How the panel is assembled

Our project follows the effect of the reorganisation of United States global health assistance on HIV services in LMICs. Measuring that effect requires data that represent two things: (1) what donors report spending, and (2) what the services themselves report doing. Neither alone is enough, because funding data are timely but say nothing about delivery, while programme data describe delivery but its trackability is not real-time and is unevenly distributed geographically.

Principles

We follow four rules that govern what enters the panel (/what data are collected).

• Structured sources only. Figures come from machine-readable feeds and from published tables that can be read deterministically. Prose is never parsed for numbers. Scanned documents are out of scope, because reading them reliably would require optical character recognition and manual checking that the project cannot guarantee at scale.
• No language models anywhere in the pipeline. Every figure is traceable to a field in a file or a cell in a published table. An honest record that a country does not publish in a usable form is more useful than a plausible number without a traceable origin. Language models are not used for data extraction to minimise data contamination and to maximise the validity and reliability of our project.
• Data origin on every row. Each observation carries its source, the URL it was read from, the source's own indicator code and the date of retrieval.
• Uncertainty is kept. Modelled estimates carry lower and upper bounds and those bounds are stored and drawn. The point estimates for the 95-95-95 targets in particular carry intervals wide enough that a single number misrepresents them.

What is read, and from where

We designed the pipeline to build the panel with two layers. The aggregator layer is near-global and gives the panel its skeleton. Data in this layer are sourced primarily from UNAIDS AIDSinfo, the WHO Global Health Observatory, the World Bank, the OECD Creditor Reporting System, and United States self-reported disbursements from ForeignAssistance.gov and programme results from PEPFAR Panorama. The country layer is the national programme record, read directly where a country publishes it in a usable form.

Going to source, country by country

The question we encountered at the pilot stage was whether national programme data can realistically be read at source, country by country. The answer varies notably, and the variation is itself a finding. All nine Southeast Asian countries run a national HIV programme database. Data exist in every country, but some do not allow foreign or non-institutional access.

• An open structured feed, as in Thailand, where the national programme's data mart answers queries directly.
• An open born-digital report whose tables can be read without interpretation, as in Cambodia and the Philippines.
• A closed database that releases only scanned documents or its submission to the international aggregators. For these countries the realistic route is an institutional data-sharing agreement, not public collection.

We decided whether a national report is worth reading at all by one factor. A report only adds something if it carries data the aggregators do not already receive, such as sub-national detail, more recent periods, key-population breakdowns, or programme flows. The Philippine quarterly surveillance report, for example, qualifies. It is regional, quarterly, and richer than the annual report the Philippines submits to UNAIDS. Malaysia's national report does not, because that report is itself the country's annual submission to UNAIDS, so reading it would return figures already in the panel through UNAIDS.

In the pilot stage (i.e., exclusively extracting data from the Southeast Asia region), three national cascades and one community layer are now read at source, by three different routes. These are a live programme interface in Thailand, published report tables in Cambodia and the Philippines, and a community-led monitoring interface covering Indonesia.

Reporting periods

Funding and programme data are filed in periods, and a period that is still open for data collection is likely subject to a figure drop when the data collected are only provisional but not full. For example, United States disbursements for fiscal year 2026 are still being reported, and PEPFAR results for 2025 cover one of four quarters. Points from an incomplete period are drawn with an open red marker and named in the note under every figure. They are never compared with complete years.

Comparing sources

Where two sources report the same country, indicator and year, both values are kept and the pair is compared. The comparison is classified before it is read, because agreement between two sources that publish the same underlying estimate is arithmetic, not corroboration. The classification method, the thresholds and every verdict with its evidence are set out on the reconciliation page.

What the panel does not support

The figures describe reported funding and reported service activity. They do not, on their own, establish that a change in one caused a change in the other. Reporting lags differ by source and by country, so a decline that appears in one series and not another may be a difference in filing but not in delivery. Programme footprints are not national totals. Here, PEPFAR reports what PEPFAR-supported sites did, which is a part of a country's response, not the entirety of it. Community-led monitoring covers a small number of facilities and is a sentinel signal, not a representative sample.

Reproducing the panel

We design the pipeline as a Python package with one connector per source and a single assembly step. The pipeline writes the panel to CSV, Parquet and SQLite, runs the reconciliation, and then we build this site from those files alone. The panel on this site was last assembled on 29 July 2026.