# K3 control: singh1998 (Singh & Fraser 1998, Am J Epidemiol 148:761-774)

Controller: K3 (independent). Full text: `Understudier WCRF/singh1998.txt`. Date: 2026-10-30.
Method: Tables 3 and 4 were compared by script against values I transcribed by hand from the text (15 rows: point, CI, cases, person-years, p trend). The text estimates (p. 764-770) and the Appendix values for Figures 1-3 (p. 774) were checked cell by cell by reading (57 rows).

## Summary
- Rows: 72. Checked: 72. Numbers confirmed: 72/72.
- Discrepancies: critical 0, major 0, minor 2.
- Missing estimates: none.

## Discrepancies

| estimate_id | field | in database | in paper (table/page) | severity |
|---|---|---|---|---|
| singh1998_007-009, 013-015, 020-023, 039-044 (16 rows) | exposure_includes_processed | no | White meat is defined only as "Current intake of poultry or fish" (Table 3 fn). The paper says nothing about processed poultry or fish | minor (the codebook says unknowns should be coded `unclear`) |
| cup2016_singh1998_colon | cases_correct | 157 | 157 is the total in the Cox model; missing intake was retained by the Woodward method (p. 763). Only 127 cases fall in the three red-meat categories with RRs (42+40+45, Table 3) | minor (interpretive). For category-based dose-response, 127 is the usable count, so CUP's 127 is not wrong. Record both: "157 total; 127 with known red-meat intake" |

## The case counts (127 vs 157)
- 157 is the analytic total: 135 colon plus 22 rectosigmoid junction (p. 762), 178,544 person-years (p. 763).
- The Table 3 category sums are: total meat 140 (32+33+75), red meat 127 (42+40+45) and white meat 138 (36+76+26).
  - The person-years behave the same way. Red meat sums to 163,858 against 178,544 in total, so the remaining cases and person-years sit in a missing-intake indicator category with no printed RR.
- The extractor's notes and studies.notes describe this correctly. For dose-response from the categories, 127 cases are the right input; 157 is the right figure for "cases in the analysis". The CUP row's `cases_correct = 157` should be reworded so it does not imply that CUP made an error.
- The outcome includes the rectosigmoid junction (ICD-9 154.0, 22 cases), so "colon" here is colon plus rectosigmoid. studies.notes already says so.

## Definitions, models and flags (checked by reading)
- Red meat = "current intake of beef or pork" (Table 3 fn); beef covers hamburger, steak and other (p. 762). Processing status is not stated, so `red_undefined` / `unclear` is correct.
- Total meat = beef, pork, poultry and fish and/or any meat, so `total_meat` / `unclear` is correct.
- m1 matches the Table 3 and Table 4 footnotes: age, sex, BMI, physical activity, parental history, current and past smoking, alcohol and aspirin. All the flags are correct.
- m2 (Figures 1-3: age, sex and parental history only) is correct.
- m3 (the 4-level text analysis, all flags unclear) is correct; the paper gives no covariates for it.
- is_main_model = yes only on Table 3 (001-009); Table 4 is a subset restricted to subjects with low intake of the other meat.
  - This is defensible, and it is what CUP used (004-006). Note, however, that the abstract presents the Table 4 subset (RR 1.90) as the red-meat result.
- Doses: Never / >0 to <1 per week / ≥1 per week are recorded correctly. In the 4-level text rows, 1-4 and >4 per week are also correct.
- Figure cells:
  - Within-stratum contrasts (reference in the same stratum) carry a set and rank. Cross-stratum joint cells carry no set. That is a sound rule.
  - Row 047, "RR = 1.00, 95 percent CI 0.52-1.95", is correctly coded as a non-reference estimate.
- Source pages are correct: Table 3 on 766, Table 4 on 767, Appendix on 774, sensitivity analyses on 768 and 770.

## Values checked from the Appendix and text
- Figure 1 (9 cells), Figure 2 (12 cells) and Figure 3 Total, Men and Women (8 cells each): every point and CI matches.
- The 4-level red and white meat estimates (p. 764/766) and the four sensitivity estimates 3.43, 3.03, 2.82 and 3.42 match.

## studies.csv
Confirmed: n = 32,051; FFQ with 55 food-frequency questions; questionnaire 1976-77, follow-up to 1 Jan 1983; outcome via hospital-record review plus registry linkage (mixed); validation r (meat index 0.83, beef index 0.49) correctly not entered as validation_r for "red meat".

## Missing estimates
None. The text reports one model with both total meat and legumes (p. 767) but gives no numbers. Sex-stratified analyses (65 men / 92 women) are reported for non-dietary factors only.

## Rows confirmed correct
All 72 numerically. Apart from the flag on the 16 white-meat rows listed above, all fields are confirmed.

## Internal inconsistencies in the paper
- The abstract gives total meat 1.85 (95% CI 1.19-2.87); Table 3 gives 1.85 (1.16-2.87). The extractor recorded the Table value and flagged the conflict. My log-symmetry check agrees that 1.19 is the more plausible lower bound (lower arm 0.467 vs upper arm 0.439 on the log scale with 1.16; 0.441 with 1.19).
- Red meat in the 4-level analysis gives 1.38 for >0-<1/week, while Table 4 gives 1.40 for the same category. White meat gives 1.55 in both. So the 4-level model is a separate fit; this is consistent with how it is coded.

## Verdict
Fit for use. All numbers are correct. Change the white-meat flag to `unclear` and reword the CUP case-count note.