# K3 control: wei2004 (Wei et al. 2004, Int J Cancer 108:433-442)

Controller: K3 (independent). Full text: `Understudier WCRF/wei2004.txt`. Date: 2026-10-30.
Method: the main-dish and processed-meat blocks of Table II (p. 439) were parsed by script and compared with all 60 category rows (cases, RR, CI, p trend, cohort/sex mapping). Rows 061-062 (Table III), definitions, units, models and flags were checked by reading.

## Summary
- Rows: 62. Checked: 62. Numbers confirmed: 62/62. The script found 0 mismatches on the 60 Table II rows.
- Discrepancies: critical 0, major 1 (the dose unit on 061/062), minor 1.
- Missing estimates: none.

## Discrepancies

| estimate_id | field | in database | in paper (table/page) | severity |
|---|---|---|---|---|
| wei2004_061, wei2004_062 | dose_unit | per_serving_week (as printed in the Table III row label) | The Methods say meat was "modeled in servings per day" (p. 435). The Table II header says "(servings/day)". The categorical data fit per day far better than per week (see below) | major (dose). The printed number is correct, but the unit is probably wrong. Set verification_status = disputed. Do not use these rows for dose-response until resolved |
| cup2016_wei2004_colon / _rectum | estimate_ids_used | empty | CUP cases (1137 colon / 338 rectum) equal the combined-cohort main-dish category sums, so CUP evidently used rows 021-025 / 026-030 | minor |

### The unit question in Table III (per serving per week vs per day)
Table III labels both meat rows "one additional serving per week". The evidence that the unit is actually per serving per day:
1. Methods p. 435: "Red meat and processed meat were modeled in servings per day". The polytomous model entered "the medians of the categories" as a continuous variable.
2. The Table II header and the Table I means are both in servings/day (e.g. 0.37 servings/day for women).
3. The Table III values do not fit the categorical results if they are per week, and they do fit if they are per day:
   - Main dish, combined colon: the categories run 1.00, 1.43, 1.29, 1.37, 1.43 (trend p = 0.25). At 1.12 per serving per week, about 6/week vs 0 would predict roughly 1.12^6 ≈ 1.97, and 2-4/week vs 1/week roughly 1.25 (observed 1.06). At 1.12 per serving per day, 6/week (0.86/day) predicts 1.10.
   - Processed meat, combined colon: the categories run 1.00, 1.02, 1.02, 1.13, 1.33. Per week, 1.20^6 ≈ 3.0. Per day, 1.20^0.86 ≈ 1.17.
   - The per-week reading implies a trend steeper than any category gradient in Table II and a trend p (≈0.07 for the main dish) lower than Table II's 0.25/0.55.

Conclusion: the "per week" label is most likely a labelling error, and the true unit is per serving/day. The extractor recorded the label as printed and wrote a UNIT CONFLICT note, which is correct practice. Because dose_unit drives the conversion to per 100 g, the field should be marked disputed, and any use of 061/062 should be run under both readings. The paper also warns that the polytomous estimates "cannot be compared directly" with the pooled logistic models (p. 435). That point is already in the notes.

## Definitions, models and flags (checked by reading)
- **What the "main dish" item includes:** Wei never lists the items. The variable is only named "beef, pork or lamb as a main dish", with no footnote. `red_main_dish` with exposure_includes_processed = `unclear` is correct for this paper.
  - Cross-paper context, outside Wei's own text: Bostick 1994 says its FFQ is "virtually identical" to the 1984 NHS FFQ, and lists that item as "(e.g., steak, roast, ham, etc.)". So ham is probably included in the HPFS (1986) and later NHS versions of the item.
  - Wei's NHS baseline is the 1980 61-item FFQ, whose item wording is not given here. I would keep `unclear` and add this cross-reference to the notes. CUP's issue_category `includes_processed` is consistent with that reading.
- **cohort_id assignment:**
  - NHS rows (001-010, 031-040) = `nhs`, women.
  - HPFS rows (011-020, 041-050) = `hpfs`, men.
  - COMBINED COHORT rows (021-030, 051-060) and Table III (061-062) = `nhs_hpfs`, all. studies.cohort_id = nhs_hpfs.
  - All of this is correct. The combined data were aggregated individual-level data analysed by pooled logistic regression, with gender as a covariate; it is not a meta-analysis.
  - cohorts.csv correctly declares nhs_hpfs as overlapping nhs and hpfs. Any pool must use either the nhs_hpfs rows or the nhs + hpfs rows, never both. Bernstein 2015 rows on nhs/hpfs must also be kept out of any pool that contains Wei.
- **Models:**
  - m1 (single cohort) and m2 (combined, plus gender) match Table II fn 1: age, family history, BMI, physical activity, main dish, processed meat, alcohol, calcium, folate, height, pack-years before age 30 and endoscopy.
  - adj_energy = no is correct: energy is not a covariate, and meat servings are not energy-adjusted. Only nutrients are energy-adjusted (p. 434).
  - adj_screening = yes (endoscopy) is correct.
  - m3: adj_fiber = unclear is a good catch. Methods p. 435 lists fiber among the polytomous variables, but Table III does not.
- is_mutually_adjusted_processed = yes on the main-dish rows and no on the processed rows is correct.
- is_main_model = yes on the m1, m2 and m3 rows does not conflict, because sex (women/men/all) or outcome (colon, rectum, colorectal) differs.
- Doses and labels: 0 / <3 per month / 1 per week / 2 to 4 per week / 5 or more per week. The bounds are recorded correctly, and leaving the "<3 per month" bounds empty is explained in the notes. Source page 439 (Table II) and 441 (Table III) are correct.

## Case-count re-derivation
| set | sum of categories | paper total |
|---|---|---|
| Main dish, NHS colon / rectum | 670 / 203 | 672 / 204 |
| Main dish, HPFS colon / rectum | 467 / 135 | 467 / 135 |
| Main dish, combined colon / rectum | 1137 / 338 | 1139 / 339 |
| Processed, NHS colon / rectum | 668 / 202 | 672 / 204 |
| Processed, HPFS colon / rectum | 467 / 135 | 467 / 135 |
| Processed, combined colon / rectum | 1135 / 337 | 1139 / 339 |

The shortfalls are all in NHS (missing intake) and are correctly documented in the row notes. All reference categories are 1.00 with an empty CI. Every point lies inside its CI.

## studies.csv
Confirmed: 87,733 women plus 46,632 men = 134,365; 2,302,712 person-years; 1139 colon and 339 rectum cases. The NHS FFQ has 61 items (1980) and the HPFS FFQ 131 items (1986). Diet is from baseline only, so diet_repeated = no. Outcome ascertainment is self-report verified by records, plus the NDI. All correct.

## Missing estimates
None. Tables II and III hold all the meat estimates. The paper has no poultry or fish estimates.

## Rows confirmed correct
wei2004_001-060 (all fields). For 061-062, every number is correct; the only problem is dose_unit (see above).

## Internal inconsistencies in the paper
- Table II header "(servings/day)" vs category labels in per week / per month.
- Table III "per week" vs Methods "servings per day" (see above).
- NHS processed meat, colon, 1 per week: "1.10 (0.81, 1.26)" is strongly log-asymmetric and is probably a typo (perhaps 1.01 or 1.10 with a different interval). It is recorded as printed and flagged in the notes, and the K3 symmetry check confirms the flag.
- The abstract and Results say main dish was "related to colon cancer". The trend p-values in Table II are 0.44 (NHS), 0.70 (HPFS) and 0.25 (combined). Only single categories approach significance.

## Verdict
Fit for use for the category rows (001-060). Hold 061/062 out of dose-response work until the unit is resolved or handled as a two-way sensitivity analysis. The cohort_id coding is correct.