← Tilbake til oversikten (norsk)

Red meat and colorectal cancer in the WCRF/AICR Continuous Update Project 2016/2017: a reproduction and re-analysis of the cohort dose–response meta-analyses

Reproducibility analysis. Version 1.0, 30 October 2026.

Abstract

Background. The 2017 colorectal cancer report of the World Cancer Research Fund/American Institute for Cancer Research Continuous Update Project (CUP) judged red meat "probably a cause of colorectal cancer", on the basis of three cohort dose–response meta-analyses in the 2016 systematic literature review (SLR): colorectal 1.12 (95 % CI 1.00–1.25), colon 1.22 (1.06–1.39) and rectum 1.13 (0.96–1.34) per 100 g/day. We tested whether these estimates reproduce and whether they survive correction of extraction deviations.

Methods. Following a pre-specified protocol we re-pooled the SLR's printed per-study estimates (DerSimonian–Laird); reconstructed each per-study estimate from the primary papers by Greenland–Longnecker; classified every difference between the SLR's inputs and the primary papers; re-ran each analysis with deviations corrected and with later follow-ups substituted; and ran a multiverse over defensible extraction choices.

Results. All three pooled estimates reproduced exactly, and 21 of 26 per-study inputs reproduced from the primary papers within ±0.03. The reconstruction showed which doses CUP had used, including total-meat medians for the men of the Japan Public Health Center study. Deviations included a total-red-meat row (with processed meat) for the Multiethnic Cohort although the paper prints an unprocessed row, a near-total-meat variable (Järvinen 2001), and one cohort entered twice in the colon analysis. Corrected for all deviation classes, estimates were colorectal 1.08 (0.93–1.26), colon 1.22 (1.05–1.42) and rectum 1.11 (0.87–1.40); with later follow-ups, colorectal 1.04 (0.97–1.13) and colon 1.07 (0.94–1.20). The colorectal lower confidence limit exceeded 1 in 16 % of 768 multiverse combinations; the colon lower limit in 81 % of 384.

Conclusions. CUP 2016's arithmetic is exact. Its colorectal estimate depends on extraction choices, chiefly which EPIC estimate is used. Its colon estimate is robust to extraction choices but not to replacing two cohorts with their later publications.

1. Introduction

The Continuous Update Project (CUP) of the World Cancer Research Fund (WCRF) and the American Institute for Cancer Research (AICR) keeps the evidence base of the 2007 Second Expert Report up to date through systematic literature reviews (SLRs) conducted at Imperial College London. The colorectal SLR of 2016 [1] covered publications to 30 April 2015 ([2], p. 14) and underpins the colorectal cancer report of 2017, revised 2018 [2], which forms part of the Third Expert Report.

For red meat ("beef, pork, lamb, and goat from domesticated animals"; [2], p. 8, footnote 10) the SLR pooled eight cohort studies for colorectal cancer, 11 study rows for colon cancer and seven for rectal cancer, per 100 g/day ([1], Figs. 125, 132 and 139, pp. 355, 363, 370). The report states that the colorectal analysis "showed no significant association (RR 1.12 (95% CI 1.00–1.25))" ([2], p. 34). The Panel concluded: "The result was positive, but not significant, for colorectal and rectal cancers and significant for colon cancer, with low heterogeneity … Consumption of red meat is probably a cause of colorectal cancer" ([2], p. 36). The colon estimate, 1.22 (1.06–1.39), is thus the only statistically significant cohort result behind the "probable" grading.

CUP 2016 is the natural centre of a reproduction exercise for two reasons. It is the most recent synthesis on which current guidance rests, replacing the 2010 SLR [3] and its journal version, Chan et al. 2011 [4] (colorectal 1.17 (1.05–1.31); [1], Table 75, p. 336). And its estimates are not summaries of published numbers. Only one of the 13 studies, EPIC, published a per-100 g estimate that CUP used directly ([1], Table 79, p. 343). For the others the SLR team either built the slope from category data with portion-size, midpoint and energy-conversion rules, or rescaled a per-serving estimate ([1], Table 79, pp. 341–346; [3], p. 155). Each input is an analytic product with its own choices, and the pooled estimate inherits them all. This paper checks those inputs against the primary papers one by one.

2. Methods

2.1 Protocol

The re-analysis follows a protocol fixed before any run (REANALYSE_PROTOKOLL.md, v1.0, 30 October 2026). It requires reproduction of each pooled cohort estimate with the synthesis's stated model and increment (rules R1–R3), derivation of unprinted per-study estimates under the synthesis's own dose rules (R4), a classified deviation list (R5), corrected runs one class at a time and all together, a run with later follow-ups of the same cohorts, and leave-one-out for every run (R6). Reproduction succeeds when the pooled point estimate and both limits match to two decimals (R3). Pooling uses DerSimonian–Laird random effects [5]; per-study slopes are derived by the Greenland–Longnecker method [6] in its generalised least-squares form [7].

2.2 The synthesis's own method, as stated

The SLR forest plots state "Weights are from random effects analysis" ([1], pp. 355, 363, 370). The conversion rules are those of the 2010 SLR, whose methodology CUP 2016 kept "largely unchanged" ([2], p. 14):

"all results were converted to a common scale (grams per day) and 120 grams was used as a standard serving or portion size in studies that reported frequency of intake … For studies that presented the results in grams per 1000 kcal per day the intakes were converted to absolute intakes using the mean or median energy intake reported by the studies" ([3], p. 155).

SLR 2016 Table 79 records what was derived per study ([1], pp. 341–346): person-years by category, "Mid-points of exposure categories", "Conversion from servings/week to g/day"; for the Multiethnic Cohort (MEC) "g/1000kcal/day converted to g/day using average energy intake per each quantile"; for EPIC "Continuous results were directly used in dose-response analysis". The exposure rule (spelling as printed) is:

"In this analysis we included red meat was classified as beef, pork, lamb, hamburgers and fresh red meat. Studies which combined processed and unprocessed red meat where excluded from this analysis, and included in the red and processed meat analysis" ([1], p. 335).

For EPIC the 2010 SLR explains: "The observed result from EPIC (Norat, T. et al., 2005) was selected instead of the calibrated results used in the SLR for comparison with other studies that had not corrected for measurement error" ([3], p. 156).

2.3 Data

CUP's 26 per-study dose–response rows (eight colorectal, 11 colon, seven rectum) were transcribed from Figs. 125, 132 and 139 into database/cup_extraction.csv with printed weights, cases, contrasts, Table 79 page, and links to the primary-paper estimates CUP used and those its rules imply. Primary-paper estimates come from database/estimates.csv, one row per published estimate (every category, model, sex and subsite) for the 13 primary papers [8–20] and their follow-ups [21–24]; doses from database/dose_conversions_cup.csv; per-100 g conversions from database/derived_per100g.csv.

2.4 Reproduction and reconstruction

Test 1 re-pooled the SLR's printed per-study values. Test 2 reconstructed each of the 26 values from the primary paper's category data by the Greenland–Longnecker method [6, 7] (scripts/gl_doseresponse.py, scripts/test2_cup_rows.py) under the SLR's stated rules; where the literal rule failed, alternative readings were tried and recorded. A row counted as reproduced when its point estimate was within ±0.03 of CUP's; limits are reported alongside.

2.5 Deviations, corrections and follow-ups

The protocol classes are exposure (row includes processed meat or is the wrong variable), dose, cases, model (observed vs calibrated; unadjusted vs adjusted for processed meat), duplicate and transcription. Following the reporting brief, a deviation is where the synthesis differs from what the primary paper prints or breaks its own stated rule; an assumption is a defensible choice within the stated method (120 g per serving, the dose for an open top category). Assumptions are not corrected but varied in the multiverse. Protocol model entries that reproduce a stated or defensible CUP choice are run in the same way but discussed as sensitivity analyses (Section 4.1).

Where the primary paper prints an alternative, it replaced CUP's row. Otherwise the row was removed (Järvinen 2001, near-total meat) or kept and documented (Takachi 2011 exposure; Wei 2004; Bostick 1994; case counts). Follow-up runs substituted EPIC → Papadimitriou 2022 [21], JPHC → Islam 2019 [22] and NHS/HPFS → Bernstein 2015, Supplement S5 [24]. Because Islam 2019 pools JPHC with the Takayama study, the Takayama row (Oba 2006) is dropped whenever Islam 2019 is used, so that no cohort is counted twice; the later Takayama publication, Wada 2017 [23], enters only the multiverse, and only in combinations that keep Takachi 2011. All runs come from scripts/reanalyses.py (pooling in scripts/pool.py, DerSimonian–Laird [5]), output kontroll/reanalyser/cup2016.md/.json.

Run names are quoted verbatim so that every number can be traced; the published-values runs are labelled by the SLR figure and page that print their inputs (Fig. 125, p. 355; Fig. 132, p. 363; Fig. 139, p. 370).

2.6 Multiverse

scripts/multiverse.py pooled every combination of choices CUP made or could have made under its own rules (kontroll/multiverse_report.md). Colorectal, 768 combinations: MEC total or unprocessed row; EPIC observed, calibrated or Papadimitriou 2022; Järvinen in/out; English unadjusted or adjusted for processed meat; Lee in/out; Larsson portion 100, 120 or 150 g or the paper's own per-100 g estimate; Tiemersma in/out; random or fixed effect. Colon, 384 combinations: JPHC with CUP's doses, correct doses or Islam 2019; Takayama as Oba 2006 or Wada 2017 (Takayama row dropped when Islam 2019 is chosen); SMC as two rows or one; EPIC observed or calibrated; Järvinen, Singh and Bostick each in/out; random or fixed effect. The outcome is the share of combinations whose lower 95 % limit exceeds 1.

2.7 Extraction and control

All extraction, coding and checking was done by AI agents (Anthropic Claude, Opus models) under the author's direction, following a written codebook (KODEBOK.md). Every row of the estimates database was re-checked against full text or supplement by independent agent instances that had not seen the extraction: of 10,383 rows, 10,353 were verified, 18 corrected and 12 left disputed because the papers' own numbers are internally impossible; no numerical error was found in the 13 CUP studies (database/estimates.csv, status 30 October 2026; corrections in kontroll/rettelogg.md). CUP's per-study inputs and the three pooled analyses were checked independently against the SLR pages on 20 October 2026, with no discrepancy in any number attributed to CUP (Understudier WCRF/kort/kontroll_rapport.md). The separate synthesis-extraction control K11 covered the earlier syntheses, not CUP 2016. All pooling is done by script.

3. Results

3.1 Reproduction of the pooled estimates (Test 1)

All three published estimates reproduced exactly (Table 2): colorectal 1.12 (1.00–1.25), k = 8 (run: CRC – published per-study values (SLR 2016 Fig. 125, p355)); colon 1.22 (1.06–1.39), k = 11 (run: Colon – published per-study values (SLR 2016 Fig. 132, p363)); rectum 1.13 (0.96–1.34), k = 7 (run: Rectum – published per-study values (SLR 2016 Fig. 139, p370)). Weights agree with the printed ones to rounding (MEC 35.0 % vs 35.10 %; EPIC 24.9 % vs 24.84 %), as does I² (23 % vs 23.6 %; colon 12 % vs 11.7 %).

The SLR's colorectal sensitivity range, "from 1.09 (95% CI=0.96-1.25) when EPIC (Norat, 2005) was omitted to 1.19 (95% CI=1.06-1.34) when MEC (Ollberding, 201[2]) was omitted" ([1], p. 334), is identical to our leave-one-out. Two other statements did not reproduce. For colon "when Larsson, 2005 was omitted" the SLR gives 1.09 (1.03–1.16) ([1], p. 335); removing both SMC rows gives 1.18 (1.05–1.34) (run: Colon – corrected: duplicate, leave-one-out of the combined SMC row). For rectum without EPIC the SLR gives 1.05 (0.94–1.17) ([1], p. 335); we obtain 1.08 (0.87–1.33) (leave-one-out, run: Rectum – published per-study values (SLR 2016 Fig. 139, p370)). Both printed intervals are narrower than the fixed-effect interval of the full analysis (colon 1.08–1.37; rectum 0.96–1.34). On the log scale no subset can have a narrower interval than the full set's fixed-effect interval, so the printed values cannot be correct as stated. We could not identify how they arose.

3.2 Reconstruction of the per-study inputs (Test 2)

Twenty-one of 26 per-study values reproduced from the primary papers within ±0.03 (Table 1A; kontroll/test2_report.md); in 18 of these both confidence limits also matched within 0.02. The exceptions are Wei colon (CUP 0.83–1.50; ours 0.90–1.34), Wei rectum (0.50–1.56; 0.59–1.30) and English rectum (0.99–2.49; 0.94–2.38).

Five rows could not be reproduced. Lee 2009 (three rows): the available manuscript version lacks confidence intervals for the middle categories (database-samledom.md, §3). Tiemersma 2002 (CUP 1.69 (0.88–3.23); ours 1.92 (0.85–4.31)) and Singh 1998 (2.31 (0.63–8.45); ours 2.49 (0.46–13.39)) turn on the unpublished dose CUP assigned to an open top category. Our literal rule for Singh reproduces exactly the value in the 2010 SLR, 2.49 (0.46–13.39) ([3], Fig. 74, p. 170), which CUP 2016 replaced without comment. Lee's values also changed between reviews without explanation: colorectal 0.80 (0.52–1.23) in 2010 ([3], Fig. 70, p. 166) and 0.81 (0.50–1.29) in 2016; colon 0.85 (0.45–1.61) and 0.86 (0.43–1.74).

The reconstruction also revealed the dose rules CUP actually applied:

1. Larsson 2005 (SMC). All four rows, including distal colon 2.59 (1.40–4.79), reproduce exactly only with the paper's category medians (1.5, 2.5, 4.0, 5.5 servings/week) × 120 g / 7 (dose_conversions_cup.csv). Table 79 says "Mid-points of exposure categories" ([1], p. 343). Medians are the better practice, but not what is stated.
2. Takachi 2011 (JPHC). Colon 1.36 (1.06–1.73) and rectum 0.96 (0.66–1.41) reproduce only when men's doses are the total-meat medians, 20–117 g/day, not the red-meat medians, 15–102 g/day, printed in the same table (Takachi 2011, Table 2, p. 607). SLR Table 79 prints "117 vs 20 g/day" for men ([1], p. 342): the paper's Q5 and Q1 total-meat medians.
3. Ollberding 2012 (MEC). Reproduces exactly when g/1000 kcal is multiplied by each quintile's energy intake, as stated ([1], Table 79, p. 341).
4. English 2004 (MCCS). CUP raised the paper's hazard ratio per serving/week to the power 5.83 (7 × 100/120), reproducing all three rows to rounding.

3.3 Deviations

Table 3 lists every deviation with page references.

Exposure. For MEC, CUP used "Total red meat"; Ollberding's Table 2 also prints "Red meat excluding processed red meat", Q5 vs Q1 0.98 (0.87–1.10) against 1.02 (0.91–1.16). CUP's rule ([1], p. 335) points to the unprocessed row, which by the same method gives 0.97 (0.83–1.13) per 100 g/day. Järvinen's "red meat" is undefined; its quartile bounds are 91–94 % of those for "Meat, all" (men <94 vs <101 and >206 vs >220 g/day; women <61 vs <67 and >134 vs >145; Järvinen 2001, Table 3). The 2010 SLR flagged it as undefined and excluded it in a sensitivity analysis ([3], p. 155); the 2016 SLR includes it without comment. Takachi's "red meat" includes ham, sausage, bacon, luncheon meat and chicken liver (p. 604), with no unprocessed row. Bostick's main-dish item includes ham; Wei's excludes hamburgers, which CUP's definition includes.

Dose. The Takachi men's doses (Section 3.2).

Duplicate. SMC enters the colon analysis as a distal row (2.59) and a proximal row (1.17) ([1], Fig. 132, p. 363), whereas the colorectal and rectal analyses use one SMC row, and Chan 2011 used a single colon row, 1.68 (1.11–2.55) ([4], Fig. 4, p. 7). Combining the two 2016 rows by fixed effect gives 1.69 (1.11–2.56).

Cases. These do not affect the estimates but make the SLR's totals wrong. The colorectal total of 6,662 ([1], p. 334) is the sum of the per-study counts only with Järvinen's 63 colon cases; Table 79 gives 109 colorectal cases ([1], p. 345), which gives 6,708. Takachi's men enter colon with 259 cases ([1], p. 342); the paper's Table 2 gives 481, so the colon total of 4,081 is 222 too low. English's 452 are tumours; the paper counts 451 persons.

Model. For EPIC, CUP used the observed estimate; Norat 2005, Table 4, also prints calibrated estimates per 100 g: colorectal 1.49 (0.91–2.43), colon 1.36 (0.74–2.50), rectum 1.75 (0.93–3.30). For MCCS, CUP used the model not adjusted for processed meat; English 2004 gives the mutually adjusted model per serving/week, colorectal 1.02 (0.98–1.08), colon 1.00 (0.94–1.06) and rectum 1.07 (0.99–1.16) (p. 1511), i.e. 1.12 (0.89–1.57), 1.00 (0.70–1.40) and 1.48 (0.94–2.38) per 100 g/day.

3.4 Corrected and follow-up estimates

Colorectal. Each correction lowered the estimate and moved the lower limit below 1: exposure 1.10 (0.96–1.25), k = 7 (run: CRC – corrected: exposure); model 1.11 (0.97–1.26), k = 8 (run: CRC – corrected: model); all classes 1.08 (0.93–1.26), k = 7, I² 25 % (run: CRC – corrected: all classes). Adding the 2022 EPIC follow-up (5,069 cases; 58.9 % of the weight) gives 1.04 (0.97–1.13), I² 1 % (run: CRC – corrected + follow-ups). In the all-classes run no single omission lifts the lower limit above 1; the range is 1.04 (0.91–1.20) without EPIC to 1.15 (0.96–1.38) without MEC.

Colon. The estimate barely moved: dose 1.22 (1.06–1.40); duplicate 1.22 (1.08–1.37); exposure 1.21 (1.04–1.41); model 1.22 (1.05–1.43); all classes 1.22 (1.05–1.42), k = 9, I² 4 % (runs: Colon – corrected: dose / duplicate / exposure / model / all classes). In the all-classes run, only omitting Takachi 2011 brings the lower limit to 1 or below: 1.16 (0.97–1.38). With later publications, and the Takayama row dropped because Takayama is inside Islam 2019, the estimate is 1.07 (0.94–1.20), k = 8, I² 16 % (run: Colon – corrected + follow-ups), the weight moving to Bernstein 2015 (NHS/HPFS, unprocessed, 0.99 (0.87–1.12); 42.7 %) and Islam 2019 (1.05 (0.88–1.26); 29.1 %). Its leave-one-out range is 1.02 (0.92–1.12) without SMC to 1.13 (0.95–1.34) without NHS/HPFS.

Rectum. Exposure 1.12 (0.94–1.33), k = 6 (run: Rectum – corrected: exposure); dose 1.14 (0.96–1.35), k = 7 (run: Rectum – corrected: dose); model (EPIC calibrated 1.75 (0.93–3.30); MCCS adjusted for processed meat 1.48 (0.94–2.38)) 1.12 (0.91–1.37), k = 7, I² 1 % (run: Rectum – corrected: model); all classes 1.11 (0.87–1.40), k = 6, I² 13 % (run: Rectum – corrected: all classes). No follow-up run was produced for rectum (Section 4.4).

3.5 Multiverse

Colorectal (768 combinations). Median pooled RR 1.08 (range 1.03–1.16); lower limit above 1 in 123 combinations (16 %). EPIC decided it: 43 % significant with the observed estimate, 2 % with the calibrated, 3 % with Papadimitriou 2022. Next came Järvinen in/out (25 % vs 7 %) and Tiemersma in/out (21 % vs 11 %). Larsson's portion size mattered little (9 %, 11 %, 13 % for 100, 120, 150 g). No combination fell below 1.

Colon (384 combinations). Median 1.22 (range 1.11–1.28); lower limit above 1 in 312 (81 %). The Takachi dose error made no difference (100 % significant with either dose set). EPIC observed vs calibrated gave 96 % vs 67 %. The one choice that changed the picture materially was the JPHC publication: with Islam 2019 instead of Takachi 2011 the median fell to 1.13 and the significant share to 44 % (kontroll/multiverse_report.md).

3.6 Figure

Figure 1. Forest plots of every run (published, single-class corrections, all classes, follow-ups) with weights and leave-one-out, in the accompanying HTML (kontroll/reanalyser/html/index.html, section "WCRF/AICR CUP 2016/2017", run selector sel-cup2016, panel forest-cup2016; deviation table dev-cup2016).

Table 1A. CUP 2016 per-study inputs (per 100 g/day) and their reconstruction from the primary papers (Test 2)

Source: kontroll/test2_report.md; database/cup_extraction.csv; SLR 2016 Figs. 125, 132, 139.

Table 1B. Exposure as used by CUP against the primary paper and CUP's own rule ([1], p. 335)

Table 2. Published, reproduced, corrected and follow-up estimates (RR per 100 g/day, DerSimonian–Laird)

No follow-up run exists for rectum.

Table 3. Deviations between CUP 2016 inputs and the primary papers

4. Discussion

4.1 What the deviations are, and whose they are

CUP 2016's arithmetic is exact: every pooled estimate, every printed weight and the colorectal sensitivity range reproduce from its own inputs, and 21 of its 26 inputs can be rebuilt from the primary papers. The questions raised here therefore concern extraction, not computation.

Three deviations depart from what the primary papers print or from CUP's own rule. The MEC row: the SLR excludes "Studies which combined processed and unprocessed red meat" ([1], p. 335), and Ollberding 2012 prints a qualifying row beside the one CUP used. MEC carries 35 % of the colorectal weight, but because both rows are near null (1.01 vs 0.97 per 100 g) the substitution moves the pooled estimate less than one might expect. Järvinen 2001 is near-total meat by the paper's own bounds; CUP's 2010 review had already flagged it as undefined ([3], p. 155). The Takachi men's doses are the total-meat medians from the adjacent block of the same table, a transcription-level error, and the least consequential: corrected doses move the JPHC colon estimate from 1.36 to 1.39 and leave the pooled colon estimate at 1.22.

The SMC duplicate gives one cohort two weights; CUP's own colorectal and rectal analyses, and Chan 2011's colon analysis [4], treat it as one. The split also adds heterogeneity (2.59 vs 1.17). Combining the rows changes the colon point estimate by less than 0.01 and narrows its interval. The case-count errors do not enter the inverse-variance weights but make the totals quoted in the report ([2], pp. 34–35) inaccurate.

Two protocol model entries are not deviations by our definition: CUP's values match what the papers print and no stated rule is broken. CUP stated and justified the observed EPIC estimate ([3], p. 156) on grounds of comparability with uncorrected studies, a defensible position; the calibrated estimate better reflects true intake but is far less precise (0.91–2.43 against 1.02–1.43). The unadjusted MCCS model is consistent with how CUP treated cohorts that do not adjust red meat for processed meat. We report both alternatives as sensitivity analyses, not corrections of errors. Likewise, 120 g per serving, the open-top doses for Tiemersma and Singh, and the interpolated energy for MEC's second and fourth quintiles are assumptions within the stated method; the multiverse shows Larsson's portion size hardly matters.

Two things remain undocumented: the Singh colon value changed from 2.49 (2010) to 2.31 (2016) without comment, and Table 79 describes the SMC conversion as using midpoints while only medians reproduce the printed values. Neither changes a pooled estimate materially (Singh has 1.05 % of the colon weight), but both show that the method description does not fully record what was done. For completeness: Teresa Norat, first author of the EPIC paper [11], is listed as Principal Investigator of the Imperial College CUP team ([2], p. 88) and first author of the 2010 SLR [3], where the observed-estimate rationale is given. We note this because it bears on the EPIC choice, and infer nothing beyond the stated rationale.

4.2 What the corrected results mean

Colorectal. CUP's report already called the colorectal estimate non-significant ([2], p. 34); the corrections confirm that reading. With all deviation classes corrected it is 1.08 (0.93–1.26); with the larger EPIC follow-up, 1.04 (0.97–1.13). The multiverse locates the borderline lower limit of 1.00 in one choice, the observed EPIC estimate: with it 43 % of defensible combinations are significant, with the calibrated estimate or the 2022 follow-up 2–3 %. Across 768 combinations the pooled RR stays between 1.03 and 1.16. The direction is stable; under most defensible choices these cohorts lack the precision to separate it from null.

Colon. The colon estimate, the significant result the Panel singled out, is robust to every extraction correction: 1.21–1.22 under each class and all together, significant in 81 % of 384 combinations. This is a fair result for CUP: the colon finding does not depend on the errors documented here. It depends instead on the version of two cohorts. Takachi 2011 (21.9 % of the weight) uses a processed-inclusive "red meat" and was succeeded by Islam 2019, which isolates unprocessed red meat in a JPHC–Takayama pool [22]; Wei 2004 was succeeded by Bernstein 2015 [24], with 0.99 (0.87–1.12) per 100 g unprocessed. With these publications the colon estimate is 1.07 (0.94–1.20), and in the multiverse Islam 2019 instead of Takachi 2011 cuts the significant share from 100 % to 44 %.

Rectum. Null as published and null after correction.

These results concern the syntheses, not the causal effect of red meat. All inputs are observational, and re-analysis removes neither residual confounding nor dietary measurement error. The post-2015 cohort literature is analysed separately (kontroll/reanalyser/update2026.md).

4.3 The SLR's text against its own figures

Two sensitivity statements (colon without SMC, rectum without EPIC; [1], p. 335) give intervals narrower than the full analysis's fixed-effect interval and are wrong as printed. The SLR also says of rectal cancer that "No association of red meat intake and rectal cancer risk was reported in all studies" ([1], p. 335), while its own Fig. 138 shows English 2004 at 2.30 (1.20–4.20) for highest vs lowest ([1], p. 369). None of this alters a pooled estimate, but it affects how the review describes its own robustness.

4.4 Limitations

1. Not reproduced. Five inputs (Lee 2009 ×3, Tiemersma 2002, Singh 1998) were used as printed by CUP in every run. Summed from the printed weights, they carry 8.0 % of the colorectal, 4.5 % of the colon and 4.35 % of the rectal weight.
2. No rectum follow-up run. Papadimitriou 2022 was not substituted for rectum, and no follow-up run was produced for that outcome.
3. Takayama inside Islam 2019. Islam 2019 pools JPHC-5y with Takayama, and its supplement does not separate the cohorts (database-samledom.md, §4). An earlier version of the follow-up run and the multiverse counted Takayama twice (Islam 2019 plus a separate Takayama row); this was corrected before the runs reported here, which drop the Takayama row whenever Islam 2019 is used. The JPHC contribution therefore cannot be isolated from Takayama.
4. Conversions. The EPIC follow-up was converted from a per-standard-deviation estimate (36.2 g), which magnifies rounding in the published value. The NHS/HPFS colon value comes from Bernstein 2015's Supplement S5.
5. Sources. Primary papers were read as text, several from OCR of scanned pages; Lee 2009 only as the author manuscript. The SLR was read from a text extraction of pp. 334–375. Figure values were taken from printed numbers, not plot positions.
6. AI extraction. All extraction was done by AI agents. Independent re-extraction, a correction log, and the rule that every number is either printed in a source or produced by a published script mitigate this risk; they do not eliminate it.
7. Uncorrected exposure deviations. Where no alternative row exists (Takachi, Bostick, Wei) the deviation is documented, not corrected. Strictly applied, CUP's p. 335 rule would move Takachi and Bostick to the red-and-processed analysis; leave-one-out in each run shows the effect of omitting them.

5. Conclusions

CUP 2016's red meat–colorectal meta-analyses are computationally exact, and most of their inputs can be traced to the primary papers, which revealed the dose rules actually used. The borderline significance of the colorectal estimate rests on the observed EPIC estimate; under the SLR's own exposure rule and the alternatives the papers print, it is 1.08 (0.93–1.26), consistent with CUP's own description of that result as non-significant. The colon estimate, the significant result the Panel relied on, survives every extraction correction, but with the later publications of the Japanese and US cohorts it falls to 1.07 (0.94–1.20).

Data and code availability

Database: database/estimates.csv, cup_extraction.csv, dose_conversions_cup.csv, derived_per100g.csv, cohorts.csv, synthesis_extraction.csv, studies.csv; codebook KODEBOK.md. Scripts: scripts/reanalyses.py (all runs), pool.py (DerSimonian–Laird, leave-one-out), gl_doseresponse.py and test2_cup_rows.py (Greenland–Longnecker), reproduce_cup.py, scenarios.py, multiverse.py. Outputs: kontroll/reanalyser/cup2016.md/.json, kontroll/test2_report.md, kontroll/scenarios_report.md, kontroll/multiverse_report.md, forest plots in kontroll/reanalyser/html/index.html. Control: Understudier WCRF/kort/kontroll_rapport.md, kontroll/rettelogg.md, database-samledom.md; protocol REANALYSE_PROTOKOLL.md.

References

1. World Cancer Research Fund/American Institute for Cancer Research. Continuous Update Project Systematic Literature Review: The Associations between Food, Nutrition and Physical Activity and the Risk of Colorectal Cancer. London: Imperial College London CUP team; 2016. Section 2.5.1.3 Red meat, pp. 334–375. Available at wcrf.org/colorectal-cancer-slr.
2. World Cancer Research Fund/American Institute for Cancer Research. Continuous Update Project Expert Report 2018. Diet, nutrition, physical activity and colorectal cancer. 2017, revised 2018. Available at dietandcancerreport.org.
3. World Cancer Research Fund/American Institute for Cancer Research. WCRF/AICR Systematic Literature Review Continuous Update Project Report: The Associations between Food, Nutrition and Physical Activity and the Risk of Colorectal Cancer. Norat T, Chan D, Lau R, Aune D, Vieira R; statistical adviser Greenwood D. London: Imperial College London; 2010.
4. Chan DSM, Lau R, Aune D, Vieira R, Greenwood DC, Kampman E, Norat T. Red and processed meat and colorectal cancer incidence: meta-analysis of prospective studies. PLoS ONE 2011;6(6):e20456. doi:10.1371/journal.pone.0020456.
5. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials 1986;7:177–188.
6. Greenland S, Longnecker MP. Methods for trend estimation from summarized dose-response data, with applications to meta-analysis. Am J Epidemiol 1992;135:1301–1309.
7. Orsini N, Bellocco R, Greenland S. Generalized least squares for trend estimation of summarized dose–response data. Stata J 2006;6:40–57.
8. Ollberding NJ, Wilkens LR, Henderson BE, Kolonel LN, Le Marchand L. Meat consumption, heterocyclic amines and colorectal cancer risk: the Multiethnic Cohort Study. Int J Cancer 2012;131:E1125–E1133. doi:10.1002/ijc.27546.
9. Lee SA, Shu XO, Yang G, Li H, Gao YT, Zheng W. Animal origin foods and colorectal cancer risk: a report from the Shanghai Women's Health Study. Nutr Cancer 2009;61:194–205. doi:10.1080/01635580802419780.
10. Larsson SC, Rafter J, Holmberg L, Bergkvist L, Wolk A. Red meat consumption and risk of cancers of the proximal colon, distal colon and rectum: the Swedish Mammography Cohort. Int J Cancer 2005;113:829–834. doi:10.1002/ijc.20658.
11. Norat T, Bingham S, Ferrari P, et al. Meat, fish, and colorectal cancer risk: the European Prospective Investigation into Cancer and Nutrition. J Natl Cancer Inst 2005;97:906–916. doi:10.1093/jnci/dji164.
12. English DR, MacInnis RJ, Hodge AM, Hopper JL, Haydon AM, Giles GG. Red meat, chicken, and fish consumption and risk of colorectal cancer. Cancer Epidemiol Biomarkers Prev 2004;13:1509–1514.
13. Tiemersma EW, Kampman E, Bueno de Mesquita HB, Bunschoten A, van Schothorst EM, Kok FJ, Kromhout D. Meat consumption, cigarette smoking, and genetic susceptibility in the etiology of colorectal cancer: results from a Dutch prospective study. Cancer Causes Control 2002;13:383–393. doi:10.1023/a:1015236701054.
14. Järvinen R, Knekt P, Hakulinen T, Rissanen H, Heliövaara M. Dietary fat, cholesterol and colorectal cancer in a prospective study. Br J Cancer 2001;85:357–361. doi:10.1054/bjoc.2001.1906.
15. Pietinen P, Malila N, Virtanen M, Hartman TJ, Tangrea JA, Albanes D, Virtamo J. Diet and risk of colorectal cancer in a cohort of Finnish men. Cancer Causes Control 1999;10:387–396.
16. Takachi R, Tsubono Y, Baba K, Inoue M, Sasazuki S, Iwasaki M, Tsugane S. Red meat intake may increase the risk of colon cancer in Japanese, a population with relatively low red meat consumption. Asia Pac J Clin Nutr 2011;20:603–612.
17. Oba S, Shimizu N, Nagata C, Shimizu H, Kametani M, Takeyama N, Ohnuma T, Matsushita S. The relationship between the consumption of meat, fat, and coffee and the risk of colon cancer: a prospective study in Japan. Cancer Lett 2006;244:260–267. doi:10.1016/j.canlet.2005.12.037.
18. Wei EK, Giovannucci E, Wu K, Rosner B, Fuchs CS, Willett WC, Colditz GA. Comparison of risk factors for colon and rectal cancer. Int J Cancer 2004;108:433–442. doi:10.1002/ijc.11540.
19. Singh PN, Fraser GE. Dietary risk factors for colon cancer in a low-risk population. Am J Epidemiol 1998;148:761–774.
20. Bostick RM, Potter JD, Kushi LH, Sellers TA, Steinmetz KA, McKenzie DR, Gapstur SM, Folsom AR. Sugar, meat, and fat intake, and non-dietary risk factors for colon cancer incidence in Iowa women (United States). Cancer Causes Control 1994;5:38–52.
21. Papadimitriou N, Bouras E, van den Brandt PA, et al. A prospective diet-wide association study for risk of colorectal cancer in EPIC. Clin Gastroenterol Hepatol 2022;20:864–873. doi:10.1016/j.cgh.2021.04.028.
22. Islam Z, Akter S, Kashino I, et al. Meat subtypes and colorectal cancer risk: a pooled analysis of 6 cohort studies in Japan. Cancer Sci 2019;110:3603–3614.
23. Wada K, Oba S, Tsuji M, et al. Meat consumption and colorectal cancer risk in Japan: the Takayama study. Cancer Sci 2017;108:1065–1070.
24. Bernstein AM, Song M, Zhang X, et al. Processed and unprocessed red meat and risk of colorectal cancer: analysis by tumor location and modification by time. PLoS ONE 2015;10(8):e0135959. doi:10.1371/journal.pone.0135959.

Generated from the open database and scripts described in the article; extraction and control performed by AI agents (Claude Opus/Fable) under the author's direction. 30 October 2026.