← Tilbake til oversikten (norsk)

What the narrative reports said: WCRF 1997 and IARC Monograph 114 checked against the primary cohort papers

Re-analysis series, companion article. Draft of 30 October 2026.

Abstract

Background. Two influential judgements on red meat and colorectal cancer rested on no pooled estimate of their own. WCRF/AICR (1997) graded red meat a "probable" cause after a narrative review of seven cohorts. IARC Monograph 114 (evaluated 2015, published 2018) found "limited evidence" in humans and classified red meat in Group 2A, drawing on narrative text, an evidence table (Table 2.2.1) and published meta-analyses. We checked what both reports printed against the primary papers.

Methods. Each cohort study cited for red meat and colorectal cancer was matched to a double-checked database of 10,383 estimates from the primary papers; a value matched when point estimate and both limits agreed to two decimals. IARC quotations are given with page and line numbers. The 1997 report itself could not be obtained; its values come from search-inside snippets.

Results. Of 178 relative-risk triples in IARC Table 2.2.1, 166 match exactly, 11 belong to studies outside the database and one differs in a confidence limit. The substantive problems lie elsewhere. Singh and Fraser 1998 is printed as colon and rectum with a restricted-subset estimate, 1.90 (1.16–3.11), beside full-cohort case counts; the paper is colon-only and its full-cohort estimate is 1.41 (0.90–2.21). An Alexander 2015 result carries the wrong unit and upper limit, one cohort is tabulated twice, Bernstein 2015's rectal rows are labelled "processed", and Chan 2011 is described as 10 cohorts instead of 8. Every WCRF 1997 number we could check (Willett 1990, Giovannucci 1994, Thun 1992) matches.

Conclusions. The reports transcribed numbers accurately but were less careful in selection, labelling and counting. No error reverses a conclusion. Read against the primary papers, the IARC tabulation supports its own grading, "limited evidence", rather than the common paraphrase that IARC found red meat to cause colorectal cancer.

1. Introduction

The rest of this series re-analyses the quantitative syntheses of red meat and colorectal cancer: Norat 2002, Larsson 2006, WCRF 2007, CUP 2010 and CUP 2016. Two other milestones are not quantitative syntheses and cannot be re-run.

WCRF/AICR 1997. Food, Nutrition and the Prevention of Cancer: a Global Perspective [1] judged that "red meat probably increases risk and processed meat possibly increases risk of colorectal cancer" (p. 246, per [2]). It pooled nothing. Its evidence was a table of seven cohort studies and 26 case-control studies, with narrative commentary.

IARC Monographs volume 114. Red Meat and Processed Meat [3] concluded that "There is limited evidence in humans for the carcinogenicity of consumption of red meat" (§6.1, p. 501, lines 24324–24327). It classified red meat as "probably carcinogenic to humans (Group 2A)" (§6.3, p. 501, lines 24346–24347). The rationale rested on "the substantial epidemiological data showing a positive association between consumption of red meat and cancer of the colorectum, and the strong mechanistic evidence" (§6.4, p. 501, right column, lines 24324–24333). The Working Group ran no meta-analysis. It summarised the cohort evidence in Table 2.2.1 (pp. 145–160), in narrative text (pp. 111–124) and through six published meta-analyses (§2.2.3, pp. 142–144). Its summary in Lancet Oncology [4] is the version most often cited.

Neither report can be re-pooled. Both can be checked in three ways: - whether each printed number matches the primary paper; - whether each study is the one the report says it is, with the label the report gives it; - whether the counts in the text ("seven cohort studies", "fourteen cohort studies", "half of those studies") correspond to the tables.

This article does that and reports what the tables support.

2. Methods

2.1 Sources

IARC 114. We used a two-column text extraction of the monograph, 24,574 lines long [5]. Page numbers are the printed page numbers. Each page's number sits on the page's last line, which we confirmed against the running headers. Line numbers are exact for Table 2.2.1 rows. Text passages interleave the left and right columns, so line numbers for them can be off by about ±5. The Lancet Oncology summary [4] was read from a separate text file.

WCRF 1997. The report was not available. The Internet Archive copy (item foodnutritionp00amer) is lending-only. On 30 October 2026 the full-text and search-inside endpoints returned HTTP 401 or "Item not available", and the details page gave only metadata (678 pp.). All WCRF values below therefore come from a research card compiled on 29 October 2026 from Internet Archive search-inside snippets [2]. Page numbers are the book reader's. They probably match the printed pages, but this was not checked. Table 4.10.10 (cohort studies) could not be confirmed as a whole, and Table 4.10.11 (case-control studies) was not seen at all.

Primary papers. Each study the reports cite was matched to the project database [6]. The database holds 10,383 estimates from 53 publications. Every row was checked against the full text or supplement by an extraction instance that had not built it: 10,353 rows verified, 18 corrected and 12 disputed, with all corrections logged [7]. Where needed, values were also read directly from the primary paper texts (for example Singh and Fraser 1998 [8]) and from the two meta-analyses on disk, Chan 2011 [9] and Alexander 2015 [10].

2.2 Matching rule and classification

• A printed value matches when a database row for the same study has an identical point estimate and identical lower and upper limits after rounding to two decimals.
• A match to a row that is not the study's main model is labelled in_db_nonmain. That signals a model choice, not an error.
• For each report we list:
• studies not in the database (values cannot be checked);
• transcription discrepancies;
• labelling and counting discrepancies;
• internal inconsistencies (table against text);
• interpretive statements that the numbers do not clearly support.
• Severity is graded major where the error changes what a reader would conclude about a study, moderate where it changes a count or label, and minor for clerical errors.

2.3 Who did the work

The comparison tables, the value-by-value check and the verbatim retrieval were done by AI agents (Anthropic Claude Opus instances) under the author's direction. The database they checked against was built and independently controlled in the same way. The scripts and tables are published (Data availability).

3. Results

3.1 WCRF 1997

Section 4.10.2.6 states that "seven cohort studies have examined meat intake" (p. 243). The card gives the report's own tally for Table 4.10.10: of 16 risk estimates, 8 were above 1.5 or statistically significant, 8 lay between 0.75 and 1.5, and none were below 0.75 (p. 243). These 16 include processed meat and adenoma outcomes. The report counts 26 case-control studies (23 on cancer, 3 on polyps).

Table 1. WCRF 1997 cohort evidence (Table 4.10.10, pp. 243–246) against the primary papers

Five of the seven cohorts are in the database. The report prints numbers for three of them (Willett 1990, Giovannucci 1994 and Thun 1992), and all match exactly.

The two "positive" cohorts have three things in common that the card snippets do not show the report mentioning: - both are from the same Harvard group; - both cover colon cancer only; - both are adjusted only for age (and, in NHS, energy).

Two of the "null" verdicts sit over point estimates above 1: Thun's 1.21 for men and Bostick's main-dish 1.21. Neither is significant, so the verdicts are defensible. They are still more categorical than the numbers.

Of the 26 case-control studies, the card names only Manousos 1983 and Miller 1983. The other 24 cannot be identified without Table 4.10.11.

3.2 IARC 114: what the text counts

The inclusion rule (§2.2.1, p. 111, lines 5660–5664) reads: "The most recent publication of a cohort study, or the publication with the highest number of cases in the analysis, was included in the review." Studies with mortality endpoints, with meat definitions including poultry or fish, on dietary patterns, or with fewer than 100 cases were excluded (lines 5664–5673).

The red meat section opens: "Fourteen cohort studies and two cohort consortia provided informative data on the association between red meat and risk of colorectal cancer (see Table 2.2.1)" (p. 111, lines 5660–5663). The fourteen are not listed. Reconstructed from the order of the narrative they are:

NYUWHS, Dutch Monitoring Project, Takayama, AHS, ATBC, IWHS, DCH, SWHS, MCCS, SMC, SCHS, EPIC, NHS/HPFS and MEC.

The two consortia are the UK Dietary Cohort Consortium (UKDCC) and GECCO/CCFR.

The summary (§5.2.1, p. 493, lines 23964–23971) states: "Positive associations between high consumption of red meat and cancer of the colorectum were observed in seven studies, including most of the larger studies: the European Prospective Investigation into Cancer and Nutrition (EPIC) study in 10 European countries, the Swedish Mammography Cohort (SMC), and the Melbourne Collaborative Cohort Study (MCCS)." The seven are not named.

If "positive" means a highest-vs-lowest point estimate clearly above 1 (here ≥ 1.1), exactly seven cohorts qualify: - Kato 1997: 1.23 - Tiemersma 2002: 1.6 - Singh 1998: 1.90 - Bostick 1994: 1.21 - English 2004: 1.4 - Larsson 2005: 1.22 - Norat 2005: 1.17

This reconstruction is ours (judgement call 6 in [15]). On that reading, only Singh's estimate is nominally significant, and only in a restricted subset (Section 3.4). Norat's per-100 g estimate is significant, and English's interval touches 1. A strict "> 1" rule would also admit SCHS (1.01) and Takayama men (1.03).

3.3 IARC 114: Table 2.2.1 against the primary papers

Table 2.2.1 has 23 rows. It has no comments column. Its columns are reference and design, population and exposure assessment, organ site, exposure category, exposed cases, risk estimate and covariates.

Table 2. IARC Table 2.2.1 rows (pp. 145–160) against the database

At the level of individual numbers the table is accurate. Of the 178 RR/CI triples printed in Table 2.2.1: - 166 are identical to a database row; - 11 belong to studies not in the database (Berndt, Sørensen, Andersen, Figueiredo, Ananthakrishnan); - 1 differs: Pietinen's second quartile upper limit.

3.4 IARC 114: discrepancies

Table 3. Discrepancies between IARC 114 and the primary papers, or within IARC 114

Every IARC number for the Chan 2011 meta-analysis matches Chan's table. These are highest vs lowest 1.10 (1.00–1.21) for colorectal cancer, 12 studies, and per 100 g/day 1.17 (1.05–1.31) for colorectal, 1.17 (1.02–1.33) for colon and 1.18 (0.98–1.42) for rectal cancer [9, lines 187–218]. The other three meta-analyses IARC summarises (Pham 2014, Carr 2016, Qiao & Feng 2013) and Alexander 2010 are not on disk. Their figures as IARC prints them are unchecked.

3.5 What could not be checked

• The WCRF 1997 report itself. Every WCRF value comes from snippets on a research card. We could not see Table 4.10.10 as a whole, the exposure label for each row, or any part of Table 4.10.11.
• Two WCRF cohorts (Phillips 1985, Knekt 1994) and five IARC Table 2.2.1 rows (Berndt 2006, Sørensen 2008, Andersen 2009, Figueiredo 2014, Ananthakrishnan 2015). These are not in the database, and their printed values were not compared with the papers. For Sørensen, Andersen and Berndt we could check only internal consistency.
• IARC's case-control evidence. IARC tabulates 55 case-control rows in Table 2.2.3. None is in the database, so the statement that seven of 15 informative case-control studies were positive [4, lines 93–97] could not be checked.
• Text-only studies. Fourteen studies IARC discusses only in the text are listed in our IARC table file [15]. Most are item-level or gene–environment analyses outside the database's scope.
• Line numbers for text passages are approximate (±5 lines) because the columns interleave. Table rows are exact.

4. Discussion

4.1 Accurate numbers, careless selection

Neither report is careless with numbers. Every WCRF figure we could check matches its source. In IARC Table 2.2.1, 166 of 167 checkable triples match to two decimals, and the remaining one differs in a single confidence limit. For a 16-page table this is a good record. It should be said plainly, because the discrepancies listed above could otherwise suggest a general sloppiness the data do not show.

The errors that matter are in what was selected and how it was labelled and counted.

Singh and Fraser 1998. This row is the most consequential. The paper's abstract leads with the subset estimate, "among subjects who favored specific types of meat … RR for >1 time/week vs. no red meat intake = 1.90" [8, lines 19–21]. Choosing that figure is therefore understandable. But IARC prints it as a colorectal result, with full-cohort case counts and covariates, in a table whose other rows are full-cohort estimates. The full-cohort estimate, 1.41 (0.90–2.21), is not significant. Under the "point estimate clearly above 1" reading of the seven positive studies, Singh stays positive either way. The table's only nominally significant highest-vs-lowest red meat result in a cohort, however, comes from a subset analysis.

Alexander 2015 in §2.2.3. A per-week increment printed as per-day understates the per-day magnitude, by roughly a factor of seven on the log scale. The upper limit 1.14 also widens an interval that Alexander prints as 1.00–1.04. In fairness, this error runs against an association rather than for one. Alexander's per-serving result is more positive per day than IARC's text implies.

"Ten cohort studies". Chan 2011's 17 % per 100 g/day came from 8 cohort studies for colorectal cancer. IARC's own meta-analysis section says 8 (p. 143, line 7252). The summary (p. 493) and the Lancet Oncology report [4, line 106] say 10. The sentence in §5.2.1 refers to the meta-analysis "of red meat and/or processed meat". A generous reading takes 10 as a description of the meta-analysis as a whole. Even on that reading, the figure attached to the red meat estimate is wrong. The error does not change the estimate. It slightly inflates the apparent breadth of its base, and the Lancet Oncology version is the one most readers saw.

Labels and duplication. The DCH duplication and the Bernstein "processed" label do not change any count or estimate the Working Group relied on. They do matter to anyone who uses Table 2.2.1 as a data source.

Text and table out of step. The text counts IWHS and UKDCC among the 14 cohorts and 2 consortia, but the table leaves them out and includes seven red-and-processed studies the text calls "less informative". A reader of the table alone sees a different evidence base from a reader of the text. Together with the absence of a list of the 14 cohorts or of the 7 positive studies, the central counts in the monograph cannot be verified from the monograph itself. They can only be reconstructed, as we have done, with some judgement.

4.2 What the tabulations support

WCRF 1997. The cohort evidence behind "probable" was thin by the report's own description. There were seven cohorts. Two showed positive associations, and these were colon-only analyses from one research group adjusted only for age (and energy). The other five were reported as null, and two of those (Phillips, Thun) used mortality endpoints, Phillips with total meat. The report's own tally of 16 estimates, half above 1.5 or significant, counts processed meat and adenomas alongside red meat. A "probable" grading therefore leaned heavily on the 26 case-control studies, which we could not inspect. Read against what we could check, the cohort evidence supports a hypothesis worth pursuing. We see no clear basis in the cohort column for "probable" as distinct from "possible", but the case-control column may have carried that difference.

IARC 114. The Working Group's own reading of its cohort evidence was cautious. It wrote that "chance, bias, or confounding could not be ruled out for consumption of red meat, as no association was observed in several of the larger studies" (§5.2.1, p. 493, lines 23999–24002). The Lancet Oncology summary adds that "residual confounding from other diet and lifestyle risk is difficult to exclude" [4, lines 134–141].

Table 2.2.1, checked against the papers, bears this out: - The unprocessed-red-meat rows from the largest cohorts are null or near-null: MEC 0.98 (0.87–1.10), NHS/HPFS 0.99 (0.87–1.13) per serving/day, SCHS 1.01 (0.82–1.26) and DCH 1.02–1.03 per 25 g/day. - Positive point estimates come from EPIC (per 100 g 1.21, 1.02–1.43), SMC and MCCS, with intervals that include or touch 1 for the highest-vs-lowest contrasts. - Singh's significant result is a subset estimate.

The human evidence was graded "limited", and the 2A classification rests explicitly on that plus "strong mechanistic evidence" (§6.4). The animal evidence was graded "inadequate" (§6.2, p. 501).

The tabulation supports the human-evidence grading IARC actually gave, "limited". It does not support the frequent public paraphrase that IARC found red meat to cause colorectal cancer, which IARC said only of processed meat. The discrepancies we found do not alter this reading. If anything, correcting Singh's row and the "ten cohort" count makes the cohort column marginally weaker than it was presented.

4.3 Being fair to the authors

Several of the "discrepancies" are reasonable choices that are under-documented rather than errors: - IARC's choice of the total-meat-adjusted model for Flood 2003 is defensible, and the covariates column says so. - Bernstein's per-serving estimates are what the paper's main tables print. - WCRF's "no increase" verdicts are fair descriptions of non-significant estimates near 1.

We grade as errors only what departs from the paper (Singh's site and subset, Alexander's unit, Bernstein's label), from the report's own rule (DCH twice), or from the report's own text (the 8-versus-10 count).

4.4 Limitations

• The WCRF 1997 check rests on snippets, not the report.
• Five IARC Table 2.2.1 studies and two WCRF cohorts are outside the database.
• The reconstruction of IARC's 14 cohorts and 7 positive studies is ours. IARC lists neither.
• The extraction and checking were done by AI agents. The database they matched against was independently controlled row by row, but the matching step itself had no second, independent pass.

Data availability

• kontroll/evidens/wcrf1997_cohorts.csv and .md
• kontroll/evidens/iarc114_cohorts.csv and .md, with 37 rows (23 tabulated, 14 text-only), and with page, line, Working Group note and database match for each
• kontroll/evidens/iarc114_table_values_check.csv: all 178 triples
• kontroll/evidens/NOTES.md: judgement calls and sources not obtained
• Database: database/estimates.csv, studies.csv, with the correction log kontroll/rettelogg.md

The IARC text file is Historikk/txt/36_IARC114.txt.

References

1. World Cancer Research Fund / American Institute for Cancer Research. Food, Nutrition and the Prevention of Cancer: a Global Perspective. Washington DC: AICR, 1997.
2. Research card Historikk 1965-2015/kort/25_WCRF1997.md (Internet Archive search-inside snippets, read 29 October 2026); also kontroll/fase0_inventar.md §1a.
3. IARC Working Group on the Evaluation of Carcinogenic Risks to Humans. Red Meat and Processed Meat. IARC Monographs on the Evaluation of Carcinogenic Risks to Humans, vol. 114. Lyon: International Agency for Research on Cancer, 2018.
4. Bouvard V, Loomis D, Guyton KZ, et al. Carcinogenicity of consumption of red and processed meat. Lancet Oncol 2015;16:1599–1600.
5. Text extraction of [3]: Historikk 1965-2015/Historikk/txt/36_IARC114.txt (24,574 lines).
6. Red meat and colorectal cancer database, version 3, 30 October 2026 (database-samledom.md).
7. Correction log (kontroll/rettelogg.md).
8. Singh PN, Fraser GE. Dietary risk factors for colon cancer in a low-risk population. Am J Epidemiol 1998;148:761–774.
9. 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:e20456.
10. Alexander DD, Weed DL, Miller PE, Mohamed MA. Red meat and colorectal cancer: a quantitative update on the state of the epidemiologic science. J Am Coll Nutr 2015;34:521–543.
11. Willett WC, Stampfer MJ, Colditz GA, Rosner BA, Speizer FE. Relation of meat, fat, and fiber intake to the risk of colon cancer in a prospective study among women. N Engl J Med 1990;323:1664–1672.
12. Giovannucci E, Rimm EB, Stampfer MJ, Colditz GA, Ascherio A, Willett WC. Intake of fat, meat, and fiber in relation to risk of colon cancer in men. Cancer Res 1994;54:2390–2397.
13. Goldbohm RA, van den Brandt PA, van 't Veer P, et al. A prospective cohort study on the relation between meat consumption and the risk of colon cancer. Cancer Res 1994;54:718–723.
14. Bostick RM, Potter JD, Kushi LH, et al. Sugar, meat, and fat intake, and non-dietary risk factors for colon cancer incidence in Iowa women. Cancer Causes Control 1994;5:38–52.
15. Evidence-table notes and files (kontroll/evidens/NOTES.md, iarc114_cohorts.md).

Primary studies in IARC Table 2.2.1 (identifiers as recorded in the IARC reference list or the database): Kato 1997 (doi:10.1080/01635589709514588); Chen 1998 (Cancer Res 1998;58:3307–11); Pietinen 1999 (Cancer Causes Control 1999;10:387–96); Järvinen 2001 (doi:10.1054/bjoc.2001.1906); Tiemersma 2002 (doi:10.1023/a:1015236701054); Flood 2003 (doi:10.1093/aje/kwg099); English 2004 (Cancer Epidemiol Biomarkers Prev 2004;13:1509–14); Chao 2005 (doi:10.1001/jama.293.2.172); Larsson 2005 (doi:10.1002/ijc.20658); Norat 2005 (doi:10.1093/jnci/dji164); Berndt 2006 (doi:10.1158/1055-9965.EPI-06-0449); Oba 2006 (doi:10.1016/j.canlet.2005.12.037); Kabat 2007 (doi:10.1038/sj.bjc.6603837); Butler 2008 (doi:10.1038/sj.bjc.6604678); Sørensen 2008 (doi:10.1016/j.canlet.2008.02.046); Andersen 2009 (doi:10.1186/1471-2407-9-407); Lee 2009 (doi:10.1080/01635580802419780); Cross 2010 (doi:10.1158/0008-5472.CAN-09-3929); Ollberding 2012 (doi:10.1002/ijc.27546); Figueiredo 2014 (doi:10.1371/journal.pgen.1004228); Ananthakrishnan 2015 (doi:10.1158/1055-9965.EPI-14-0897); Bernstein 2015 (doi:10.1371/journal.pone.0135959); Wei 2004 (doi:10.1002/ijc.11540); Spencer 2010 (doi:10.1007/s10552-010-9569-7). WCRF 1997 cohort in the database since R65: Thun et al., J Natl Cancer Inst 1992;84:1491–500. WCRF 1997 cohorts not in the database: Phillips & Snowdon, J Natl Cancer Inst 1985;74:307–17; Knekt et al., Int J Cancer 1994;59:756–60.

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.