Decomposing U.S. Crude Death Rates, 2010-2024: Population Aging Dominates, and the Denominator Is Not Stable
A reproducible analysis of United States mortality, 2010-2024, making two contributions. The decomposition. The change in the crude death rate is split into an age-specific mortality component and a population age-structure component using the Kitagawa method. Over the full interval the age-structure component exceeds the entire observed rise in the crude rate, while age-specific mortality improved: the death rate went up over a period in which dying at a given age became less likely. Pandemic-era excess mortality is estimated against a baseline projected from the pre-pandemic trend in the age-adjusted rate rather than the raw count, and the age distribution of COVID-19 deaths is characterised. The denominator. Four vintage boundaries falling inside a single fifteen-year series are measured rather than assumed: the per-year vintage chain CDC WONDER carries, the Vintage 2024 restatement of 2023, the 2010 April 1 measurement basis, and the bridged-race to single-race seam at 2017/2018. The seam is exactly zero in every age band, in deaths and population alike, measured against a purpose-run export. The restatement is not: it accounts for roughly a quarter of the published 2023-to-2024 crude-rate decline, most of which books as a spurious improvement in age-specific mortality. NCHS has itself published two different crude death rates for 2020, in two of its own reports, naming denominator rebasing as the cause — which turns the argument that a crude rate is a claim about a denominator into a documented instance. Inputs are four committed CDC WONDER exports and three U.S. Census population vintages. Each export carries WONDER's own query-parameter footer and a SHA-256 that the test suite recomputes, so a reviewer receives the exact bytes the results were computed from rather than a description of a query to reconstruct. Every value in the manuscript is generated by the analysis code and substituted into a template; no number is typed into the prose. Three claims are kept separate and are not interchangeable. Provenance records where a value came from. Attestation records that a person checked it against that source, and no automated path in this repository can write it. Corroboration records that a separate publication reports the same figure, and is deliberately partial: 14 of 15 annual totals are corroborated against NCHS's published NVSR reports, 2023 on the death count only, and 2024 has no published source at all. A blank corroboration field means not corroborated, never that corroboration failed. Stated limits. The NVSR corroboration is not independent confirmation: NVSR and WONDER are both NCHS products drawing on the same mortality file and the same Census-derived denominators, so agreement shows the query returned what NCHS published, not that NCHS is correct. Age groups are collapsed to six rather than the NCHS eleven, so age-adjusted rates here are not comparable to NCHS's published figures. Licensing is split, because one licence covering all three components would have to be wrong about at least one. Software (src/, tests/, bootstrap scripts, notebooks) is BSD-3-Clause. The manuscript, figures and results.json are CC BY 4.0. The raw federal data is not licensed, because it is not the author's to license: U.S. federal government works are in the public domain under 17 U.S.C. § 105. The record below lists BSD-3-Clause because this is a software deposit; see LICENSE, paper/LICENSE and DATA.md in the repository for the full statement. AI assistance. This work was produced with AI assistance (Anthropic's Claude, via Claude Code), including the analysis software, its tests and documentation, and drafting of the manuscript. The author directed the work, verified every data value against the CDC WONDER export it was taken from, and is solely responsible for the analysis, its interpretation and its conclusions. No AI system meets authorship criteria and none is listed as a creator. See the repository README for the full disclosure.