← Papers

A 2025 USAID funding freeze exposed maintenance fragility in open-source health software: revenue-funded projects slowed bug fixes after the shock, donor-funded projects stayed stable, and mixed-funded projects weathered the immediate hit but weakened over time.

Maintenance Dynamics in Digital Health Public Goods: A Quasi-Experimental Study of Issue Resolution During the 2025 USAID Funding Disruption

Jinyou Sheng, Amy Finnegan · September 04, 2026 · Oxford Open Digital Health

openalex quasi_experimental medium evidence 7/10 relevance Summary only summary available; pdf_status=paywall DOI Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation: September 7, 2026

Jinyou Sheng provider ID
Amy Finnegan provider ID

Using the January 2025 USAID funding freeze as a natural experiment, the study finds that revenue-driven open-source maintainers experienced post-shock deterioration in bug-resolution efficiency, donor-driven maintainers remained stable, and hybrid organizations showed short-term resilience but gradual decline over time.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

0 cumulative citations

OpenAlex · Observed September 7, 2026

View corpus context

0 cumulative citations

Semantic Scholar · Observed September 7, 2026

View corpus context

Abstract Open-source digital public goods play a growing role in point-of-care health service delivery in low- and middle-income countries (LMICs). Many of these systems are developed and maintained by organizations that rely heavily on donor funding, making their long-term sustainability vulnerable to funding instability. The January 2025 U.S. Agency for International Development (USAID) funding freeze created a rare natural experiment to observe how different organizations maintaining these systems respond under external funding shocks. Using 7,922 issues from eight repositories created between June 2023 and June 2025, a DistilBERT model was fine-tuned to classify bug severity. Organizations were categorized into three funding-based models: donor-driven (primarily grant-funded), revenue-driven (sustained through commercial or internal investment), and hybrid (combining grants and revenue). Interrupted time series (ITS) and Cox proportional hazards models were then applied to evaluate changes in bug-resolution efficiency. Revenue-driven organizations did not exhibit a statistically significant immediate level change but experienced significant deterioration in the post-intervention period. Donor-driven organizations exhibited no substantial changes following the intervention, maintaining stable baseline levels. Hybrid organizations showed resilience to the immediate funding shock, with a gradual decline in bug resolution as time went on after the shock. These findings suggest that the funding shocks function as a stress test for digital public goods and highlight the resilience advantages of various governance models.

Summary

Main Finding

A natural experiment using the January 2025 USAID funding freeze shows that funding model matters for the operational resilience of open-source digital public goods used in point-of-care health delivery in LMICs. Revenue-driven maintainers experienced significant post-shock deterioration in bug-resolution efficiency, donor-driven maintainers remained broadly stable, and hybrid organizations showed short-term resilience but gradual decline over time.

Key Points

Dataset: 7,922 issue reports from eight repositories (June 2023–June 2025).
Natural experiment: USAID funding freeze (Jan 2025) used as an exogenous shock.
Organizations classified into three funding models:
- Donor-driven (primarily grant-funded)
- Revenue-driven (commercial or internally financed)
- Hybrid (combining grants and revenue)
NLP: a DistilBERT model was fine-tuned to classify bug severity from issue text, enabling comparable outcome measurement across repositories.
Outcomes and models:
- Interrupted time series (ITS) to detect level/trend changes around the shock.
- Cox proportional hazards models to assess time-to-bug-resolution (bug-resolution efficiency).
Empirical results:
- Revenue-driven: no statistically significant immediate level change at the shock, but significant deterioration in bug-resolution efficiency in the post-intervention period (bugs took longer to resolve).
- Donor-driven: no substantial changes detected — baseline levels of bug resolution remained stable through the shock.
- Hybrid: resilient to immediate shock (no abrupt drop), but showed a gradual decline in bug-resolution efficiency as the post-shock period continued.
Interpretation: the funding shock acted as a stress test revealing differences in resilience across governance/funding models.

Data & Methods

Sample: 8 open-source repositories supporting point-of-care health systems in LMICs; issues dated June 2023–June 2025; total issues = 7,922.
Outcome: bug-resolution efficiency operationalized via classified bug severity and time-to-resolution of issues.
Text processing & classification:
- Fine-tuned DistilBERT (transformer-based language model) to label issues by severity, allowing stratified analysis and adjustment for issue-criticality.
Causal inference / time series methods:
- Interrupted time series (ITS) to estimate immediate (level) and slope (trend) changes pre- vs post- Jan 2025 funding freeze.
- Cox proportional hazards models to estimate hazard ratios for bug closure before vs after the shock, controlling for covariates (severity, repo fixed effects, etc.).
Robustness and limitations noted in the study:
- Potential misclassification from automated severity labeling.
- Limited number of repositories (n=8) — limits external generalizability.
- Possible confounding co-occurring events and heterogeneity within funding categories.
- Relatively short post-intervention follow-up for long-run effects.

Implications for AI Economics

Funding model matters for operational resilience: diversified (hybrid) funding provides short-term insulation; reliance on revenue alone may risk longer-term maintenance degradation under funding stress.
Policy and donor design:
- Grants and donor arrangements can stabilize maintenance outcomes if structured to support continuity (not only feature development).
- Encourage funding diversification (hybrid models) and mechanisms to underwrite maintenance risk (contingency reserves, insurance-like funds, milestone-based multi-year commitments).
Value of automated monitoring: ML/NLP (e.g., DistilBERT) can cost-effectively monitor software-health metrics (bug severity, time-to-resolution) at scale — useful for real-time performance-based funding and early warning systems.
Incentives and private sector engagement:
- Revenue-driven models may prioritize product development or monetizable features and underinvest in routine maintenance when resources tighten; contract and incentive designs should explicitly cover maintenance public-good externalities.
Research and evaluation:
- Use exogenous shocks and ML-enabled operational metrics to evaluate sustainability interventions and the marginal value of different funding mixes.
- Further work needed on causal mechanisms (why donor-driven remained stable), generalizability beyond USAID and the eight repos, and extension to other operational outcomes (security, uptime, feature backlog).
Practical recommendations:
- For donors: build multi-year, maintenance-focused funding lines and encourage co-financing to create hybrid resilience.
- For maintainers/platforms: implement automated monitoring dashboards and contingency planning to detect and mitigate degradation early.

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — The design leverages a plausibly exogenous, well-defined shock and uses appropriate time-series and survival models with granular issue-level data (n=7,922), which strengthens causal claims; however, only eight repositories (limited clusters), a single shock event, potential co-occurring confounders, measurement error from automated severity labeling, and relatively short post-period reduce confidence in strong causal generalization. Methods Rigormedium — Appropriate and complementary methods (ITS and Cox models) and careful adjustment for severity via NLP improve internal validity, but concerns remain about small number of repositories (low cluster count), potential violations of ITS assumptions (no concurrent shocks), classifier misclassification, and limited robustness checks described in the summary. SampleIssue-level data (N=7,922) from eight open-source repositories that support point-of-care health systems in low- and middle-income countries, covering June 2023–June 2025; repositories are categorized by maintainer funding model into donor-driven, revenue-driven, and hybrid organizations; outcomes include NLP-classified bug severity and time-to-issue-resolution. Themesgovernance adoption org_design IdentificationUses the January 2025 USAID funding freeze as an exogenous shock in an interrupted time series (ITS) design, comparing pre/post level and slope changes in bug-resolution outcomes; complements ITS with Cox proportional hazards models for time-to-resolution, repo fixed effects, and controls (including automated severity labels from a fine-tuned DistilBERT) to adjust for issue criticality and observed heterogeneity across funding-model groups. GeneralizabilitySmall number of repositories (n=8) limits external validity to the broader universe of digital public goods, Sample restricted to point-of-care health systems in LMICs; findings may not generalize to other domains or high-income contexts, Shock is USAID-specific; other donor or market shocks may have different effects, Heterogeneity within funding-model categories (size, user base, technical architecture) could limit applicability of group-level results, Potential measurement error from automated severity classification may bias estimates for severity-stratified analyses, Relatively short post-intervention follow-up limits inference about long-run effects

Claims (6)

Claim	Direction	Outcome	Confidence & Evidence	Details
Revenue-driven maintainers experienced a significant post-shock deterioration in bug-resolution efficiency after the January 2025 USAID funding freeze. Task Completion Time	negative	Time from issue opening to bug resolution or closure	Reading fidelity high Study strength medium	n=7922 0.48
Revenue-driven maintainers did not show a statistically significant immediate level change in bug-resolution outcomes at the time of the funding freeze. Task Completion Time	null_result	Immediate post-freeze level of bug-resolution efficiency	Reading fidelity high Study strength medium	n=7922 0.48
Donor-driven maintainers remained broadly stable in bug-resolution outcomes through the USAID funding freeze, with no substantial changes detected. Task Completion Time	null_result	Bug-resolution efficiency and time-to-resolution	Reading fidelity high Study strength medium	n=7922 0.48
Hybrid organizations showed short-term resilience to the funding freeze, with no abrupt decline in bug-resolution efficiency, but experienced a gradual deterioration as the post-shock period continued. Task Completion Time	mixed	Bug-resolution efficiency and time-to-resolution over the post-freeze period	Reading fidelity high Study strength medium	n=7922 0.48
The funding model was associated with differences in operational resilience of open-source digital public goods used in point-of-care health delivery in low- and middle-income countries. Organizational Efficiency	mixed	Changes in bug-resolution efficiency following the funding shock	Reading fidelity high Study strength medium	n=7922 0.48
A fine-tuned DistilBERT model was used to classify bug severity from issue text, enabling severity-stratified and severity-adjusted comparisons of bug-resolution outcomes across repositories. Other	positive	Bug severity classification from issue reports	Reading fidelity high Study strength low	n=7922 0.24