# K6 control – etemadi2018 (NIH-AARP + PLCO + AHS, Int J Cancer 2018, accepted manuscript)

Controller: K6 · 2026-10-30 · Source: `post2015/etemadi2018.txt` (manuscript pages "x of 17").

## Method
- Table 3 (ms p. 16) was parsed by script. All 35 continuous rows were compared on:
  - point, CI and p_value (from the "\*" = p<0.001 footnote);
  - n_cases against the Table 2 totals;
  - the subsite-trend p in notes.
- Linear- and log-scale CI symmetry was computed, along with the z-value implied by the CI (as if it were HR ± 1.96·SE).
- The 4 text HvL rows were checked by reading (ms pp. 7–8).

## Results of the number check
- **35/35 Table 3 rows match exactly** on point, CI, \* → p_value, n_cases, and the subsite-trend p in notes (0.04, 0.6, 0.05, 0.04, 0.05, 0.7, 0.8).
- **All 4 text rows match**, with correct pages:
  - heme 1.14 (1.05–1.24), 1.20 (1.02–1.42), 1.27 (1.07–1.52), ms p. 7;
  - nitrate 1.18 (1.08–1.28), ms pp. 7–8.
- **0 mismatches.**
- **n_cases:** 6640 / 4997 / 3060 / 1807 / 1643 match Table 2. Person-years 4,605,311 is the pooled total, not outcome-specific, as noted.

## Linear-scale CIs (the specific check requested)
- **The CIs are symmetric on the linear scale**, mostly within ±0.01. Two examples:
  - processed red meat, colorectal: 1.30 (1.10, 1.49), ±0.20/0.19;
  - poultry, proximal colon: 0.73 (0.37, 1.08).
- **19 rows have log-asymmetry above 15 %:** _006–010, _013–015, _020–030. This is exactly the 19 the extractor reported.
- The pattern fits HR ± 1.96·SE(HR) after random-effects pooling, not exp(β ± 1.96·SE).
- Total processed meat (_031–035) is the exception: it is log-symmetric, which suggests it was computed differently.
- **Consequence for pooling:**
  - A log-scale SE must not be derived from the printed limits naively.
  - Use the SE on the HR scale: (hi − lo)/3.92, then delta method, SE(lnHR) ≈ SE(HR)/HR.
  - Otherwise, use the average of the two log half-widths.
  - Whichever is chosen should be documented.

## Discrepancies

| estimate_id | field | in database | in paper | severity |
|---|---|---|---|---|
| etemadi2018_010 | notes / p_value | p_value `<0.001`; not flagged | Printed with \* (p<0.001), but the CI 0.99–1.83 crosses 1. This is the same internal conflict the extractor flagged on _031/_032, but it is unflagged here | minor |
| etemadi2018_003, 004, 006, 013, 014, 021, 022, 024, 027, 028, 030 | p_value | `<0.001` | The asterisk is printed, but the implied z (from the CI on the HR scale) is 2.5–3.2, which is not consistent with p<0.001. The p_value is "as printed" but unreliable; a note would help | minor |

## Checked by reading
- **Exposure definitions (ms p. 5):**
  - Unprocessed red meat includes liver and mixed dishes → `red_unprocessed`.
  - Processed red meat = bacon, beef cold cuts, ham, hotdogs, sausage.
  - Total red = sum → `red_total`, includes_processed = yes.
  - White meat includes poultry cold cuts, low-fat poultry sausages and hotdogs, so includes_processed = yes is correct.
  - Total processed = red + white processed.
- **is_main_model.**
  - Total processed meat = yes and processed red meat = no. This is a documented judgment (one main row per `processed` × outcome).
  - For CUP-type "processed red" analyses the analyst may want to flip it. Not an error.
- **is_mutually_adjusted_processed = unclear** on the Table 3 rows. Correct: the model is substitution adjusted for *total meat*.
- **adjustments.csv.**
  - m1 (Table 3 footnote): sex, age, family history, ethnicity, NSAIDs, education, smoking, BMI, alcohol, fibre, calcium, energy and total meat. energy_method density. Correct.
  - m2 (Fig 2 legend): the same without total meat, plus cohort. Correct.
  - adj_physical_activity = no is correct (not listed).
- **dose_unit `per_50g_per_1000kcal`.** Correct. dose_conversions correctly declines to convert without mean energy.
- **studies.csv:**
  - n = 407,270; PY 4,605,311; median follow-up 13.8 (mean 11.3).
  - DHQ 124 items (AARP); adenocarcinomas only; ascertainment mixed (PLCO by medical record).
  - All correct.
  - NIH-AARP overlaps cross2010 and zouiouich2025. Noted.

## Missing estimates
1. **HCA highest-vs-lowest quintile rows, ms p. 8, Results text.** These are printed with CIs, and KODEBOK 4.1 explicitly lists `hca_phip`, `hca_meiqx` and `hca_dimeiqx` for "Cross 2010, Etemadi 2018". The extractor's NOTES call HCAs out of scope (R1), but the codebook vocabulary anticipates them, and cross2010 was coded with HCAs.
   - MeIQx, colorectal: 1.21 (1.12–1.31)
   - DiMeIQx, colorectal: 1.16 (1.08–1.24)
   - PhIP, colorectal: 1.09 (1.00–1.18)
   - PhIP, all colon: 1.09 (1.00–1.20)
2. **B(a)P, colorectal: 1.00 (0.93–1.08).** There is no vocabulary value for it, so it is optional.
3. **Not codable** because no numbers are printed:
   - nitrite ("weaker than nitrate");
   - per-cohort estimates (Suppl Table 1);
   - quintile series (Suppl Tables 2–3; Fig 1/2 are plots only);
   - lag and residual-energy sensitivity analyses ("data not shown").

## Confirmed rows
All 39 rows are confirmed on the numbers. The only issues are the p_value annotations above.

## Internal inconsistencies in the paper
- **Table 3 "\* p<0.001"** is attached to estimates whose CIs touch or cross 1:
  - _010: 0.99–1.83;
  - _031: 1.00–1.30;
  - _032: 1.00–1.36;
  - plus about 10 more where the CI implies p ≈ 0.001–0.01.

  The asterisk appears to be applied loosely, or it refers to something else.
- **The CIs are on the linear scale**, which is unusual for HRs (see above).
- **Mean follow-up 11.3 y is less than the median of 13.8 y.** This is possible with a left-skewed distribution, but worth noting.
- **Table 2:** colon + rectum (4997 + 1643 = 6640) equals the CRC total, while proximal + distal (3060 + 1807 = 4867) is less than colon (4997). This is consistent with the footnote "a few overlapping or unknown tumor locations".

## Verdict
**Fit for use, with care.** Every number matches.
- Users must handle the linear-scale CIs when deriving SEs.
- The asterisk-derived p_values should not be relied on.
- Consider adding the 4 HCA rows, which the vocabulary already supports.