# Data choices — heat-deaths story

**Why this story changed.** The DataPressr draft "Heat Is the Quiet Killer" led on the National Weather Service ranking (heat first, 5,366 deaths in 1997–2025) and brought in death certificates only as a 2006–2010 aside. Two reviewers noted that on death certificates cold outnumbers heat, which undercuts the title. Querying CDC WONDER directly for every year confirmed it: cold appears on more certificates than heat in all 26 final years, 1999–2024. So the central claim is now the reversal: heat leads the weather service's count, cold leads death certificates, and the gap is narrowing.

**Searched.** NWS national summaries (via the DataPressr dataset) and NWS per-state heat summaries (via the draft's snapshot); CDC WONDER Multiple Cause of Death, final (D77 1999–2020, D157 2018–2024) and provisional (D176); the published death-certificate studies (Berko et al. 2014 for 2006–2010; Howard et al., *JAMA* 2024, heat 1999–2023; Liu et al., *JAMA* 2025, cold 1999–2022); EPA's heat- and cold-related deaths indicators (not used: they repackage the same CDC data with a lag).

**Chose WONDER over quoting the studies.** One code set, one source, every year, both hazards, to 2024. It reproduces Liu et al.'s 40,079 cold deaths for 1999–2022 exactly. For heat 1999–2023 it gives 21,557 against Howard et al.'s 21,518 (the DataPressr draft outline records their code set as X30, T67 and P81, a newborn code; an earlier file vintage may also differ). Berko reports 3,332 heat and 6,660 cold deaths; this extract gives 3,408 and 6,733. Its cold definition omits T69; the remaining differences are unresolved.

**Definitions used.**

- *NWS heat* is the summary's `Heat` row and *NWS cold* its `Cold` row; the summaries do not say how they map Storm Data's Heat/Excessive Heat and Cold/Wind Chill/Extreme Cold event types onto these rows. Winter storm, winter weather and ice are separate rows; adding winter storm and ice to cold (1,573) is reported once as a robustness check.
- *Certificate heat:* X30 or T67 anywhere on the certificate. *Certificate cold:* X31, T68 or T69 anywhere. T67/T68 are contributing-cause codes, not underlying-cause codes. Certificates carrying W92/W93 (excessive heat/cold of man-made origin) are not excluded, so these are not exclusively weather deaths; X30 and X31 are natural only. The heat and cold groups are not mutually exclusive and should not be added together. Their union was deliberately not queried: subtracting it would reconstruct counts of 1–9, which WONDER suppresses.
- *Final years:* 1999–2024. 2025 is excluded from matched-source comparisons but included in NWS ranking totals (it is preliminary there). Provisional 2025 certificate counts are shown as open circles on the chart only.
- *Comparison window:* 1999–2024, the years both sources have final data. The NWS ranking itself uses all 29 NWS years, 1997–2025, as the dataset does.

**Definitional changes checked.**

- *Death certificates:* ICD-10 applies to deaths from 1999, so the codes are stable across the series (earlier ICD-9 years are not used). The 2018 move to single-race categories (D157) does not affect cause coding; D77 and D157 agree exactly for 2018–2020.
- *NWS:* the counting rule is NWS Instruction 10-1605. The current version (26 July 2021) supersedes the 16 July 2018 version; its summary of revisions (links, wording, damage sources, flood, debris-flow, dust-storm, hail and snow-squall guidance) touches neither heat nor cold. Versions before 2018 were not checked, so a change in how heat or cold deaths were entered between 1997 and 2018 cannot be ruled out. The heat rule (an entry generally needs heat-advisory or warning conditions in that office's own criteria) means coverage varies by office and year. The national summary's event vocabulary is stable for heat and cold over 1997–2025 (the DataPressr dataset README lists the changes that do occur: thunderstorm wind, hurricane spelling, rip current from 2002, small stream/urban flood to 2006).
- *Arizona:* Arizona's NWS heat deaths jump from at most 9 a year (except 51 in 2005) to 143 in 2019. No source found documents a rule change; the 2021 directive post-dates it. In the 2020 heat summary, 299 of Arizona's 337 deaths have location "Other" and unknown age and sex, which looks like a bulk entry from a county tally, but this is not documented and the post does not claim it.

**Not used.** Death rates for heat (the counts are compared within one year, and the rise from 2015–2019 to 2020–2024 is far beyond population growth; the post quotes Liu et al.'s age-adjusted cold rate instead). State-level certificate counts (the WONDER API serves national data only). Weinberger and Gasparrini provide model-based context; neither estimate is included in the certificate or NWS counts.

**Size.** Three CSVs, about 60 KB.