# K10 control – giovannucci1994 (Giovannucci et al., Cancer Res 1994;54:2390-7)

Source: 22_Giovannucci1994.txt only (two-column text layer with heavy layout loss; in this file a "5" is systematically rendered as "3" inside numbers, e.g. "337 (138-8.06)" for the abstract's 3.57 (1.58-8.06)). No PDF checked.
Rows checked: 41.

## Discrepancies
| estimate_id | field | in database | in paper | severity |
|---|---|---|---|---|
| giovannucci1994_011 | category_label | <1/month | Table 2 lowest header printed "0" for processed meats; "<1/month" is supported by prose only for the main-dish item (abstract). It could equally be "0/never". | minor (unsupported assumption, flag) |
| giovannucci1994_027 | dose_point | empty | printed "036". Given the file's 5→3 substitution, 0.56 is more likely than 0.36, but this is unresolved. Empty is acceptable. | minor |
| adjustments giovannucci1994_m2 | energy_method | covariate | energy controlled by stratified analysis (Table 2 fn c). "covariate" is the closest vocabulary value; add a note. | minor |
| giovannucci1994_031/032 | model_label | Multivariate (text) | text says "multiple logistic model"; methods name proportional hazards. The note already covers this. | none |

No numeric discrepancies were found. Each printed number was checked against the text file and, where the abstract or Results prose repeats it, against that prose.

## Rows resting on OCR correction
| estimate_id | correction | supported by prose? |
|---|---|---|
| _002 | upper CI "134" → 1.54 | **No**. Supported only by log-symmetry (1.54 is symmetric with 0.62 around 0.97; 1.34 is not) and the file's 5→3 pattern. UNSUPPORTED BY PROSE, plausible. |
| _003 | upper CI "136" → 1.56 | **No**. Same basis as _002. UNSUPPORTED BY PROSE, plausible. |
| _004 | median "883" → 88.3 | decimal-point restoration only; consistent with the monotone medians. Acceptable. |
| _005 | median "1293" → 129.3; upper CI "235" → 2.55 | CI is **yes**: abstract gives 1.71 (1.15–2.55), P trend 0.005. Median is decimal restoration. |
| _005 | person_years left empty (printed "529,689") | n/a. Probably 52,968 or 52,969, so leaving it empty is correct. |
| _006 | "12124,870" → 12 / 24,870 | **No prose**. A crude rate check supports it: 12/24,870 vs 44/49,307 gives a crude RR of about 1.85, against a printed age-adjusted 1.92. Accept. |
| _006 label | header "0" → <1/month | **Yes**: abstract says "less than once per month". |
| _010 | "337 (138-8.06)" → 3.57 (1.58–8.06) | **Yes**: abstract, word for word. |
| _017 – _020 | poultry lower CIs "034/030/032/034" → 0.54/0.50/0.52/0.54 | **No**. Symmetry only; the alternatives (0.34 etc.) would be grossly asymmetric. UNSUPPORTED BY PROSE, plausible. |
| _024 | fish lower CI "031" → 0.51 | **No**. Symmetry only. UNSUPPORTED BY PROSE, plausible. |
| column-5 placements (all Table 2 sets) | the Q5 cell is printed first in each row in the text file and was moved to column 5 | **Partially**. It matches the abstract for red meat (RR 1.71) and main dish (3.57). For the other rows, a consistent pattern across all 6 rows supports it. Accept. |
| _028 | person_years left empty (printed "428,232") | n/a. Correct to leave it empty. |

Summary: 6 rows (_002, _003, _017, _018, _019, _020) plus _024, so 7 rows in all, carry CI digits that rest only on symmetry. This is low risk, but they should carry `verification_status = disputed` or a note until someone sees a PDF image.

## Missing estimates
None within the codebook scope. Not extracted, and out of scope: Table 3 "red meat protein" quintiles (a nutrient); the text "fat from red meat" RR 1.39 and "fat from chicken" RR 0.60 (fat sources). Hamburger (P=0.04), bacon (P=0.11) and chicken with skin (P=0.06) have P values only.

## Confirmed
All remaining fields in rows 001–041 are confirmed: case counts and person-years in Table 2, medians, RRs, P trends (0.005, 0.01, 0.06, 0.27, 0.79, 0.015), the text estimates for the multivariate model (1.66, 1.04–2.65; 3.07, 1.35–6.98), distal 1.78 (0.97–3.25), proximal 0.87 (0.43–1.76), rectum 1.22 (0.36–4.14), colorectal 1.66 (1.14–2.42) and ratio 1.81 (1.22–2.67), and the smoking strata (2.18, 5.11, 1.38, 1.60). Source pages are correct (Table 2 p.2392; multivariate text p.2391; subsite and smoking text p.2393). studies.csv is correct: 47,949 men aged 40–75, 131-item FFQ, 251 colorectal / 205 colon / 46 rectal cases, follow-up 1986 to Jan 1992. Red meat includes hot dog, preserved meats and bacon (Table 2 fn a), so includes_processed=yes is right.

Case sums: red meat 201, main dish 201, processed 202, poultry 205, fish 205, ratio 172 (subset eating ≥1 serving/day). All are ≤205, consistent with fn b ("may not add up to 205").

## Internal consistency
Text and tables agree. Methods name proportional hazards, while Results call the multivariate model "multiple logistic"; this is a wording inconsistency in the paper.

## Verdict
Fit for use. Seven rows carry CI digits restored by symmetry alone; these should be confirmed from a PDF image if the study enters any computation that uses the CIs of Q2/Q3 red meat or of poultry/fish.