# K7 control – chan2005 (Chan et al., Int J Cancer 2005;115:648–652)

Full text: `drive/Understudier WCRF/chan2005.txt` (no PDF available; ≤/≥ symbols are lost in the text file, while `<` and `>` survive).
Rows checked: 17 (chan2005_001–017). Every number was compared with Table II (p. 650) and the Results text (p. 649) and recomputed by script where possible.

## Discrepancies

| estimate_id | field | in database | in paper | severity |
|---|---|---|---|---|
| – | – | – | – | none found |

All 17 rows reproduce the printed values exactly: OR, CI, cases/controls, model, page and table.

## The meat row that duplicates a smoking row (chan2005_001–004)

I confirm the extractor's flag independently:
- Table II, All women, ">0.5 serving/day" prints 17/30, age-adjusted 1.15 (0.82–1.63) and MV 1.21 (0.85–1.72). The "Any smoking before age 30" row prints the same two estimates digit for digit (99/226).
- Crude OR from the meat counts: (17/30)/(166/413) = **1.41**. Crude OR from the smoking counts: (99/226)/(84/217) = **1.13**. The printed 1.15 fits the smoking counts and does not fit the meat counts.
- The genotype strata are internally consistent. Slow: crude (7/19)/(100/248) = 0.91, printed 0.91. Rapid: crude (10/11)/(66/165) = 2.27, printed 2.40. So the error is limited to the All-women meat row.
- The counts sum correctly: 166+17 = 183 cases and 413+30 = 443 controls; slow 107/267 and rapid 76/176 match Table I.

**Verdict:** rows 001–004 are correctly transcribed as printed, but the HRs in 002 and 004 are almost certainly the smoking estimates. They must never be used. The setting is_main_model = no is correct.

## Definitions, models and flags
- Exposure "beef, pork or lamb as a main dish" → `red_main_dish`, includes_processed = `unclear`. This is correct (the paper does not say).
- Category bounds: the table prints "[?]0.5" for the reference and ">0.5" for the top category. The abstract and text contrast "0.5 serving/day" with "<0.5 serving/day". Because the symbols are lost, the boundary convention cannot be settled from the text file. The database's ≤0.5 / >0.5 follows the table and is acceptable as printed.
- Rows 013–015 (three-category analysis, text p. 649): 0.96 (0.66–1.42) and 1.42 (0.74–2.74) are confirmed. The reference (≤1 serving/week) is implied by the text. Labelling these rows "Multivariate" with m2 is an assumption: the text says "multivariate ORs" but does not restate the covariates. is_main_model = yes is defensible given the corrupted Table II all-women row.
- Rows 016 and 017 (joint effects): 5.85 (1.54–22.2) and 17.6 (2.0–158.3) are confirmed. The reference strata match the text and abstract (abstract: "<0.5 serving/day", ≤35 pack-years). Both CIs are log-symmetric. Coding them as `hvl` is a vocabulary stretch, but the notes document it and forbid pooling.
- adjustments m1 (age-only) and m2 (age, BMI, family history, HRT, endoscopy → screening, multivitamin, aspirin → NSAID, plus pack-years for meat analyses) match the Table II footnote. adj_energy = no is correct.
- studies: nested_cc, 626 participants (183 cases, 443 controls), NHS blood subcohort 1989–1990, ages 44.7–70.9 at blood draw and 96 % medical-record confirmation are all confirmed. diet_items = 61 comes from "among over 61 foods"; strictly that means more than 61, which is a trivial point.

## Missing estimates
None. Table II's smoking and genotype rows are not meat exposures. No colon/rectum split is reported.

## Confirmed rows
All 17 rows are confirmed as transcribed. Rows 002 and 004 are confirmed as printed but are a paper error (see above).

## Internal inconsistencies in the paper
1. The Table II All-women meat row duplicates the smoking-before-30 row (above).
2. The Table II footnote describes the meat adjustment variable as "0.5 servings/**week** and >0.5 servings/week". Everywhere else the unit is per **day**.
3. The table gives the top category as ">0.5"; the text and abstract give it as "0.5 … compared to <0.5".