# K10 control – phillips1975 (Phillips, Cancer Res 1975;35:3513-22)

Source: 14a_Phillips1975.txt only (text layer of the AACR PDF, with Table 5 scrambled). No PDF image checked.
Rows checked: 16.

## Method of verification
Table 5 prints the raw 2×2 data "(case+, case−, cont+, cont−)" for each dichotomy and extreme-thirds contrast, and fn a defines RR = ad/bc. I recomputed every point estimate from the printed raw data:

| row(s) | raw data | ad/bc | printed | DB |
|---|---|---|---|---|
| meat 20 yr ago (001/002) | 26,9,51,50 | 2.83 | "2.8'" (displaced) | 2.8 ✓ |
| beef products, any (003/004) | 22,19,36,70 | 2.25 | "2.Y" | 2.3 ✓ |
| beef products, extremes (005/006) | 11,19,13,70 | 3.12 | 3.1' | 3.1 ✓ |
| hamburger, any (007/008) | 21,20,31,74 | 2.51 | 2.5' | 2.5 ✓ |
| hamburger, extremes (009/010) | 11,20,18,74 | 2.26 | 2.3 | 2.3 ✓ |
| lamb, any (011/012) | 12,29,14,91 | 2.69 | 2.7' | 2.7 ✓ |
| fish, any (013/014) | 18,23,34,71 | 1.63 | "I .6" | 1.6 ✓ |
| fish, extremes (015/016) | 12,23,11,71 | 3.37 | 3.4' | 3.4 ✓ |

The case and control counts in the category rows (n_cases = case+/case−, n_in_category = cont+/cont−) match the raw data in every row. The significance markers (') are coded correctly as p_value <0.05: 002, 004, 006, 008, 012 and 016 are marked; 010 and 014 are not.

## Rows resting on OCR correction
| estimate_id | correction | supported by prose? |
|---|---|---|
| _001/_002 | the meat-20-yr row is scrambled; raw set (26,9,51,50) and RR 2.8 were assigned to it | **Yes**: abstract says "relative risks for colon cancer of 2.8 for past use of meat", and ad/bc = 2.83. Of the printed raw sets near the scrambled rows, this is the only one that reproduces 2.8. Accept. |
| _004 | "2.Y" → 2.3 | **Yes**: abstract says "2.3 for beef", and ad/bc = 2.25. |
| _012 | 2.7 | **Yes**: abstract says "2.7 for lamb". |
| _014 | "I .6" → 1.6 | ad/bc = 1.63. There is no prose number, but the text says "any use of beef, lamb, fish ... significantly related". That conflicts with the absence of a significance marker on 1.6 (see below). |

## Discrepancies
| estimate_id | field | in database | in paper | severity |
|---|---|---|---|---|
| studies.csv | design | case_control | not in the design vocabulary (cohort, nested_cc, case_cohort, pooled_cohorts, meta_analysis). It is a hospital case-control study among Adventists, not cohort follow-up. Either extend the vocabulary or flag it. Keep the study out of cohort pools. | major (design/eligibility) |
| all | effect_measure | RR | printed as "relative risk" but defined as ad/bc, which is an odds ratio (fn a). OR is more accurate; RR is "as published". | minor |
| _001/_002 | exposure | total_meat (includes_processed unclear) | fn e, "Includes meat, poultry, and fish", most plausibly attaches to "Meat, any type, 20 yr ago". So the exposure is "any flesh food 20 yr ago" (vegetarian vs non-vegetarian status in the past). Add a note that fish and poultry are included. | minor |
| _002 | p_value | <0.05 | abstract says "statistically significant". The marker itself is lost in the scramble ("2.8'" carries the mark). Consistent. | – |

## Missing estimates
None that are required by R1. Out of scope or optional: "High saturated fat foods (heavy+medium vs light)" 2.1 / extremes 2.V(?) is a composite of cheese + eggs + all meats and poultry, not a meat variable. Current use of meat/poultry/fish, beef steak, other beef and poultry are listed only as "not associated" (Table 6, p.3519), with no numbers.

## Confirmed
All 16 rows numerically. source_page 3519 is correct (Table 5 sits between the 3518 and 3519 page footers). studies.csv is correct: 41 cases, 3 matched controls each, discharges 1969–73. The note that outcome = colon (Table 5 "colon cancer cases"; text "colon-rectal") is accurate.

## Internal consistency
The text says any use of "beef, lamb, fish" is significantly related, but fish any-vs-none (1.6) carries no significance marker in Table 5; only fish extremes (3.4) does. Minor text/table inconsistency in the paper.

## Verdict
Numerically correct. Not a cohort, so it is not fit for cohort pooling; use it only as historical context. These are crude ad/bc odds ratios with no CIs.