In 1982–84 a Taiwan steel mill recycled orphaned cobalt-60 sources into reinforcing rebar that ended up in ~180–200 Taipei-area buildings. Roughly 6,000–10,000 residents were chronically irradiated for years before the contamination was discovered. The resulting cohort has been read as proof of two opposite things: that low-dose radiation is harmless (even beneficial), and that it measurably causes cancer. This project asks a narrower, cleaner question — what does a causal-inference reading of the public data actually license us to conclude?

1. The data, and what is actually public

There is no public individual-level dataset for this cohort. The microdata is held under restricted access by Taiwanese institutions (Taipei Medical University / National Yang-Ming; the Taiwan Cancer Registry; the Atomic Energy Council for dosimetry) and released only on-site through Taiwan’s Health and Welfare Data Science Center to IRB-approved, Taiwan-based researchers.

What is public are the aggregate contingency tables in the peer-reviewed papers. This reanalysis digitizes those tables (with provenance headers, validated against reported totals) and fits causal models to them:

Note: PMID 9600303 (Tung 1998), sometimes cited as “the cohort study,” is the dose-reconstruction methodology paper, not the cancer-outcome analysis.

2. Methodology

A. Encode the data-generating process as a DAG

The core of a Pearl-style analysis is being explicit about the causal structure. Dose was effectively assigned at the building level (everyone in a unit shared the gamma field). Socioeconomic status is an unmeasured common cause of who ended up in a new 1982 Taipei apartment and of baseline cancer risk. Crucially, the external-SIR design conditions on a selection node — membership in this particular, non-random population.

        SES (unmeasured) ───────────────┐
          │           │                 ▼
          ▼           ▼               Cancer ◄── AgeAtExposure, AttainedAge,
       Building    Selection            ▲          Sex, CalendarPeriod
          │       (SIR design)          │
          ▼                             │   Smoking, ReproHistory (unmeasured)
       TrueDose ──────────────────────►─┘        │
          │                                       ▼
          ▼                                     Cancer
     AssignedDose ◄── DoseError (reconstruction)

B. Check identifiability — honestly

A self-contained d-separation routine confirms the causal effect is not identified by the covariates available in the published data (age at exposure, attained age, sex, calendar period). The backdoor paths through SES, smoking, and reproductive history stay open. Identification therefore rests on the authors’ implicit quasi-randomization assumption — that because residents were unaware of the contamination, their dose is independent of lifestyle given age. We treat that assumption as a first-class object to be probed, not a footnote.

C. Estimate, reproduce, and stress-test

• Reproduce the published external SIR from counts (validation that the digitized data and machinery agree with the source).
• Fit the internal dose–response — the radiation-standard linear excess-relative-risk model and a log-linear hazard model — adjusting for age at exposure where the tables permit.
• Bias analysis: E-values for unmeasured confounding; a classical-vs-Berkson measurement-error simulation; sensitivity to the assumed within-group dose; and false-discovery-rate control for multiple comparisons.

3. Results

The two headline findings are both real — and not in conflict

A health-selected cohort with a compressed age distribution can simultaneously have fewer total cancers than the general population and a real internal gradient in which higher-dose members get more cancer than lower-dose members of the same cohort. These answer different causal questions. The external comparison conditions on a non-exchangeable reference; the internal comparison does not.

Why the deficit looks “protective” — age structure, then selection

The widely-quoted “~35% lower” is the crude, all-ages comparison: 247 cancers across 97,106 person-years (254 per 100,000) against Taiwan’s national 390 per 100,000 → SIR 0.65. That national rate is inflated by an elderly tail this cohort barely has, so age-standardizing to the cohort’s own age distribution — what Doss does — moves it to SIR 0.84. Roughly half to two-thirds of the crude gap is age structure — but a deficit of ~16% (Doss 0.84), up to ~25% (Hwang 2008, 0.75), survives standardization.

That surviving deficit is not protection, and not the cohort simply “being young” — by 2012 its attained age (~45) is a few years older than Taiwan’s mean. It is healthy-cohort selection: the people who moved into new 1982 Taipei apartments were not a random draw of Taiwan. (For scale, the within-cohort rate gradient across age at exposure is ~14×, versus only ~2.6× across dose — which is why a crude comparison misleads so badly.)

Is the internal signal robust?

4. In response to Southwood & Chalmers

Southwood and Chalmers argue that the case for low-dose radiation harm is weak, and use this cohort as a centrepiece: the affected group has fewer cancers overall, so the site-specific “excesses” look like statistical noise. This reanalysis agrees with much of their critique but parts ways on the load-bearing inference.

First, what is not in dispute. Nobody here is arguing about low dose radiation — a CT scan’s worth of exposure is roughly fine, which is exactly why we use CT so freely. The live question, as Southwood puts it, is chronic low-dose-rate exposure that nonetheless accumulates to a high total. That is precisely what this cohort is: years of continuous exposure, with a high-dose tail reaching ~2.4 Sv cumulative. So the external deficit is beside the point — the evidence that actually bears on the question is the internal dose–response, and that is what the rest of this section is about.

Where we agree

• The deficit is real, and Chen & Luan’s hormesis mortality claim is the weakest reading of it (mortality endpoint, young cohort, short follow-up).
• Hwang 2006’s site-specific SIRs are multiple-comparison noise. Their point — 77 subtypes screened, every “significant” site resting on ≤7 cases, ~4 false positives expected by chance — is exactly what our FDR step reproduces.
• Pre-registration concerns are legitimate, and the magnitude of any effect is soft.

Where the reasoning breaks — the crux

The article treats the overall deficit as evidence against causation and the internal dose–response as p-hacking that “contradicts” it. Under confounding, that is a non-sequitur — and two numbers show why:

The internal comparison is exactly the estimand the external deficit cannot speak to, because high- and low-dose members share the selection that drives the deficit. So the deficit is uninformative about the dose–response, rather than a rebuttal to it.

Where the article has force — and where it is too dismissive

In fairness, our own analysis says the internal signal is fragile: modest E-values mean a moderate unmeasured confounder (e.g. reproductive history, unmeasured here) could weaken the breast-cancer result. So “unconvincing” is not unreasonable. But the article lumps together two very different things: Hwang’s subtype dredge, and Hsieh’s pre-specified continuous-dose model — especially the young-breast finding (HR 1.38, p = 0.0008), which is a directed hypothesis with strong external corroboration in the atomic-bomb survivor data. That external consistency is the article’s blind spot.

5. Verdict

The reanalysis lands close to Southwood & Chalmers on the practical temperature — harm at these doses is small and hard to pin down — but rejects the route they take to get there. The overall cancer deficit, once you concede (as we do) that a ~16–25% deficit survives age-standardization, is healthy-cohort selection, and a selection-driven external deficit is structurally incapable of testing a within-cohort dose–response.

The one comparison that can test the low-dose-rate/high-cumulative-dose question — the internal gradient — leans against “fine”: positive, roughly linear, no visible threshold or hormesis. But it is fragile — modest E-values (1.2–1.6), key confounders (smoking, reproductive history, cluster SES) unmeasured, and no clean point-identification from aggregate public data. So the honest bottom line is not “radiation caused these cancers” and not “probably fine”: it is that the best-identified estimate available leans toward harm, the external deficit does nothing to offset it, and the question cannot be settled either way without the individual-level HWDC microdata. Their conclusion might still be right — but the evidence that could show it currently points the other way, and getting a real answer needs the data.

6. Reproduce it

Everything above is generated by a runnable pipeline: ./run.sh digitizes the tables, runs the QC gate, builds the DAG, reproduces the published numbers, fits the causal models, runs the bias analyses, and writes the synthesis. Code, data, and methods:

DAG + d-separation linear-ERR & log-linear fits E-values measurement-error simulation FDR

→ github.com/riemannzeta/co60-causal-reanalysis