19% increase in consumption, with larger gains for larger facilities (36%) and previously under-served facilities (32%). The authors also establish an equivalence between maximizing consumption and minimizing unmet demand under a dispensing model without waste, and report consistent results under

Load-bearing premise

The causal estimate assumes the five treated districts received the same quantity and timing of central supply as the eleven control districts, so that any difference in consumption reflects only how the stock was allocated.

Editorial extensions

If this is right

  • A 19% rise in consumption implies the same volume of donated stock met more patient need, since total supply was fixed before allocation.
  • The equity analyses show the gains are not a rich-district effect: previously stockout-prone facilities improved 32% and rural facilities showed significant gains.
  • The compliance data (normalized overlap 0.89–1.00 with algorithmic allocations) mean the reported effect reflects actual implementation, not just a recommendation.
  • At $30 per month in server costs and an estimated $5.24 per DALY, the intervention is cheap enough that the financial barrier to replication elsewhere is minimal.

Reading between the lines

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

  • A natural next test is whether the same decision-aware reweighting improves allocation of other scarce health commodities—vaccines, insecticide-treated nets, diagnostics—where censored demand and missing data are also pervasive.
  • The consumption-based evaluation cannot separate use by genuinely needy patients from provider-induced demand; linking allocations to health outcomes (e.g., stockout days, facility visits) would strengthen the access claim.
  • The 19% estimate would shrink if treated districts received more or earlier supply; checking warehouse invoices by district and delivery date would quantify how much of the gap is allocation efficiency.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes a decision-aware machine learning framework for allocating essential medicines in Sierra Leone, combining multi-task learning, catalytic priors, and a stochastic optimization layer. The system was deployed as a decision-support tool in five districts in 2023 Q2 and later scaled nationwide. The central empirical claim, evaluated with Synthetic Difference-in-Differences, is an estimated 19% increase in normalized consumption of allocated products in treated districts, robust to several alternative specifications (DiD, geographic matching, imputation, alternative controls, substitution-robust analyses, LATE). The paper also reports distributional benefits concentrated in larger, rural, and previously under-served facilities, and a very low cost per DALY.

Significance. If the causal estimate is credible, this is a substantial and rare field demonstration of an ML-based allocation system in a low-resource public-health setting, with meaningful policy implications. The paper's strengths include the real-world deployment with government partnership, multiple consistent estimators (SynthDiD, DiD, matching, imputation, product-based controls), explicit robustness checks for missingness and substitution, and a cost-effectiveness analysis. These features make the manuscript valuable. However, the headline causal claim currently rests on an unverified assumption about supply-quantity balance between treated and control districts, and on a post-period that includes a quarter in which the control group also received the intervention. These are load-bearing for the central result and need to be addressed before the claim can be accepted.

major comments (3)
  1. [Supplement §2.1 and §4] The allocation design creates a direct channel for supply-quantity confounding. The text states that the government first fixed total amounts for control districts and that 'the remaining supply was then assigned to the treatment districts, maintaining independence of supply quantities between the two groups.' This independence is asserted, not demonstrated. If treated districts received more, less, or earlier supply, the estimated 19% consumption effect could reflect supply volume rather than allocation efficiency. The authors already have mSupply invoice records (used for compliance analysis in §2.4), so it is feasible to check balance on quantities received, delivery timing, and per-product receipts across treatment status. Please provide such a balance table or an alternative control that adjusts for supply quantities.
  2. [§5, Eq. (8), Fig. 3, Fig. SI 3] The main SynthDiD analysis uses a balanced panel from 2022 Q3 through 2023 Q3, but the system was rolled out nationwide at the beginning of 2023 Q3. Thus the 'control' districts also become treated in the final post-period. Including 2023 Q3 in the post-period either contaminates the control group (if coded as untreated) or attenuates the ATT (if coded as treated). The clean pilot post-period is only 2023 Q2. Please report the SynthDiD estimate using only 2023 Q2 as the post-treatment period, or explicitly model the staggered rollout. The 'Alt. Control' analysis using nationwide roll-out has the same issue and should be presented only as a supplementary design.
  3. [§5 and Supplement §2.2] Inference is based on only 5 treated districts and 3 pre-treatment quarters. The jackknife standard errors reported for SynthDiD may not be reliable with such a small number of treated clusters. Please add cluster-robust inference (e.g., wild cluster bootstrap) or a randomization/permutation test that randomly reassigns the 5 treated districts and recomputes the ATT. This is especially important given the borderline significance of the main SynthDiD coefficient (p<0.01 with a standard error of 0.046).
minor comments (6)
  1. [§5, first paragraph] Typo: 'Instead, examined the equivalent objective' should read 'Instead, we examined the equivalent objective.'
  2. [Table 1 and Table SI 6] The column 'Improv.%' is somewhat ambiguous; consider renaming to 'Improvement %' or 'Relative change %' and stating the denominator precisely in the notes.
  3. [Fig. SI 3 and Fig. SI 4] The y-axis label says 'between treated & control district' — should be 'between treated and control districts' (plural) for consistency.
  4. [Supplement Table SI 12] The row 'Vaccines storage: refrigerators2–8◦C' is missing a space: should be 'refrigerators 2–8°C.'
  5. [Supplement §2.4] The compliance measure is described as 'normalized absolute difference' averaged across facility-product pairs; please clarify whether higher overlap corresponds to lower difference, and define the range (0 to 1) explicitly.
  6. [§4 and Supplement §2.1] The paper says districts were 'selected by the central government based on a randomized allocation schedule,' but details of the randomization (e.g., stratified? simple random sample?) are not provided. This is relevant for assessing the validity of the causal design.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the 19% ATT is a measured deployment outcome, not a fitted model output; minor self-citations are not load-bearing.

full rationale

The headline claim—'Our evaluation finds an estimated 19% increased consumption of allocated products in treated districts'—is a field ATT estimated from observed DHIS2 consumption around a randomized staggered deployment, not an output of the fitted demand model. The model's fitted components (catalytic prior C, r, sigma, random forest weights) shape the allocations, but the outcome is measured independently in the post-deployment period; the SynthDiD equation (Eq. 8) estimates tau from outcomes, and no step in Section 3 or Supplement Section 1.5 feeds the model's predictions into the ATT calculation. The consumption/unmet-demand equivalence in Supplement Section 2.3 is an accounting identity (U + C = xi by construction) used only to justify the choice of outcome metric; it is a tautology, not a circular derivation of the causal effect. The self-citations present (e.g., refs. 26, 51, 52) are general methodological references for prediction-plus-optimization and multi-task learning; the load-bearing method citations for catalytic priors [42], SynthDiD [43], and decision-aware learning [38-41,54] are external or validated on held-out historical data. The supply-quantity independence assumption in Section 2.1 ('The remaining supply was then assigned to the treatment districts, maintaining independence of supply quantities between the two groups') is an untested identifying assumption and a legitimate confounding concern, but it is an assumption about the experiment, not a step that reduces the conclusion to the model's own fitted values. Overall, there is no significant circularity; the score of 2 reflects only minor, non-load-bearing self-citations.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical or structural entities are introduced. The 'catalytic prior' consists of synthetic observations generated from a fitted population model, not an independent invented object. The main free parameters are the catalytic prior C, the at-risk population multiplier r, and the fixed demand standard deviation σ; all are estimated from historical or census data.

free parameters (4)
  • Catalytic prior scale C (per product) = not reported
    Defined in §1.5.2 as ξ0_{t,n}=r·C·p_n; C is a product-specific constant 'estimated from our historical data' (average quarterly demand per unit of at-risk population) and directly controls the synthetic prior that regularizes predictions.
  • At-risk population share r = not reported
    In §1.5.2, r converts catchment population to women/children and is 'estimated using 2015 Sierra Leone Census data at the chiefdom level'.
  • Constant demand std dev σ = not reported
    The model fits mean with random forest and σ separately from historical data (§1.5.1); σ appears in the stochastic optimization objective and in the decision-aware weights.
  • Decision-aware loss constant c and small-σ approximation = not reported
    In §1.5.3, the final reweighted loss includes an unspecified constant c and replaces a probability with an indicator under a small-σ approximation; no values are given.
assumptions (5)
  • domain assumption Facilities do not waste stock and dispense at most patient need (a_t ≤ ξ_t).
    The consumption-unmet demand equivalence in Theorem 1 requires a_t ≤ ξ_t; over-dispensing/waste is excluded (§2.3, Theorem 1). The paper tests for rationing/over-dispensing but notes the tests are underpowered.
  • domain assumption Random allocation and supply independence between treatment and control districts.
    The causal interpretation of the SynthDiD ATT assumes treatment status is as-if random and that supply quantities assigned to treated districts do not differ systematically from controls (§2.1).
  • domain assumption DHIS2 consumption records and mSupply stock/expiry records are reliable after preprocessing.
    The prediction model and the outcome panel are built from DHIS2 monthly consumption and mSupply warehouse data (§1.1); preprocessing removes inconsistent/zero entries but cannot verify entries.
  • domain assumption Dropping censored observations (stockout months) leaves training features unbiased enough with catalytic priors correcting covariate shift.
    §1.2 drops censored observations because consumption equals demand only when no stockout occurs; the paper relies on catalytic priors to mitigate the resulting covariate shift.
  • domain assumption SynthDiD identification conditions hold with 5 treated clusters.
    The estimator requires stable pre-treatment differences between treated units and synthetic control and valid jackknife inference; with only 5 treated clusters and 3-4 pre-periods, this is a strong assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Access to Essential Medicines via Decision-Aware Machine Learning." pith.science (2026). https://pith.science/paper/CGURWG7U

@misc{pith2026260720542,
  author       = {Pith},
  title        = {Pith review of: Improving Access to Essential Medicines via Decision-Aware Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CGURWG7U}},
  note         = {Machine review of arXiv:2607.20542}
}
read the original abstract

A critical challenge in healthcare systems in low- and middle-income countries (LMICs) is the efficient and equitable allocation of scarce resources, particularly essential medicines. This problem is complicated by limited high-quality data, which restricts the applicability of traditional data-driven techniques. We propose a novel decision-aware machine learning framework for essential medicines allocation, which additionally leverages multi-task learning to ensure sample efficiency and catalytic priors to ensure equitable allocation. In collaboration with the Sierra Leone national government, we performed a staggered, nationwide deployment of our system as a decision support tool. Our econometric evaluation finds an estimated 19% increase in consumption of allocated products in treated districts, demonstrating its efficacy at improving access to essential medicines. Our tool was subsequently scaled nationwide, covering an estimated 2 million women and children under five. Our work demonstrates how machine learning methods can improve efficiency at very low cost in resource-constrained global health settings.

Figures

Figures reproduced from arXiv: 2607.20542 by the authors.

Figure 1
Figure 1. System Overview. Every quarter, our system extracts and processes data including the total available central stock and historical facility-level consumption records from mSupply and the DHIS2 government database. It then trains a decision-aware prediction model that informs a stochastic optimization procedure to make allocation decisions. The system further provides a picking list for frontline workers to collect su… view at source ↗
Figure 2
Figure 2. Map of treatment distribution in 2023 Q2. Yellow dots denote treated facilities (i.e., those in the Tonkolili, Falaba, Karene, Kono, and Pujehun districts), and purple dots denote control facilities (i.e., those in the Kailahun, Kenema, Bombali, Koinadugu, Kambia, Port Loko, Bo, Bonthe, Moyamba, Western Area Rural, and Western Area Urban districts). Instead, examined the equivalent objective of maximizing patient co… view at source ↗
Figure 3
Figure 3. Average normalized consumption time trends. Panel (a) compares treatment (green) versus control (orange) groups using raw data. Panel (b) compares treatment (green) versus synthetic control (orange) groups constructed via SynthDiD. The x-axis shows time in quarters and the y-axis denotes average normalized consumption. The vertical dashed line at 2023 Q1 marks the quarter immediately before deployment. was deployed.… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

90 extracted references · 1 canonical work pages

  1. [1]

    Yadav,Health systems & reform1, 142 (2015)

    P. Yadav,Health systems & reform1, 142 (2015). 16

    2015

  2. [2]

    UNICEF,et al.,The State of the World’s Children 2015: Reimagine the Future-Innovation for Every Child (Executive Summary)(UN, 2015)

    2015

  3. [3]

    Gallien, N.-H

    J. Gallien, N.-H. Z. Leung, P. Yadav,Production and Operations Management30, 4501 (2021)

    2021

  4. [4]

    Merkuryeva, A

    G. Merkuryeva, A. Valberga, A. Smirnov,Procedia Computer Science149, 3 (2019)

    2019

  5. [5]

    Yenet, G

    A. Yenet, G. Nibret, B. A. Tegegne,ClinicoEconomics and Outcomes Researchpp. 443–458 (2023)

    2023

  6. [6]

    De Fauw,et al.,Nature medicine24, 1342 (2018)

    J. De Fauw,et al.,Nature medicine24, 1342 (2018)

    2018

  7. [7]

    Esteva,et al.,Nature medicine25, 24 (2019)

    A. Esteva,et al.,Nature medicine25, 24 (2019)

    2019

  8. [8]

    K. Y . Ngiam, W. Khor,The Lancet Oncology20, e262 (2019)

    2019

Show all 90 references
  1. [9]

    Bastani,et al.,Nature599, 108 (2021)

    H. Bastani,et al.,Nature599, 108 (2021)

    2021

  2. [10]

    Yang,et al.,NPJ digital medicine5, 194 (2022)

    X. Yang,et al.,NPJ digital medicine5, 194 (2022)

    2022

  3. [11]

    S. M. Zuma, L. M. Modiba,World J Pharm Res8, 1532 (2019)

    2019

  4. [12]

    rep., World Health Organization (2018)

    World Health Organization, Service availability and readiness assessment (SARA): An annual monitoring system for service delivery,Tech. rep., World Health Organization (2018)

    2018

  5. [13]

    Droti, K

    B. Droti, K. P. O’Neill, M. Mathai, D. Y . T. Dovlo, J. Robertson,BMJ Global Health4, e001306 (2019)

    2019

  6. [14]

    Tetteh,Research in Social and Administrative Pharmacy5, 286 (2009)

    E. Tetteh,Research in Social and Administrative Pharmacy5, 286 (2009)

    2009

  7. [15]

    Terwiesch, G

    C. Terwiesch, G. Cachon,Matching supply with demand: an introduction to operations management(McGraw-Hill, 2006). 17

    2006

  8. [16]

    T. G. Fenta,Ethiopian Journal of Health Development31, 266 (2017)

    2017

  9. [17]

    Johnson, K

    A. Johnson, K. Peter, M. Shital,Rwanda Journal of Medicine and Health Sciences4, 257 (2021)

    2021

  10. [18]

    Mwashiuya, A

    S. Mwashiuya, A. D. Mchopa, F. A. Shayo,African Journal of Procurement, Logistics & Supply Chain Management8, 01 (2025)

    2025

  11. [19]

    J. Braa, H. Muquinge,IST Africa3(2007)

    2007

  12. [20]

    Donnelly,The Lancet377, 1393 (2011)

    J. Donnelly,The Lancet377, 1393 (2011)

    2011

  13. [21]

    Linnander,et al.,PloS one12, e0186832 (2017)

    E. Linnander,et al.,PloS one12, e0186832 (2017)

    2017

  14. [22]

    Linnander, K

    E. Linnander, K. LaMonaca, M. A. Brault, M. Vyavahare, L. Curry,International Journal of Multiple Research Approaches10, 136 (2018)

    2018

  15. [23]

    rep., Project Last Mile (2023)

    Project Last Mile, 2023 annual report,Tech. rep., Project Last Mile (2023)

    2023

  16. [24]

    X. Zhu, A. Ninh, H. Zhao, Z. Liu,Production and Operations Management30, 3231 (2021)

    2021

  17. [25]

    Mbonyinshuti, J

    F. Mbonyinshuti, J. Nkurunziza, J. Niyobuhungiro, E. Kayitare,Processes10, 26 (2021)

    2021

  18. [26]

    Bastani, O

    H. Bastani, O. Bastani, T.-H. Chung,Responsible and Sustainable Operations: The New Frontier(Springer, 2024), pp. 187–202

    2024

  19. [27]

    Bertsimas, A

    D. Bertsimas, A. Thiele,Operations research54, 150 (2006)

    2006

  20. [28]

    C. Yang, Z. Hu, S. X. Zhou,Management science67, 185 (2021)

    2021

  21. [29]

    Aviv,Management science47, 1326 (2001)

    Y . Aviv,Management science47, 1326 (2001)

    2001

  22. [30]

    USAID — DELIVER PROJECT, Task Order 4,Quantification of Health Commodities: A Guide to Forecasting and Supply Planning for Procurement, Arlington, Va (2014). 18

    2014

  23. [31]

    Reproductive and Maternal Health Services Unit, Ministry of Health,National Guidelines for Quantification, Procurement, and Pipeline Monitoring for Family Planning Commodities in Kenya, Ministry of Health, Nairobi, Kenya (2016)

    2016

  24. [32]

    M. W. Fredrick, W. Muturi,IOSR J Business Manag18, 63 (2016)

    2016

  25. [33]

    G.-Y . Ban, C. Rudin,Operations Research67, 90 (2019)

    2019

  26. [34]

    Dhis2 — digital health information system 2,https://dhis2.org/(2025)

    2025

  27. [35]

    Caruana,Machine learning28, 41 (1997)

    R. Caruana,Machine learning28, 41 (1997)

    1997

  28. [36]

    Donti, B

    P. Donti, B. Amos, J. Z. Kolter,Advances in neural information processing systems30 (2017)

    2017

  29. [37]

    A. N. Elmachtoub, P. Grigas,Management Science68, 9 (2022)

    2022

  30. [38]

    Bertsimas, N

    D. Bertsimas, N. Kallus,Management Science66, 1025 (2020)

    2020

  31. [39]

    Wilder, B

    B. Wilder, B. Dilkina, M. Tambe,Proceedings of the AAAI Conference on Artificial Intelli- gence(2019), vol. 33, pp. 1658–1665

    2019

  32. [40]

    K. Wang, B. Wilder, A. Perrault, M. Tambe,Advances in Neural Information Processing Systems33, 9586 (2020)

    2020

  33. [41]

    S. Shah, K. Wang, B. Wilder, A. Perrault, M. Tambe,Advances in Neural Information Processing Systems35, 1320 (2022)

    2022

  34. [42]

    Huang, N

    D. Huang, N. Stein, D. B. Rubin, S. Kou,Proceedings of the National Academy of Sciences 117, 12004 (2020)

    2020

  35. [43]

    Arkhangelsky, S

    D. Arkhangelsky, S. Athey, D. A. Hirshberg, G. W. Imbens, S. Wager,American Economic Review111, 4088 (2021). 19

    2021

  36. [44]

    worldbank.org/indicator/SP.DYN.CBRT.IN?locations=SL (2025)

    World Bank, Crude birth rate (per 1,000 people) - sierra leone, https://data. worldbank.org/indicator/SP.DYN.CBRT.IN?locations=SL (2025). Ac- cessed: 2025-08-21

    2025

  37. [45]

    UNICEF, Sierra leone: Demographics, health & infant mortality — country profile, https: //data.unicef.org/country/sle/(2025)

    2025

  38. [46]

    J. P. Souza,et al.,The Lancet Global Health12, e306 (2024)

    2024

  39. [47]

    World Health Organization, Maternal mortality: Fact sheet (2024)

    2024

  40. [48]

    Shafiq,et al.,BMJ open14, e076256 (2024)

    Y . Shafiq,et al.,BMJ open14, e076256 (2024)

    2024

  41. [49]

    J. R. Birge, F. Louveaux,Introduction to stochastic programming(Springer Science & Business Media, 2011)

    2011

  42. [50]

    Gilbert,Management Science51, 305 (2005)

    K. Gilbert,Management Science51, 305 (2005)

    2005

  43. [51]

    Bastani,Management Science67, 2964 (2021)

    H. Bastani,Management Science67, 2964 (2021)

    2021

  44. [52]

    K. Xu, H. Bastani,Management Science(2025)

    2025

  45. [53]

    D. B. Rubin,Biometrika63, 581 (1976)

    1976

  46. [54]

    Kallus, X

    N. Kallus, X. Mao,Management Science(2022)

    2022

  47. [55]

    D. Card, A. B. Krueger, Minimum wages and employment: A case study of the fast food industry in new jersey and pennsylvania (1993)

    1993

  48. [56]

    Abadie, A

    A. Abadie, A. Diamond, J. Hainmueller,Journal of the American statistical Association 105, 493 (2010). 20

    2010

  49. [57]

    Clarke, D

    D. Clarke, D. Paila˜nir, S. Athey, G. Imbens, Synthetic difference in differences estimation (2023)

    2023

  50. [58]

    Candes, B

    E. Candes, B. Recht,Communications of the ACM55, 111 (2012)

    2012

  51. [59]

    Https://www.msupply.org.nz

    mSupply (2024). Https://www.msupply.org.nz

    2024

  52. [60]

    Sierra Leone Ministry of Health and Sanitation (MoHS), Center for International Earth Science Information Network (CIESIN), Columbia University, Sierra leone national health facilities dataset version 01 (2023)

    2023

  53. [61]

    D. W. Martin,et al.,Health security18, S (2020)

    2020

  54. [62]

    WorldPop, Worldpop global project population data: Estimated residential population per 100x100m grid square,https://www.worldpop.org(2021)

    2021

  55. [63]

    Weiss,et al.,Nature medicine26, 1835 (2020)

    D. Weiss,et al.,Nature medicine26, 1835 (2020)

    2020

  56. [64]

    Didan, MOD13Q1 MODIS/Terra Vegetation Indices 16-Day L3 Global 250m SIN Grid V006 [Data set],https://doi.org/10.5067/MODIS/MOD13Q1.006(2015)

    K. Didan, MOD13Q1 MODIS/Terra Vegetation Indices 16-Day L3 Global 250m SIN Grid V006 [Data set],https://doi.org/10.5067/MODIS/MOD13Q1.006(2015)

    2015 doi

  57. [65]

    Besbes, A

    O. Besbes, A. Muharremoglu,Management Science59, 1407 (2013)

    2013

  58. [66]

    W. H. Greene,International edition, New Jersey: Prentice Hallpp. 201–215 (2000)

    2000

  59. [67]

    php/the-nmsa-act/(2017)

    NMSA, National medical supplies agency act, 2018, https://nmsa.gov.sl/index. php/the-nmsa-act/(2017)

    2018

  60. [68]

    G. A. Stevens,et al.,The Lancet388, e19 (2016)

    2016

  61. [69]

    Statistics Sierra Leone, 2015 population and housing census,Census report, Statistics Sierra Leone, Freetown, Sierra Leone (2015). 21

    2015

  62. [70]

    Humanitarian Data Exchange (HDX), HDX HAPI Population Dataset (2023)

    2023

  63. [71]

    Kotary, F

    J. Kotary, F. Fioretto, P. Van Hentenryck, B. Wilder,arXiv preprint arXiv:2103.16378 (2021)

    2021 arXiv

  64. [72]

    Agrawal,et al.,Advances in neural information processing systems32(2019)

    A. Agrawal,et al.,Advances in neural information processing systems32(2019)

    2019

  65. [73]

    W. H. O. W. Team, Allocation logic and algorithm to support allocation of vaccines secured through the covax facility,Tech. rep., World Health Organization (2021)

    2021

  66. [74]

    G. D. Lab, Global data lab — open data on human development, health, education, wealth, and more,https://globaldatalab.org/(2025)

    2025

  67. [75]

    rep., GHNDR and GEDDR Africa, World Bank (2018)

    Africa health service delivery in sierra leone: Results of 2018 service delivery indicator survey,Tech. rep., GHNDR and GEDDR Africa, World Bank (2018)

    2018

  68. [76]

    Last updated May 11, 2025

    SimpleMaps, World cities database, https://simplemaps.com/data/ world-cities (2025). Last updated May 11, 2025. License: Creative Commons Attribution 4.0

    2025

  69. [77]

    Y . Yang, A. V . Diez-Roux,American journal of preventive medicine43, 11 (2012)

    2012

  70. [78]

    O. C. Ashenfelter, D. Card, Using the longitudinal structure of earnings to estimate the effect of training programs (1984)

    1984

  71. [79]

    Angrist, G

    J. Angrist, G. Imbens, Identification and estimation of local average treatment effects (1995)

    1995

  72. [80]

    WHO, Global health estimates: Leading causes of dalys, https://www.who.int/ data/gho/data/themes/mortality-and-global-health-estimates/ global-health-estimates-leading-causes-of-dalys(2024). 22

    2024

  73. [81]

    ICER, Perspectives on cost-effectiveness threshold ranges,White paper / seminar report, Institute for Clinical and Economic Review (2019)

    2019

  74. [82]

    U. D. o. V . A. Health Economics Resource Center (HERC), Cost-effectiveness analysis, https://www.herc.research.va.gov/include/page.asp?id= cost-effectiveness-analysis

  75. [83]

    P. J. Neumann, J. T. Cohen, M. C. Weinstein,et al.,N Engl J Med371, 796 (2014)

    2014

  76. [84]

    Alt. Control

    J. Ludwig, S. Mullainathan, A. Rambachan,AEA Papers and Proceedings(American Economic Association 2014 Broadway, Suite 305, Nashville, TN 37203, 2024), vol. 114, pp. 623–627. End Notes AcknowledgmentsWe are grateful for the close partnership offered by the Sierra Leone Ministr...

    2014

  77. [85]

    First, we collected data on geographic coordinates of health facilities in Sierra Leone from several sources, including Google Maps and Geo-Referenced Infrastructure and Demographic Data for Development (GRID3) (60)

  78. [86]

    NDVI serves as a proxy for vegetation density, which can indicate human activity.2

    Next, we used Google Earth engine’s satellite imagery datasets to compute the normalized difference vegetation index (NDVI) on a 10km ×10km patch around each facility at monthly resolution between January 2022 and December 2022. NDVI serves as a proxy for vegetation density, w...

    2022

  79. [87]

    friction surface

    Then we used “friction surface” data from Google Earth (63) to obtain the travel time between every facility and pixel of the area with potential human activity. This allows us to define the catchment area on the basis of minimal travel time

  80. [88]

    overshooting

    Finally, we estimated pn for health facility n by using data from WorldPop (62), which provides population count estimates for each 100m×100m grid cell. We estimated r, the proportion of women and children, using 2015 Sierra Leone Census data (69) at the chiefdom level.3 Then,...

    2015

  81. [89]

    Missingness

    Missingness as the outcome variable: We used a missingness indicator as the dependent variable in our main specification; results are shown in the “Missingness” row of Table SI 6, and show no significant treatment effect. These results suggest that our intervention did not cau...

  82. [90]

    No Missingness Imbalance

    Analysis on a sub-sample with no missingness imbalance: We also re-ran our main analysis on a subset of products that showed no significant difference in missing data rates between the treatment and control groups after the treatment. Results are shown in the “No Missingness I...

    2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.