Where health-economic evidence exists — and where it is missing
The evidence base at a glance
Country landscape
This tab is unfiltered: it always shows the whole corpus and does not read or change the Explorer filters. Click a bubble to select a country and fill the two mix panels and the disease-group × evaluation-type grid below. Click a heatmap cell to read the studies behind it. You can also search for a country with the picker below.
Each cell is a share of its disease-group row, so a row with any studies sums to 100% across the evaluation types. A study can sit in several disease groups, so rows can overlap and their totals exceed the country’s study count. Click a cell to open the studies behind it.
The map covers full economic evaluations of health-care delivery published from 1 January 2010 onward: studies that compare two or more ways of delivering care on both costs and consequences. The unit of analysis is the study; the analysis is descriptive and, where it compares research against disease burden, ecological.
Title/abstract screening was machine-learning based, with a measured sensitivity of 0.959 against the current corrected human reference vintage (93/97; 95% CI 0.899–0.984). Partial evaluations (cost-minimisation, budget-impact, cost-of-illness) and records unclear at extraction are excluded, so the analysis population is full evaluations by construction. Studies whose consequence is a non-clinical metric (e.g. diagnostic accuracy, utilisation) are kept and flagged. Reviews synthesising published economic evaluations are retained by extraction design (framework v6.8 codes them as literature/registry synthesis); 1,602 of the 38,199 (4.2%) are title-flagged reviews, 65% of which name no single country, so country-level panels draw mostly on articles.
| Selection funnel | Records |
|---|
| Input | Source |
|---|---|
| Study metadata & bibliometrics | CrossRef + OpenAlex + PubMed + NLM MeSH harvest (~93% coverage of the eligible population) |
| Income groups | World Bank API classification |
| Population | World Bank SP.POP.TOTL (latest observation, mostly 2025) |
| Disease burden | IHME GBD 2023 DALYs (country totals + 22 Level-2 causes) |
| Health spending | IHME Health Spending 1995–2022 (total health expenditure per capita, PPP 2022; national total = per-capita × latest population) |
| Disease coding | MeSH C-tree any-topic headings (2026 vocabulary, incl. the C12.050 pregnancy-complications branch and F03) mapped to 12 groups, crosswalked to GBD Level-2 causes |
| Research topics | OpenAlex topic taxonomy plus an emergent theme model (abstract embeddings → UMAP → HDBSCAN) |
Intervention type and care setting were extracted for every eligible study by a single-pass language model (validated against a keyword lexicon, the emergent-topic clusters and a second model lineage; no human validation available).
Reference data are fetched from dated sources, never bundled snapshots.
Extracted geography only. The author-affiliation proxy is accurate where it applies, but its availability is income-biased — merging it would inflate the headline burden gap from 17.8× to 25.8×, so it is excluded from every income figure.
Multi-country studies count once per country in the country panels, and every figure declares whether its denominator is studies or study–country pairs.
Countries without a World Bank income group are visibly absent, never zero-filled as “0 studies”.
Burden proportionality is a reference point, not a normative target: research allocation also reflects tractability, intervention cost and the stock of existing evidence.
Abstract decision-signals measure what abstracts report, not what studies did — “not reported” is not “not done”.
MeSH uses any-topic headings: major-topic marking jumps at the 2020 harvest boundary and would manufacture a trend.
The corpus is effectively English-only (99.8%). Geography is missing for ~33% of studies; country-level figures cover the placeable subset. Harvest coverage is partial (47–93% depending on field) and declared per figure. Comparisons of research to burden are ecological. Extraction fields disagree for 8.3% of studies, handled by conservative disqualification. The 50,661 records unsure at title/abstract are held out of the population pending a handling policy.