# K5 control – islam2019 (pooled analysis of 6 Japanese cohorts)

Controller: K5 · 2026-10-30 · Full text: `Understudier WCRF/oppfolginger/islam2019.txt` (main text only; Supplementary Tables S1–S4 not in file)
Method: Tables 2–5 parsed by script (394 table cells: every category HR and CI, every trend P, and the continuous column in Tables 4–5). Every value was compared with estimates.csv. I also checked that each row's `source_table` matches its pool and sex (Table 2 = frequency/men, 3 = frequency/women, 4 = quantity/men, 5 = quantity/women). The 7 text rows were read by hand.

Rows checked: 401.

## Which pool is which (verified)
| Pool | Tables | Cohorts (Table 1) | Participants | CRC cases | Exposures |
|---|---|---|---|---|---|
| **Frequency** | 2 (men), 3 (women) | JPHC I, JPHC II, JACC, MIYAGI, OHSAKI | 107,777 + 124,626 = 232,403 | 3,428 + 2,266 = 5,694 | beef, pork, processed red meat, chicken; <1/wk, 1–2/wk, 3–4/wk, almost daily (3–4/wk+ combined for proximal/distal) |
| **Quantity** | 4 (men), 5 (women) | JPHC‐5y, TAKAYAMA | 57,672 + 65,963 = 123,635 | 2,108 + 1,442 = 3,550 | unprocessed red meat (beef+pork), beef, pork, processed red meat, chicken; cohort-specific quartiles + per 100 g (per 50 g processed) |

The database gets this right throughout:
- All Table 2/3 rows have frequency units (`times_week`) and notes naming the frequency pool.
- All Table 4/5 rows have quartiles, per_100g/per_50g and notes naming the quantity pool.
- No category set mixes tables.
- `red_unprocessed` occurs only in the quantity pool, as in the paper.

## Discrepancies

| estimate_id | field | in database | in paper | severity |
|---|---|---|---|---|
| studies.csv islam2019 | n_participants / cases_colorectal / notes | 356,038 / 9,244, described as "two separate pools" | Table 1: JPHC‐5y population 140,420 = JPHC I 61,595 + JPHC II 78,825. JPHC‐5y is the 5-year survey of the same JPHC I+II residents (10 vs 5+6 PHC areas). The pools therefore **share participants**. 356,038 and 9,244 double-count JPHC people and cases (the paper itself prints these sums in the Discussion). | **major** (not a wrong number, but the overlap is undocumented; frequency- and quantity-pool rows must never be combined, and the JPHC overlap with takachi2011 applies to BOTH pools) |
| islam2019_m1–m4 (adjustments) | adj_stratified_center | yes | Area is a covariate ("area‐adjusted"), not a stratum. Methods: area applies within JPHC‐I, JPHC‐II and JACC only, so for m3/m4 (JPHC‐5y, TAKAYAMA) it probably applies at most to JPHC‐5y. Other extractors coded area-as-covariate = no (e.g. the study at NOTES_all.md l.928). | minor (cross-study flag inconsistency) |
| islam2019_395 | sex | women | Sentence follows "In women (Table 3)"; sex not stated explicitly. Inference is reasonable: Table 3 multivariable counterpart is 1.31 (0.90‐1.91), men's 1.15. | minor (inferred; note says so) |
| all 401 | source_page | "page not marked in txt" | – | minor |

No critical discrepancies. All 394 table values (point, CI, p_trend) match exactly.

## Specific checks
- **Printed anomalies, recorded as printed and flagged:**
  - islam2019_261 (Table 4, processed, distal colon, men, Q2) prints "0.58 (0.70‐1.04)", a point estimate outside its own CI.
  - islam2019_084 "1.33 (1.001‐1.76)" was split across two lines in the txt and is correctly re-joined.
  - islam2019_354 (beef, proximal colon, women, per 100 g): "0.54 (0.03‐11.29)" as printed.
- **is_main_model:** quantity pool = yes, frequency pool = no.
  - This is a defensible tie-break, but the paper does not rank the pools. Neither one is "less adjusted": both use the same covariate set.
  - For beef, pork, processed and chicken, analysts should treat the frequency-pool rows as an alternative main estimate from a different (partly overlapping) set of cohorts, not as a secondary model.
- **Text rows:**
  - 396: 1.42 (1.04‐1.96), P .18.
  - 397: chicken 1.04 (0.66‐1.63), P .03. The inconsistency flag is correct: the text calls it "decreased risk".
  - 398: 0.71 (0.55‐0.91).
  - 399: 1.29 (1.03‐1.61), P .04.
  - 400/401: excl. TAKAYAMA, 1.11 (0.93‐1.33) and 1.23 (1.00‐1.52), women. `cohort_id = jphc` is correct, because only JPHC‐5y remains.
  - 395: age/area model, 1.48 (1.03‐2.11).

  All verified.
- **Adjustment flags m1–m4** match the footnotes: age, diabetes, BMI, smoking, alcohol, PA, log energy (covariate), calcium and fiber quartiles. No education.
- **Reference rows** = 1.00 with empty CI in all 98 sets.
- **Case sums:** not checkable. The paper prints no cases per category; `n_cases` is empty throughout, which is correct.

## Missing estimates
None in the main text. Every Results-text value is either a table cell or one of the 7 coded text rows. Supplementary Tables S1–S4 (subsite results excluding early cases, etc.) are not available. The database states this, but they remain an open gap.

## Confirmed rows
All 401 rows confirmed numerically (islam2019_001–401). Only the flag and label notes above apply.

## Internal inconsistencies in the paper
1. **islam2019_261:** "0.58 (0.70‐1.04)". The point estimate lies outside its CI (probably 0.85 or 0.88).
2. **Chicken, CRC, women, excl. 3 y:** the text says "decreased risk" but the HR printed is 1.04 (0.66‐1.63).
3. **Double-counted totals:** the abstract and Discussion add the two pools (356,038 participants, 9,244 cases) although JPHC I+II and JPHC‐5y are the same population.
4. **Takayama differs from Wada 2017.** Table 1 gives 14,224 men / 16,575 women and 449 / 352 CRC; Wada 2017 gives 13,957 / 16,374 and 429 / 343 for the same end date. The extractor notes this.