# Confounding structure: how heavy red-meat eaters differ (baseline tables)

Data: `baseline_by_intake.csv` (225 rows; lowest vs highest printed intake category; table and page for every number). Sources: only the full texts on disk (paths are in the final section). Page = journal page; "n/a" = PMC/XML text without page numbers, so the section is cited. Where I have done any arithmetic, it is marked **(my calc)**.

## 1. Synthesis

**Eleven baseline tables by intake, from seven cohorts.** These are NIH-AARP twice (Cross 2010, Zouiouich 2025), UK Biobank twice (Knuppel 2020 Suppl. S2, Bradbury 2020 T1–2), NHS, HPFS (Bernstein 2015), SMC (Larsson 2005), MEC (Ollberding 2012) and the Danish DANSDA (Mejborn 2021, within diet-quality strata). The same pattern recurs in all the US and UK cohorts:

| Difference (high vs low red meat) | Typical size | Where it holds / breaks |
|---|---|---|
| **More men** | 45→73 % (AARP); 36→58 % (MEC); 29→61 % (UKB); 44→68 % and 29→53 % (DK) | Universal where printed |
| **Higher BMI** | +1.8 to +2.8 kg/m² (AARP +2.6/+2.5; UKB +2.2, +1.8 men, +1.8 women; NHS +1.9; HPFS +1.8; MEC +2.8) | SMC only +0.4. DK obese 12.6→14.3 % and 11.2→14.3 % |
| **More smoking** | Current smoking roughly doubles in AARP (7.7→17.1 %, Cross). UKB men 10.5→15.5 %, women 7.9→10.1 % (Bradbury). Heavy smokers (≥15/d) 2.3→6.2 % (Knuppel). Pack-years 7.6→13.2 (MEC) | Small in NHS/HPFS, which print only pack-years before age 30 (7.0→7.2; 10.6→12.1). **Absent within DK diet-quality strata** (39.6→37.4 %; 19.7→17.2 %). SMC and MCCS: smoking not reported |
| **Less physical activity** | Vigorous PA >5/wk 27.7→15.8 % (AARP); NHS 19.8→13.7 MET-h/wk; UKB low PA 28.2→32.3 % | Flat in HPFS (36.8→34.9) and MEC (0.39→0.42 h/d). Higher in DK high-meat groups |
| **Lower fibre / fruit / vegetables / whole grains** | Fibre density −30 to −35 % (AARP 13.8→9.1 g/1000 kcal; NHS 20.6→14.3 g/d; HPFS 26.9→18.7; MEC 14.8→9.7). F&V −23 to −39 % (UKB 5.59→4.33 servings/d; AARP-Z 2.82→1.71 cups/1000 kcal). Whole grains −43 % (AARP-Z) **(my calc for the %)** | **Opposite in SMC**: fruit, vegetables and whole grains rise slightly with red meat. Vegetables also rise in NHS/HPFS (3.01→3.28; 3.46→3.78). UKB cereal fibre is nearly flat (4.66→4.44). EPIC: r(fibre, red+processed) = −0.18/−0.21 |
| **Higher energy intake** | +16 to +19 % in AARP and MEC (1685→1978; 1554→1846; 1962→2269 kcal) and +37 % in SMC (1139→1555) **(my calc for %)**. MCCS r = 0.37 | Not available in UKB. Not printed in NHS/HPFS T1 |
| **Lower calcium, folate, vitamin D** | Calcium −13 to −35 % (SMC 745→648; MEC; AARP 501→361/1000 kcal; NHS 1279→837 mg) | Consistent |
| **Lower education** | College+ 46.7→37.3 % (AARP); degree 64.7→58.9 % (UKB). DK: 24.3→22.9 % and 37.6→35.5 % | SMC flat (11.7→11.7 %). **UK socio-economic status runs the other way**: highest Townsend fifth 14.7→19.6 % (men) and affluent fifth 16.0→20.1 % |
| **Alcohol** | Mixed. Higher in UKB (men 19.7→28.7 g/d), HPFS (9.1→13.4) and NHS (5.1→6.1) | Flat in AARP (11.9→11.0) and MEC (8.8→8.4). UKB (Knuppel S2, labels corrected after Papier 2021, R108): ≥20 g/d rises 15.3→37.8 % and non-drinkers fall 17.0→6.3 %, so UKB belongs under "Higher" |
| **Processed meat travels with red meat** | AARP processed red meat 1.87→14.06 g/1000 kcal; NHS 0.06→1.03 svg/d; SMC 9.6→38.3 g/d; UKB 24-h 9→29 g/d (men) | Universal |

**Other striking differences:**
- **Less colonoscopy/endoscopy screening:** NHS 21.2→17.0 %, HPFS 28.2→21.9 %. This is a direct CRC-detection confounder, and only NHS/HPFS report it.
- **Worse overall diet quality:** AHEI 52.0→38.5 and 48.1→37.3; DASH 26.4→20.5. Saturated fat is +56 % in NHS **(my calc)**.
- **Supplement use:** multivitamin use is lower in NHS/HPFS (56→45 %, 54→44 %) but higher in MEC (41→57 %).
- **Other health factors:** diabetes 7.9→15.0 % (MEC); HRT use is lower in MEC and NHS; coffee >4 cups/d 11.6→19.4 % (AARP-Z).
- **NSAID/aspirin use** is slightly *higher* in heavy eaters (AARP 63→68 %; UKB 25.6→29.8 %; NHS 47→53 %). This is protective confounding, i.e. it biases the HR towards the null.
- **Age** runs both ways: heavy eaters are younger in SMC (−7.8 y), MEC and AARP, and older in UKB (+2 y).

**Interpretation of the pattern.** Every classic CRC risk factor except NSAID use and age moves *against* the heavy meat eater in the US/UK cohorts: male sex, BMI, smoking, inactivity, low fibre, low calcium, alcohol (UK) and less screening. The **Swedish Mammography Cohort is the clear exception**. There, high red-meat women eat more fruit, vegetables, whole grains and fish, have similar BMI and education, and the gradient is mostly energy (total intake scale). The **Danish data** show that once diet quality is held fixed, most of the smoking, education and activity gradient lies *between* diet-quality groups, not between meat groups (Mejborn, Results 3 and Table 1).

**How much does adjustment remove?** (These are model results, not baseline numbers.)
- MEC: Q5 total red meat 1.26 basic → **1.02** multivariable (Ollberding T2 p.5).
- EPIC: red+processed 1.57 sex/energy-adjusted → **1.35** full (Norat p.909).
- UKB (red AND processed meat combined): per 50 g/d 1.26 → **1.17**; waist circumference and alcohol account for nearly all of the attenuation, smoking only 1.21 → 1.20 (Bradbury T3 p.255, text p.252 and p.255).
- NLCS: vegetarians vs 6–7 d/wk 0.73 → **1.06** with all diet and lifestyle factors; fibre alone moves it 9 % (Gilsing T3).
- AARP: BMI, fibre and calcium attenuated most (Cross p.2408).
- NHS/HPFS: unprocessed red meat goes from weakly positive (age-adjusted) to **0.99** (Bernstein p.5). The inverse distal association appears only after adjustment for calcium, folate and fibre (p.10, p.12).
- Million Women: red+processed 1.08 → 1.06 / 1.07 / 1.06 / 1.06 when calcium, milk, fruit or wholegrains are added one at a time (Papier T3 p.7; no model with all four). The authors call it "minimally affected" (p.6).

## 2. Studies with no baseline table by intake

| Study | What the paper has instead |
|---|---|
| **Norat 2005** (EPIC) | Table 1 is by centre and Table 2 by case status (p.910). Only a fibre–meat correlation is printed (p.912). |
| **Wei 2004** (NHS/HPFS) | Table I is by case status (p.436). |
| **English 2004** (MCCS) | Table 1 shows case counts by intake only (p.1511). Table 2 gives Spearman correlations (p.1512); minus signs were lost in extraction, see the CSV notes. |
| **Islam 2019** (Japanese pooled) | Table 1 describes the cohorts. Text: `Understudier WCRF/oppfolginger/islam2019.txt`; the `post2015/islam2019_tfa_WRONGPAPER.txt` file is a different paper. |
| **Tammi 2024** (Finnish pooled) | Baseline table by cohort. |
| **Gilsing 2015** (NLCS-MIC) | Table 1 by case status. Characteristics by meat group are "described previously" (ref. 9), which is not on disk. |
| **Papier 2025** (Million Women) | Table 1 by case status (p.2). |
| **Papadimitriou 2022** (EPIC/NLCS) | Supplementary Table 1 by case status (p.873.e4). |
| **Vulcan 2017** (Malmö) | Table 1 by case status. |
| **Cross 2007** | Not on disk; only Cross 2010 was found. |

**Knuppel 2020:** main-text Table 1 is by cancer status only. The by-intake table is **Suppl. Table S2**, and it is by red *and processed* meat combined.

## 3. Adjustment in the main model

Y = yes; N = not in main model; s = sensitivity analysis only.

| Study | Smoking | BMI | PA | Alcohol | Energy | Fibre/F&V | Education | Source |
|---|---|---|---|---|---|---|---|---|
| Norat 2005 | Y | weight+height | Y (occupational) | Y | Y | fibre Y | N | p.909 |
| Cross 2010 | Y | Y | **N** | **N** | Y | fibre Y, F&V N | Y | T2 fn p.2409 (kept only covariates changing HR ≥10 %) |
| Zouiouich 2025 | Y | Y | Y | Y | Y | F&V + whole grains Y | Y | p.3 |
| Knuppel 2020 | Y | Y (Model 2) | Y | Y | **N** (not available) | F&V + cereal fibre Y | Y | p.3–4, p.11 |
| Bradbury 2020 | Y | waist (BMI s) | Y | Y | **N** | **N** (bread/cereal fibre s) | Y | p.251, p.257 |
| Bernstein 2015 | pack-yrs <30 only | Y | Y | Y | Y | fibre Y, F&V N | **N** | T2 fn p.7 |
| Wei 2004 | pack-yrs <30 only | Y | Y | Y | nutrients energy-adj. | **N** | **N** | T fn p.440 |
| Larsson 2005 | **N** (s, 1997 data) | Y | **N** | Y | Y | F, V, whole grains Y | Y | p.830, p.831 |
| Ollberding 2012 | pack-yrs | Y | Y (vigorous) | Y | Y | fibre Y | **N** | p.3 |
| English 2004 | **N** (not measured) | **N** (tested) | **N** (tested) | **N** (tested) | Y | cereal products Y | **N** (tested) | p.1510 (≥5 % rule) |
| Mejborn 2021 | Y | Y | Y | Y | Y | **N** (stratified by diet quality) | Y | Methods 2.6 |
| Gilsing 2015 | Y | Y | Y | Y | Y | **N** (single-factor in T3) | Y | T2 fn b |
| Islam 2019 | Y | Y | Y | Y | Y | fibre Y | **N** | T4–5 fns |
| Tammi 2024 | Y | Y | Y | Y | Y | N (plant foods are the substitutes) | Y | Methods, Stat. analysis |
| Papier 2025 | Y | Y | Y | Y | Y | **N** (s, T3) | Y | p.9 |
| Papadimitriou 2022 | Y | Y | Y | **N** | Y | **N** | Y | p.866 |
| Vulcan 2017 | Y | Y* | Y | Y | Y | N (tested, "did not affect") | Y | Methods; *BMI is in Methods text but missing from table footnote b |

## 4. Author statements on residual confounding and the "unhealthy meat-eater" pattern

**Explicit "unhealthy meat-eater" framing:**
- **Mejborn 2021.** Introduction: "persons with different diets may differ on other characteristics, cohort studies on associations between meat intake and health have many confounders". Discussion: comparing low vs high meat "without considering dietary quality and what foods replace the meat" "will simultaneously be a comparison of healthy and unhealthy diets". BMI is "likely … a mediator", so the models are possibly over-adjusted.
- **Knuppel 2020, p.8.** The processed meat–colon association attenuated to null in never-smokers, which "might suggest residual confounding by smoking or an overall unhealthy lifestyle".
- **Bernstein 2015, p.12.** "Many diet and lifestyle behaviors are associated with red meat consumption … cannot exclude … residual or unmeasured confounding". The inverse distal finding appears only "after adjustment for specific nutrients associated with a healthier diet".
- **Ollberding 2012.** Residual confounding from "incomplete covariate assessment and adjustment" (p.6). The basic→multivariable attenuation may reflect covariate measurement error or over-adjustment "given the inherent difficulty in separating out the effects of such highly correlated behaviors" (p.7).
- **Gilsing 2015, Discussion.** The dietary and non-dietary factors that characterise low-meat diets attenuate the estimates by ≥17 % when combined, "approaching the null".

**Generic limitation sentences:**
- **Cross 2010, p.2412:** "some residual confounding may remain".
- **Zouiouich 2025:** BMI and PA measured only at baseline, "leaving the possibility for residual confounding" (p.10). E-values were 1.45–2.31 (p.8) for the point estimates of the substitution models (red meat in place of white meat; Suppl. Table 7); the authors read them as indicating robustness.
- **Knuppel 2020, p.11:** "there may still be some residual confounding".
- **English 2004, p.1513:** physical activity was measured imprecisely, so there "may therefore be some residual confounding". Dietary confounding was judged unlikely.
- **Larsson 2005, p.833:** inability to control for physical activity.
- **Tammi 2024, Limitations:** residual confounding (e.g. genetic). "residual confounding by smoking likely remains" in ATBC.
- **Islam 2019, Discussion:** limitations paragraph.
- **Papadimitriou 2022, p.871.**
- **Vulcan 2017, Discussion.**

**Arguments against confounding:**
- **Bradbury 2020, p.256:** the fully adjusted models were adjusted for 18 variables that are known or suspected confounders. Adding milk, cheese and cereal fibre did not change the results.
- **Papier 2025, p.6:** the red+processed meat association was "minimally affected by confounding by diet and lifestyle factors". By contrast, the fibre, fruit and wholegrain associations "may have been influenced by residual confounding".
- **Wei 2004 and Norat 2005:** no explicit residual-confounding statement for meat.

## 5. Source files

All under `/home/claude/drive/`:
- `Historikk 1965-2015/Historikk/txt/30_Norat2005.txt`
- `Understudier WCRF/{cross2010,wei2004,larsson2005,ollberding2012,english2004}.txt`
- `Understudier WCRF/oppfolginger/islam2019.txt`
- `post2015/{zouiouich2025,knuppel2020,knuppel2020_suppl,bradbury2020,bernstein2015,papier2025,tammi2024,mejborn2021,gilsing2015,papadimitriou2022}.txt`
- `post2015/sok2026/vulcan2017.txt`

`/home/claude/db/database/` was not touched.


## K16b (30.10.2026)
Independent control of all 225 rows against the full texts: 0 value errors, 2 page errors (Knuppel oily fish, suppl. p.11 – fixed). NOTES corrections applied: MWS stepwise adjustment, UKB exposure = red+processed and attenuation drivers, E-value scope, Mejborn conditional clause, cohort count. Counter-findings noted by the controller: in Zouiouich (AARP) NSAID use is *lower* in heavy eaters (53.9→51.6 %); UKB women show *more* HRT use (30.4→40.3 %); Cross quote "some residual confounding may remain" is p.2413. Bradbury rows 65/80/96/101 lost a "≥" in extraction (highest category ≥3 and ≥2 times/wk). Full report: ../K16b/RAPPORT.md.