The Fabric of Global Health Funding

Nodes —

Edges —

Density —

Clustering † —

WCC (Frags) —

Modularity —

          † Clustering computed on unipartite projection (hover label for details)
        

Introduction

A Network in Transition

Global health financing usually appears as a map of countries or a list of budgets. Beneath that surface sits a network: financial threads connecting donors to the programmes and regions they support.

This essay traces how that fabric has evolved through expansion, pandemic shock, and geopolitical shifts.

Methodological Caution:

N = 5 biennia;
descriptive visual reading;
no causal claim; no prediction;
no formal proof of tipping point, hysteresis or critical slowing down.

Scene 01 — 2016–17: Integrated baseline

A Cohesive Starting Point

In the 2016–17 biennium, the network is relatively integrated. Most actors are connected within a few degrees of each other. The "Weakly Connected Components" (WCC) count is low, indicating a single primary fabric.

WCC: 2 (A nearly single connected component)
Clustering: 0.043 (Higher local density)

⚠ Methodological note — Clustering: This is a bipartite (2-mode) network. Triangles are structurally impossible. Donors A and B cannot form a triangle through a shared recipient because they belong to the same mode. The clustering coefficient was computed treating the network as unipartite (1-mode projection) and reflects co-funding overlap patterns rather than triadic closure in the strict graph-theoretic sense. Values are retained for comparability with the published paper.

Scene 02 — 2018–19: Early fragmentation

First Signs of Structural Drift

Even before the pandemic, the network begins to show early signs of fragmentation. While the volume of funding grows, the number of isolated fragments (WCC) starts to rise.

WCC: 2 → 123
Modularity: 0.466

The shift is structural. The network is sorting into a core-periphery pattern (Knoke, Diani, Hollway & Christopoulos, 2021): a dense inner cluster of large donors that co-finance the same cross-cutting programmes, surrounded by a growing periphery connected only to narrow thematic or regional niches. The 123 isolated fragments are peripheral actors with no path into the core.

Scene 03 — 2020–21: Expansion without binding

More knots. More threads. Less fabric.

During the COVID-19 period, the network expands dramatically: 60% more actors, 36% more financial links. Every measure of structural cohesion deteriorates at the same time.

Nodes: 624 → 999 (+60%)
Edges: 3,277 → 4,444 (+36%)
Density: 0.008 → 0.004 (−50%)
Clustering: 0.050 → 0.012 (−76%)
WCC: 123 → 546 (+344%)

The fabric expands and thins at once. New emergency actors enter; existing ties multiply. The structure loses its reinforcing patches and breaks into 546 loosely connected pieces.

⚠ Methodological note — Modularity: The Walktrap modularity values are computed on the bipartite network treated as unipartite. Modularity is sensitive to network size and node composition. In 2018–19 and 2020–21, significantly fewer nodes carry the "Recep_State" prefix due to a source-data formatting inconsistency; in 2020–21 many new emergency actors also enter. These compositional shifts mean cross-biennium modularity values are directionally informative, not strictly comparable in magnitude.

Scene 04 — 2022–23: Partial recovery

The fabric partly re-stitches, but not to its original texture

Post-pandemic, the network contracts and partially reorganises. Density and average degree recover (more threads per knot). The recovery is genuine but incomplete. What the data show looks more like hysteresis, or structural drift, than a return to the 2016–17 baseline.

Avg. degree: 4.45 → 7.69 (highest observed)
Density: 0.013 (above 2016–17 baseline of 0.009)
Clustering: 0.014 (vs. 0.043 in 2016–17)
WCC: 121 (vs. 2 in 2016–17)

Scene 05 — 2024–25: A New Stability

Stabilizing in a Modular Form

The 2024–25 biennium (provisional) shows the network settling into this more fragmented shape. Some connectivity has returned, but 126 weakly connected components remain. The structure did not bounce back.

Density: 0.012
WCC: 126
Clustering: 0.018

2025: Exposure without collapse

US entities (USAID, CDC, NIH, DTRA, USFDA, USDA, State Department) contributed $525M in the 2024–25 period (provisional): 10.9% of total flows, across 1,530 financing relationships.

At this scale of aggregation, removing US flows does not collapse the network visually. The fabric holds. But some nodes become exposed once US-linked flows are faded.

Click to reveal US-linked exposure

Baseline (Full 2024–25)

Exposed (US Flows Faded)

Layout stable. US-linked flows faded to reveal structural exposure.

Finding: No node in this dataset receives flows exclusively from US entities. However, four recipient nodes rely on US sources for 50–74% of their observed flows: Poland, Czechia, Slovakia, and Iraq. The fabric holds, but its texture thins in specific geographic and thematic patches.

Recomputed network metrics after US node removal (full network, 2024–25)

Metrics recomputed on the full subgraph with all 9 US-linked nodes and their 1,530 edges removed. This is a full structural recalculation, not a visual approximation.

Nodes remaining

431

of 564 (−24%)

Edges remaining

10,562

of 12,092 (−13%)

WCC after removal

1

vs 126 before (more coherent)

Flow removed

$525M

10.9% of Q1 2025 total

The WCC paradox. Removing US nodes should fragment the network: the largest donor cluster leaves, the fabric should break apart. The opposite happens. WCC drops from 126 to 1.

The drop is a structural trap. Those 125 isolated components existed because of US-linked ties. Remove USAID and its constellation of bilateral connections, and the satellites disappear with it. What remains is the European-multilateral core: tightly connected, but on its own.

The WCC count conceals the real loss. USAID's Participation Coefficient was 0.85, meaning it distributed funding across more programme communities than any other actor in the network: health emergencies, routine immunisation, non-communicable diseases, capacity building, regional programmes. No remaining actor comes close to that breadth.

Middle Powers: Potential Bridges?

These actors hold positions in the network that may carry bridging potential after the US withdrawal. The question is whether any of them already functions as a cross-community connector, in the way USAID did. The data give a mixed answer.

The Participation Coefficient (PC) measures how broadly each actor spreads its funding across the network's structural communities (Guimerà & Amaral, 2005). A score near 0 (a bonder) means almost all funding flows into one cluster. A score near 1 (a bridge) means funding reaches across many communities. In the brokerage framework of Gould & Fernandez (1989), a bridge actor connects communities that would otherwise have no path between them; remove it, and those communities lose contact (see also Knoke, Diani, Hollway & Christopoulos, 2021, ch. 3).

The table shows structural roles, not a ranking by size or importance.

Actor	Type	Flow 2022–23	PC (2022–23)	Structural role
USAID	US Entity	$437M	0.782	Bridge — global coordinator
EU DG-INTPA	Intergovernmental	$167M	0.722	Bridge
UN CERF	UN Organization	$140M	0.725	Bridge
EU ECHO	Intergovernmental	$149M	0.694	Bridge
France (MEAE)	Member State	$63M	0.695	Bridge
Canada (DFATD)	Member State	$149M	0.687	Bridge
Germany	Member State	$335M	0.646	Bridge — thematic concentration
Rotary International	NGO	$153M	0.464	Moderate
India	Member State	$66M	0.049	Bonder — regional cluster
China	Member State	—	0.108	Bonder — regional cluster

India and China are often proposed as candidates to fill the gap left by US withdrawal. By volume, that case has some surface appeal. By structure, it breaks down quickly. PC=0.049 and PC=0.108 mean that nearly all of their WHO financing flows into one modularity community: the cluster of programmes and recipient states tied to their immediate geopolitical neighbourhood. This is strategic, concentrated giving, and a coherent policy choice. The problem is that concentrated giving and cross-community bridging are structurally incompatible. A bonder can grow larger; it cannot become a bridge without changing where it gives, not how much.

The relevant question is who gives differently. Currently, no one does.

PC = 1 − Σ_c(k_ic/k_i)², where k_ic = weighted flow to community c, k_i = total weighted out-flow (Guimerà & Amaral, 2005). Communities from Walktrap modularity classes on the full network. Hover donor nodes to see PC values.

Based on the article

Santos Domínguez, A. B., & Ballesteros Pérez, C. E. (2026). Network Fragmentation and the 2025 Funding Shock: Early Warning Signs of a Systemic Risk in Global Health Governance. McGill Journal of Global Health, 15(1), 23–31.
DOI: 10.26443/mjgh.v15i1.1926

Read the full article

Methodological references

Csardi, G. & Nepusz, T. (2006). The igraph software package for complex network research. InterJournal, Complex Systems, 1695.
Gould, R. V. & Fernandez, R. M. (1989). Structures of mediation: A formal approach to brokerage in transaction networks. Sociological Methodology, 19, 89–126.
Guimerà, R. & Amaral, L. A. N. (2005). Functional cartography of complex metabolic networks. Nature, 433, 895–900.
Knoke, D., Diani, M., Hollway, J. & Christopoulos, D. (2021). Multimodal Political Networks. Cambridge University Press.