Methodology

Every China figure on these pages is an estimate, not an observation. This page says who built it, how, and what it leaves out.

The sources

The debt-stock panel (126 borrowers, 2000 to 2021)

Horn, S., Reinhart, C. and Trebesch, C. (2021), “China’s Overseas Lending”, Journal of International Economics 133, 103539. The file used here is the November 2025 refresh, which builds on that paper and on the updated methodology in Franz, L., Horn, S., Parks, B., Reinhart, C. and Trebesch, C. (2025), “China as an International Lender of Last Resort”, mimeo. Downloaded from Christoph Trebesch’s data page at the Kiel Institute. The panel is balanced: one row per country-year, 126 countries times 22 years, with no missing values in any debt column.

The earlier debt-stock comparison (108 observed borrowers, 2000 to 2017)

The hrt_china_debt view comes from the Mendeley replication archive for Horn, Reinhart and Trebesch (2021), retrieved August 2026 under CC BY 4.0. Its country-year workbook grid begins in 1950, but the two China debt fields contain published observations only from 2000. It is therefore an earlier vintage comparison, not a 1950 historical China debt series.

The restructuring deals (55 deals, 29 countries, 2000 to 2020)

Horn, S., Reinhart, C. and Trebesch, C. (2022), “Hidden Defaults”, World Bank Policy Research Working Paper 9925, from the “Chinese Debt Restructurings” sheet of the authors’ 1,000-restructurings workbook.

Reserves and borrower-reported debt

Gross international reserves are World Bank World Development Indicators, FI.RES.TOTL.CD, total reserves including gold, in current US dollars. Reported external debt is the World Bank International Debt Statistics, filed by borrowers through the Debtor Reporting System. The GDP denominator used for the percent-of-GDP figures is WDI NY.GDP.MKTP.CD as distributed inside the China workbook itself, not re-joined from elsewhere.

How the earlier-vintage overlap is constructed

The archive’s own estimation code defines ExternalDebt_China as total external debt to all Chinese creditors in nominal US dollars. It defines PublicDebt_China as external public debt to all Chinese creditors, also in nominal US dollars. The refreshed workbook instead supplies external debt and public and publicly guaranteed debt in billions of current US dollars. Those public-sector concepts and the two loan universes are not assumed to be identical.

A row enters the earlier publication coverage only when at least one of its two China debt fields is non-null. That leaves 1,927 published records and excludes 5,840 rows where both fields are blank. A blank is never replaced with zero. For each field, the refreshed aggregate includes only the same borrower-years for which that earlier field is published.

Borrowers are joined first by an exact, case-normalized RecipientCountry label, then by the same explicit name fixes used for the earlier HRT loan records. There is no fuzzy matching. The overlap contains 1,909 borrower-years for 107 mapped borrowers. The reconciliation reports 0 unmapped observations and 18 mapped observations outside the refreshed panel, rather than silently dropping either category.

The reporting chart keeps the 2021 replication and November 2025 refresh as separately labelled levels over their 2000–2017 overlap. Their distance is not called a revision, correction, or hidden-debt gap, and the lines are never spliced into a combined series.

How the estimate is built

The authors start from loan-level records in AidData’s Global Chinese Development Finance Dataset, which documents individual Chinese lending commitments. Those are commitments, not outstanding balances, so the authors impute the missing maturities and grace periods by creditor entity and instrument type, then run each loan forward through a disbursement and amortisation profile to produce an estimated stock outstanding in each year. Every figure on these pages inherits that imputation.

The estimated stock decomposes into public and publicly guaranteed debt, private non-guaranteed debt, and drawn People’s Bank of China swap-line balances. The first two sum to the external stock, and all three sum to the total. That additivity was verified against the panel here rather than taken on trust: the largest absolute discrepancy across all 2,772 rows is 0.000011 USD bn.

What it excludes

Stated by the source itself: short-term trade debt and portfolio debt are out of scope, and the stock does not account for missed principal payments, because arrears are not systematically available at the loan level. A borrower that has stopped paying still carries its full imputed amortisation schedule here.

Units

Three scales meet on these pages and mixing them would be a fabrication, so they are converted once, at the query. The November 2025 China columns are already in billions of current US dollars. The earlier replication’s two China debt fields are in units of nominal US dollars according to its estimation code, and are divided by a billion for display. Reserves and reported-debt columns also arrive in units of current US dollars and are divided by a billion before they are used.

Zeros, blanks, and what they mean

A zero in a China column is a zero. The panel is complete: no debt column has a single missing value. A zero means the estimate found no loans outstanding, either because none were made or because the stock amortised out. It is rendered 0, not “n/a”. Only 15 of the 126 borrowers ever carry a drawn swap balance, so the swap column is a reported zero in every year for the other 111, and legitimately so.
A blank is not a zero. GDP is missing for some country-years, and where it is missing the source correctly leaves the percent-of-GDP columns empty. Those render “n/a” and are excluded from percent-of-GDP rankings; they are never filled with zero, and a country is never shown at the bottom of a ranking because its denominator was unknown.
The earlier workbook has a null grid, not early China debt observations. Both China debt fields are blank in 5,840 of its 7,767 country-year rows, including the pre-2000 grid. Those rows do not enter observed borrower counts or aggregates. Published numeric zeros within the observed window remain zeros.
A missing denominator renders n/a. 24 of the 126 borrowers have no row in the reported-debt table at all, and of the 102 that do have one in 2021, 24 leave private non-guaranteed debt blank while only 1 reports a genuine zero. Where the reported stock is unknown, the share of it owed to China is unknown, and the page says so rather than printing a zero or a hundred percent.

Why nothing here is subtracted

The reported-debt table carries 24 indicator codes, and none of them splits a borrower’s debt by creditor country. There is therefore no reported “debt owed to China” to subtract an estimate from, and any dollar figure presented as a hidden debt gap would be invented. The two sources bound each other and nothing more: an estimate of what one creditor is owed cannot exceed a borrower’s reported total to all creditors, and where it does, the two records disagree. Which one is wrong is not something this data can settle. The full argument, and the countries where the bound breaks, are on the two-sources page.

The same logic keeps drawn swap balances out of every comparison with reported debt. A central-bank swap drawing is not a category the Debtor Reporting System’s long-term debt concepts carry, so the ratios use the estimated external stock and never the total.

Restructuring deals are shown on country pages as narrative context. The source records face-value reduction as a 0/1 flag, not as a percentage, so no haircut magnitude exists for these deals and none is displayed. Nothing is ever netted off a stock.

Data quirks carried through

The restructurings file had two country-code typos in the source, corrected in the build step and confirmed against the paired country names: Zimbabwe and Serbia.
The Republic of the Congo and the Democratic Republic of the Congo are separate borrowers with separate codes, and are not merged.
The panel ends in 2021. The reported-debt and reserves series in this estate run later. Every cross-source join derives the comparison year from the China panel rather than hard-coding one, so a later reported stock is never silently compared against an earlier estimate.

License

The earlier HRT Mendeley replication archive is licensed CC BY 4.0 and is cited to the 2021 paper. The November 2025 workbook and restructuring file carry no machine-readable license. They are published for open download with no paywall or registration under the standard academic norm: free for research use with citation. Those files are treated as citation-required rather than public domain, and are not redistributed from this site.