# K5 control – wada2017 (Takayama Study, Cancer Sci 2017;108:1065–1070)

Controller: K5 · 2026-10-30 · Full text: `Understudier WCRF/oppfolginger/wada2017.txt`
Method: Tables 2 and 3 parsed by script (every row: median, No. of subjects, person-years, cases, HR and CI for colorectal, colon and rectum, plus trend P). I compared every value with estimates.csv. Definitions, models, flags and the seafood text rows were checked by reading.

Rows checked: 152 (144 from tables + 8 seafood rows from the text).

## Discrepancies

| estimate_id | field | in database | in paper | severity |
|---|---|---|---|---|
| wada2017_145–152 | model_label / is_main_model | "Multivariable (as Tables 2/3, assumed)", main = yes | Text says only "evaluated similarly" (Results, last paragraph). The model is not restated. | minor (assumption is reasonable and labelled as one) |
| all 152 | source_page | "1065-1070 (page not marked in txt)" | No page markers in txt | minor (unavoidable; check against PDF) |

No critical or major discrepancies.

## Specific checks requested
- **Person-years per quartile:** all 72 person-year cells (3 exposures × 2 sexes × 2 analyses × 4 quartiles; each repeated on colorectal, colon and rectum rows) match Tables 2/3 exactly. Totals: men 182,985 and women 226,681, identical for all three exposures.
- **No. of subjects (`n_in_category`)** matches for all rows, including the excl-2y analysis (3481/3480…).
- **Case sums:** colorectal 429 men / 343 women, colon 276 / 244, rectum 153 / 99, for every exposure. The excl-2y sets sum to 393 / 316, i.e. 63 cases removed, as the text says. These sums agree with studies.csv (772 / 520 / 252).
- **Reference rows:** Q1 = 1.00 with empty CI everywhere.
- **Exposure coding:**
  - Red meat = beef and pork, incl. grilled offal and liver. Ham, sausage, bacon and yakibuta are defined as processed, so `red_unprocessed` / `exposure_includes_processed = no` is correct.
  - Total meat includes processed, coded `yes`. Correct.
  - Seafood = fish and shellfish, coded `fish`. Correct.
- **Adjustment sets m1 (men) and m2 (women)** match the Table 2/3 footnotes. Flags are correct: NSAID = aspirin; energy is handled by the residual method on the exposure (`energy_method = residual`); menopausal status for women only.
- **Dose:** `dose_point` = printed quartile medians (g/day, energy-adjusted). dose_conversions c01–c36 agree.
- **studies.csv:**
  - n = 30,331 (13,957 + 16,374). Correct.
  - Follow-up 15.5 y, `max` (Sept 1992 to March 2008). Correct.
  - 169-item FFQ. Correct.
  - `mixed` ascertainment (registries + colonoscopy histology). Correct.
  - `includes_processed = no`. Correct.

## Missing estimates
None. The Discussion's pooled meta-analysis of six Japanese cohorts (e.g. 1.22 for red meat) is not a Takayama estimate and is correctly not coded. Iron and fat results are "data not shown".

## Confirmed rows
All 152 rows confirmed (wada2017_001–152), apart from the minor labelling notes above.

## Internal inconsistencies in the paper
1. **Person-years in the excl-2y analysis are larger than in the full analysis in 11 cells**, although subjects and cases are fewer. Examples:
   - Total meat men Q1: 46,272 vs 46,270.
   - Processed men Q1: 46,237 vs 46,197.
   - Total meat women Q2: 55,452 vs 55,387.

   Removing early cases should not add person-time. The database records the values as printed, which is correct. Analysts should not rely on the excl-2y person-years being exact.
2. **Covariate form conflicts between Methods and table footnotes.** Methods: fiber, calcium and vitamin D entered as energy-adjusted g/day. Table footnotes: "intakes (quartiles)". This is already noted in adjustments.csv.