← Tilbake til oversikten (norsk)

Red meat and colorectal cancer in the WCRF/AICR Second Expert Report (2007): a reproduction and re-analysis of the cohort meta-analysis

Re-analysis series on the red meat – colorectal cancer syntheses, 2002–2016. Synthesis identifier: wcrf2007.

Abstract

Background. The 2007 WCRF/AICR Second Expert Report judged red meat a "convincing" cause of colorectal cancer. Its quantitative support was a random-effects meta-analysis of cohort studies: relative risk (RR) 1.43 (95 % CI 1.05–1.94) "per time/week" (seven studies) and 1.29 (1.04–1.60) per 100 g/day (three studies). We tested whether these summaries can be reproduced from the printed per-study values, and how those values compare with the primary papers.

Methods. Following a protocol fixed in advance, we re-pooled the nine rows of Figure 4.3.2 by DerSimonian–Laird random effects, compared each row with its primary paper, classified the deviations, and re-ran the analysis with each class corrected, all corrected, and a later publication substituted, with leave-one-out throughout.

Results. The summary reproduces exactly: 1.43 (1.05–1.94). Its unit is per time/day, not per time/week as printed. One study (Chen 1998) is entered twice as genotype strata. The Nurses' Health Study and Health Professionals Follow-up Study enter through their earliest colon-only reports (Willett 1990; Giovannucci 1994), although the report cites their later combined publication (Wei 2004). Combining the Chen strata gives 1.41 (1.04–1.90); removing the two superseded reports gives 1.14 (0.88–1.49); both corrections give 1.10 (0.89–1.36). Adding the cohorts' 2015 follow-up gives 1.02 (0.92–1.13). The per 100 g/day summary, printed in three versions, reproduces once a misprinted confidence limit is read off the plot: random effects gives the text's 1.29 (1.04–1.60), fixed effect the figure's 1.29 (1.05–1.59).

Conclusions. The published estimate is arithmetically reproducible, but its statistical significance rests on the earliest, smallest publications of two cohorts that the report itself cites in updated form. Without them the summary is 1.14 (0.88–1.49), and 1.10 (0.89–1.36) when the Chen strata are also combined; both intervals include 1. The corrected runs lean heavily on one row (Bostick 1994) whose printed interval is itself doubtful. The unavailability of the 2005 systematic literature review prevents a full audit.

1. Introduction

The WCRF/AICR Second Expert Report, Food, Nutrition, Physical Activity, and the Prevention of Cancer: a Global Perspective (2007) [1], concluded:

"A substantial amount of data from cohort and case-control studies showed a dose-response relationship, supported by evidence for plausible mechanisms operating in humans. Red meat is a convincing cause of colorectal cancer." (p121)

The quantitative basis is section 4.3.5.1.1 (pp120–121). It reports a highest-versus-lowest forest plot without a summary (Fig 4.3.1). It reports two dose-response meta-analyses of cohort studies, split by the unit in which the studies measured intake:

• 1.43 (1.05–1.94) "per times/week", seven studies (Fig 4.3.2);
• 1.29 (1.04–1.60) per 100 g/day, three studies (Fig 4.3.3 and text, p121).

It also shows a dose-response graph without values (Fig 4.3.4).

Case-control studies were "not summarised" because of "the abundant prospective data from cohort studies" (p121).

The 2007 judgement became the baseline of the Continuous Update Project (CUP). The 2010 CUP systematic literature review (SLR) [2] compared its new estimate against the 2007 figures (SLR 2010 p156, Table 36 p157). The CUP Colorectal Cancer 2011 Report [3] repeated the "convincing" judgement. Chan et al. 2011 [4], the peer-reviewed version of the 2010 SLR, opens by citing the 2007 judgement as its starting point (Chan 2011, abstract, p1).

The 2007 report also cites the meta-analysis by Larsson and Wolk [5], 1.28 (1.18–1.39) per 120 g/day, as supporting evidence (p121). That paper is re-analysed separately in this series.

The 2007 analysis therefore sits at the root of the subsequent WCRF evidence chain. Because the underlying earlier SLR (labelled the "2005 SLR" in SLR 2010 Tables 35–36) is not available in our files, the printed forest plots are the only record of what was pooled. This paper asks three questions:

1. Can the printed summaries be reproduced from the printed per-study values?
2. Do those values correspond to what the primary papers report?
3. How sensitive is the summary to the deviations found?

2. Methods

2.1 Protocol

The analysis follows a written protocol, version 1.0, dated 30 October 2026 and fixed before the re-analyses were run (REANALYSE_PROTOKOLL.md). The protocol's rules R1–R7 govern model choice, study selection, reproduction criteria, deviation classes and corrected runs. For WCRF 2007, the protocol states two limits in advance:

• only the cohort stratum is reproduced, because the report did not pool case-control studies;
• the 2005 SLR (the protocol calls it the 2006 SLR) is not on disk, so the description of it in the 2010 SLR is used instead (protocol §7).

2.2 Data

The synthesis. We used the report's text and figures (SER pp120–121) and the chapter 4.3 reference list (refs 8–24, p436). The values were first extracted from a text version of the report (32_WCRF2007.txt). The author later supplied PDF pages 116–125, against which every value in Figs 4.3.1–4.3.3 was confirmed (correction log R66).

The SLR 2010 description of the earlier review. We used the 2010 SLR's account of how the earlier review used each study: Table 35 (p157), Table 36 (p157) and the "SLR" column of Table 37 (pp159–165) [2].

The nine rows of Fig 4.3.2. The figure lists:

The report's reference list also contains Wei et al. 2004 [13] (ref 19), the combined NHS + HPFS publication.

Fig 4.3.3 contains three studies: Pietinen 1999 [14], Norat 2005 [15] and Larsson 2005 [16].

Extraction files. The printed per-study values are in database/synthesis_extraction.csv (27 wcrf2007 rows). The primary-paper estimates are in database/estimates.csv, with per-100 g conversions in database/derived_per100g.csv.

2.3 The synthesis's own method, as stated

• Model. "Unless indicated otherwise, random effects models … were used to generate the forest plots presented in this Report" (Box 3.3, p53).
• Outcome. Colorectal cancer or, "if not available, colon cancer". This rule is stated in the 2010 SLR's description of the earlier review (SLR 2010 p156; Table 35 p157).
• Analyses by unit. Analyses were "run separated by exposure measurement unit (gram or time/day)" (SLR 2010 p156). The 2007 report describes the same split as "seven studies that measured red meat intake in 'times per week' and three studies that measured grams per day" (p121).
• Not stated in our files. We could not find in the available documents how frequencies were converted, which category rule was used, or which publication was chosen when a cohort had several. These rules would be in the 2005 SLR.

2.4 Reproduction rule

The rule follows protocol R1–R3:

• We used the printed per-study RR and 95 % CI exactly as in Fig 4.3.2.
• The log standard error was derived from the CI width.
• We pooled by DerSimonian–Laird [17].
• Reproduction is declared when the pooled point estimate and both confidence limits match the printed values to two decimals.

2.5 Deviation classification and corrected runs

Each row was compared with its primary paper. Differences were classified under protocol R5 as exposure, dose, cases, model, duplicate, transcription or inclusion.

• A deviation means that the row differs from what the primary paper prints, or breaks a rule the synthesis states.
• An assumption is a defensible analytic choice, such as a portion size. We report assumptions but do not treat them as errors.

Under R6 we re-ran the pooling once per correctable class, then with all classes corrected together. Two classes were corrected:

• duplicate. The two Chen 1998 genotype rows were combined into one study estimate by fixed-effect inverse-variance weighting: 1.14 (0.63–2.05).
• inclusion. The rows from Willett 1990 and Giovannucci 1994 were removed. A replacement from Wei 2004 (1,139 colon + 339 rectal cases) was considered and rejected. Its only continuous estimate, 1.12 (0.99–1.27) for beef, pork or lamb (Table III, p441), is labelled "one additional serving per week", but the paper's Methods (p435) state that meat was modelled in servings per day. The database marks the row disputed, and the correction log (R15) excludes it from pooled analyses.

Deviations without a usable replacement are documented but not corrected. These are the unit of the analysis (transcription, English 2004 row; §3.4), Singh 1998 (dose), Bostick 1994 (transcription) and the nested case-control design of Chen 1998 (inclusion).

Increment. Following §3.4, all runs treat the increment as one time (serving) per day, and the run names say so ("Per time/day – …"). The printed per-study values are used unchanged; only the interpretation of the scale, and the rescaling of the follow-up row, depend on this.

Follow-up run. To the all-corrections set we added the latest NHS/HPFS publication in our database, Bernstein et al. 2015 [18], in place of Willett 1990 and Giovannucci 1994. Its estimate for unprocessed red meat is 1.00 (0.89–1.11) per 100 g/day (Suppl. Table S5; 2,731 cases). We rescaled it to one 120 g serving per day by raising the RR and its limits to the power 1.2, giving 1.00 (0.87–1.13).

2.6 Software

All runs come from scripts/reanalyses.py, which calls scripts/pool.py (DerSimonian–Laird and inverse-variance fixed effect; I² [19]; leave-one-out; a cohort-duplicate check, relaxed only for the published-value run). Per-100 g conversions come from scripts/derive_per100.py, using Greenland–Longnecker [20] as implemented by Orsini et al. [21] (scripts/gl_doseresponse.py) where only categories exist; for WCRF 2007 this method enters no pooled run and serves only to judge the Bostick row.

An independent controller (K11) recomputed the pooled values with a separately written script, kontroll/meta_check.py, without access to pool.py.

2.7 Extraction and control

All data extraction and checking was carried out by AI agents (Anthropic Claude, Opus models) working under the author's direction, to a written codebook (KODEBOK.md).

• Primary-paper database. Every row in estimates.csv was checked against the full text by an independent agent instance that had not seen the extraction. Corrections are logged in kontroll/rettelogg.md.
• Synthesis extraction. The extraction of the WCRF 2007 figures was checked line by line against the printed pages by a separate controller (K11, kontroll/K11/K11_rapport.md). It found all 24 numeric wcrf2007 rows to match Figs 4.3.1–4.3.3 as printed.
• Page images. The PDF pages 116–125 supplied by the author confirm the extracted values of Figs 4.3.1–4.3.3, the axis label of Fig 4.3.2 ("Relative risk, per time/week"), and that the misprint "1.43 (1.05–0.95)" for Larsson 2005 in Fig 4.3.3 is in the report itself. Fig 4.3.4 is a line graph with no printed values (R66).
• Figure reading. Larsson 2005's upper limit in Fig 4.3.3 was read off the plotted interval by calibrated pixel measurement: log axis, 106.5 px per natural-log unit, calibrated on the printed ticks 0.2–5. As a check, the readings for Pietinen and Norat were within 0.02 of their printed values. The reading gives ≈ 1.94. This is the only figure-read value in any run, and it is labelled as such.

3. Results

3.1 Reproduction

Pooling the nine printed rows of Fig 4.3.2 gives 1.43 (1.05–1.94), I² 58 %, k = 9 (run: Per time/day – published per-study values).

This matches the published summary and the I² of 58 % that SLR 2010 attributes to it (Table 36, p157); the fixed-effect estimate is 1.15 (1.00–1.32). K11's script gives the same (Q 19.00, τ² 0.092), and all nine inputs are confirmed against the report's PDF pages (R66).

The report prints no weights. Our largest random-effects weights are:

The gap between fixed (1.15) and random effects (1.43) arises because heterogeneity flattens the weights, shifting influence from the precise Bostick and English rows to imprecise, strongly positive ones (Singh, Chen rapid, Tiemersma men).

3.2 Table 1. Per-study inputs: as printed vs the primary papers

Sources: SER Fig 4.3.2 (p121), confirmed against the PDF page. Primary values are from database/estimates.csv and database/studies.csv.

Four of the seven studies are colon-only, as the stated outcome rule permits. Two are nested case-control studies within cohorts.

3.3 Deviations

Seven deviations were recorded (wcrf2007.md, deviation table). The unit deviation (English 2004, transcription) is treated in §3.4; the others follow.

D1 – Chen 1998, duplicate. One study is entered as two rows, split by acetylator genotype ("NAT rapid", "NAT slow"; Fig 4.3.2). The strata do not share participants, but the split gives one study two places in the random-effects weighting (the text's "seven studies" are nine rows). Replacement: the strata combined by fixed effect, 1.14 (0.63–2.05).

D2 – Chen 1998, inclusion (documented only). The Physicians' Health Study analysis is a nested case-control study. The protocol admits nested designs, so this is recorded but not corrected. The 2010 SLR later moved Chen 1998 to "red and processed meat" (SLR 2010 p156), consistent with its exposure definition, which includes hot dogs.

D3 – Willett 1990 and Giovannucci 1994, inclusion.

• What the rows are. NHS and HPFS enter the analysis through their first reports. Willett 1990 covers NHS 1980–86 with 150 colon cancers. Giovannucci 1994 covers HPFS 1986–92 with 205 colon cancers.
• What the report also cites. The reference list includes Wei et al. 2004 (ref 19): NHS and HPFS combined, with 1,139 colon and 339 rectal cancers. The report uses it only in the highest-versus-lowest plot, labelled "Wei 2003 Men/Women" (Fig 4.3.1; SLR 2010 Table 37, "SLR" column, pp159–165).
• How CUP later handled it. The 2010 SLR states: "the studies by Willett, W., et al., 1990 and Giovannucci, E., et al., 1994 were replaced by the most recent publication by Wei, E.K. et al., 2004" (SLR 2010 p156).
• Classification. Inclusion deviation. Without the 2005 SLR we cannot say whether it broke a rule, or why the choice was made.
• Correction. Both rows removed. Wei 2004's continuous estimate cannot substitute, because its unit is disputed (§2.5).

D4 – Singh 1998, dose (documented only). The value 4.51 (0.38–53.27) is a slope estimated from a three-level scale (never / <1 / ≥1 time per week) whose top category is open (Singh 1998, Table 3). It depends on the unknown dose assigned to "≥1 time per week"; the row carries 1.5 % of the weight.

D5 – Bostick 1994, transcription (documented only).

• The printed CI, 0.80–1.14, spans 0.35 on the log scale. That is four to five times narrower than a Greenland–Longnecker slope from Bostick's categories can yield (212 cases; per 100 g/day 0.94 (0.43–2.04), log width 1.56; derived_per100g.csv).
• With 24.4 % of the weight in the published run, and 56.1 % once the superseded NHS/HPFS rows are removed, this row is the largest single counterweight to the positive rows.
• We could not determine what the 2005 SLR computed. The row is therefore left unchanged in all runs.

D6 – Fig 4.3.3, per 100 g/day (see §3.7).

• The Larsson 2005 row is printed "1.43 (1.05–0.95)", an upper limit below the point estimate. The PDF shows that the misprint is in the report, not in our extraction.
• The summary is printed in three versions:
• 1.29 (1.05–1.59) in the figure;
• 1.29 (1.04–1.60) in the text (p121);
• 1.29 (0.94–1.78), I² 32 %, in SLR 2010 Table 36 (p157).
• The Norat 2005 row, 1.49 (0.91–2.43), is EPIC's measurement-error-calibrated estimate (Norat 2005 Table 4, p912). The 2010 SLR later chose the observed estimate instead (p156).

3.4 The unit of the analysis: per time/week or per time/day?

The figure axis and the text label the summary "per time/week" and "per times/week" (p121). The PDF confirms the axis label as printed: "Relative risk, per time/week". The 2010 SLR, describing the same analysis, labels it "per time/day" twice:

• in the text: "1.43 (95% CI = 1.05-1.94, seven studies) per time/day increase" (p156);
• in Table 36 (p157).

The 2010 SLR also prints sex strata of this analysis "per time/day": men 1.79 (0.98–3.27), women 1.27 (0.74–2.19).

The one row that can be checked arithmetically favours the SLR's label. English 2004 prints fresh red meat per one time/week as 1.03 (0.98–1.08) (Table 3, p1512). Raising this to the seventh power, which converts it to one time per day, gives 1.23 (0.87–1.71) (our arithmetic). Fig 4.3.2 prints 1.23 (0.88–1.73), which is consistent within rounding of the published inputs.

Two other rows give the same impression. Willett 1990's 1.81 and Giovannucci 1994's 2.20 are hard to reconcile with a per-weekly-serving slope, given their categorical results (≥1/day 2.49; ≥5/week 3.57). As per-daily-serving slopes they are plausible.

We therefore conclude that the per-study values, and so the summary, are per time/day, and that the figure and text label is wrong. The printed label itself is not in doubt, since the PDF confirms it. The conclusion is therefore an inference against a confirmed printed label. The deviation table records this as a transcription deviation on the English 2004 row (SER Fig 4.3.2 p121; SLR 2010 pp156–157; English 2004 Table 3). The 2005 SLR would settle the question; without it, the conclusion rests on one exact arithmetic check, two plausibility arguments and the SLR 2010 label. No printed value is altered by it. It determines how the 1.43 should be read, and how a later per-gram estimate must be rescaled to join the analysis (§3.6).

3.5 Table 2. Published, reproduced, corrected and follow-up estimates

* I² as given in SLR 2010 Table 36 (p157). The 2007 report says "moderate" and "low" heterogeneity respectively.

In the per 100 g/day run, one input (Larsson's upper limit) is read off the plot rather than printed; all other inputs are printed values.

3.6 Corrected and follow-up runs

Duplicate. Combining the two Chen strata changes little: 1.41 (1.04–1.90). Leave-one-out in this run shows that significance depends on each of the two early US colon reports:

Inclusion. Removing Willett 1990 and Giovannucci 1994 gives 1.14 (0.88–1.49), k = 7, with I² falling from 58 % to 28 %. The two rows were the most strongly positive of the precise rows; without them, Bostick 1994 (47.1 %) and English 2004 (30.2 %) carry over three-quarters of the weight.

All corrections. 1.10 (0.89–1.36), I² 17 %, k = 6. Bostick 1994 now carries 56.1 % of the weight and English 2004 28.0 %. Leave-one-out ranges from 1.03 (0.88–1.20) without Tiemersma men to 1.28 (0.97–1.70) without Bostick 1994. No single omission gives a lower limit above 1. The corrected estimate therefore depends heavily on the Bostick row, whose printed interval we question (D5). If its interval is too narrow, the row is overweighted and the corrected estimate would probably be higher than 1.10; we cannot say by how much.

Follow-up run. Adding the NHS/HPFS follow-up, Bernstein 2015, rescaled to one 120 g serving per day (1.00 (0.87–1.13)), gives 1.02 (0.92–1.13), I² 3 %, k = 7. The Bernstein row carries 54.1 % of the weight and Bostick 32.3 %; together they hold 86 %, so the pooled value lies close to these two near-null rows. Without the Bernstein row the result is the all-corrections value, 1.10 (0.89–1.36); without Bostick it is 1.09 (0.91–1.29). The run is driven by these two rows, and above all by the Bernstein row.

3.7 The per 100 g/day analysis

The three-study analysis cannot be reproduced from printed numbers alone, because Larsson 2005's upper limit is misprinted as "1.05–0.95" in the report itself.

Reading the limit off the plot. The plotted interval gives an upper limit of ≈ 1.94 (calibration in §2.7). With it, the run Per 100 g/day – Fig 4.3.3 (3 studies; Larsson's upper limit read off the figure as 1.94) gives:

• random effects: 1.29 (1.04–1.60), the text's summary (p121);
• fixed effect: 1.29 (1.05–1.59), the figure's summary;
• I² 6 %, consistent with the report's "low heterogeneity".

The weights are:

Leave-one-out.

What remains unexplained. The third published version, 1.29 (0.94–1.78) with I² 32 % (SLR 2010 Table 36), is not reproduced by either model. We cannot explain it.

Prior evidence. Before the PDF was available, K11 imputed the upper limit symmetrically on the log scale (1.43²/1.05 = 1.95) and obtained the same two summaries (K11 §3). The figure reading, 1.94, confirms that inference.

Interpretation. The figure thus appears to print a fixed-effect summary while the text prints the random-effects one. If so, the figure departs from the report's stated default model ("Unless indicated otherwise, random effects models … were used to generate the forest plots", Box 3.3, p53). The difference lies only in the confidence interval.

3.8 The highest-versus-lowest plot (Fig 4.3.1)

Fig 4.3.1 has no summary, but it contains inconsistencies (synthesis_extraction_NOTES.md §3; K11):

• Study count. The text says "Sixteen cohort studies" but cites 17 references (8–24).
• Phillips 1975 [22]. Plotted at 2.70 (1.60–4.55). The 2010 SLR calls it a case-control study and dropped it (Table 35 footnote, p157).
• "Chen 2003". Plotted at 1.48 (0.85–2.59). It has no entry in the chapter's reference list (ref 20 is Chen 1998). Its value equals a Jiashan (China) pork estimate [23] listed in Chan 2011's Table S2 [4].
• "Slattery 2001 Women". Has no reference in the chapter.
• Claim of uniformity. The text states that "all of the cohort studies that reported analyses of risk for the highest intake group when compared to the lowest showed increased risk" (p120). The 2010 SLR's summary of the same earlier review (Table 35, p157) says that nine of eleven studies showed increased risk and one no association, but that cell, in the red-meat table, refers to the "processed meat comparison", so the count may not concern red meat.
• Label years. "Wei 2003" and "Norat 2004" do not match the cited papers.

None affects a pooled number.

3.9 Figure

Figure 1. Forest plots of each run (published, corrected per class, all classes, follow-ups), with random-effects weights. See the interactive page kontroll/reanalyser/html/index.html, tab "WCRF 2007" (#forest-wcrf2007; tables #tbl-wcrf2007, #loo-wcrf2007, #dev-wcrf2007).

4. Discussion

4.1 What the deviations are

The WCRF 2007 dose-response summary is correctly computed, but its inputs represent the two largest US cohorts by their first, colon-only reports (150 and 205 cases), although the report cites Wei 2004 (1,139 colon and 339 rectal cases).

Once the superseded reports are removed, the summary falls from 1.43 to 1.14, and to 1.10 with the Chen strata combined; the confidence interval includes 1. The CUP team replaced the same two reports with Wei 2004 three years later and said so (SLR 2010 p156). Our corrected run reconstructs, in the 2007 framework, the first half of a change WCRF itself accepted. The second half, a per-day estimate from Wei 2004, could not be added because its unit is disputed.

Colon-only studies follow the stated outcome rule, Tiemersma's sex strata do not double count, and nested case-control studies are commonly admitted; these are defensible choices. The Chen genotype split is a duplication under protocol R5, but correcting it moves the estimate only to 1.41.

4.2 Errors, ambiguities and what cannot be decided

Four problems cannot be resolved from the printed record:

1. Bostick's narrow interval. The row is the most heavily weighted in the published run and pulls the estimate towards null. If it is mis-transcribed, the published summary would be higher, not lower. This underlines that we are not hunting only for errors that favour one direction.
2. The unit. The report says per time/week, the 2010 SLR says per time/day, and the one checkable row matches per day. We treat the 2007 label as the error. It changes how the 1.43 should be read: a 43 % higher risk per daily serving (about 120 g/day by the later CUP convention) is a much weaker claim per gram than a 43 % higher risk per weekly serving.
3. The per 100 g/day summary. It exists in three versions. With the misprinted Larsson limit read off the plot (1.94), the figure and text versions reproduce as fixed and random effects respectively. The SLR 2010 version, 1.29 (0.94–1.78), does not reproduce.
4. The 2005 SLR. Without it we cannot see the conversion rules, the publication-selection rule or the per-study derivations. We cannot say whether the choice of Willett and Giovannucci broke a rule or followed one (for example, a preference for papers reporting enough categories to derive a slope).

4.3 What the corrected and follow-up results mean

The all-corrections estimate, 1.10 (0.89–1.36), is not evidence of no association. It is less precise than the published estimate and centred on a smaller effect. What it shows is that the statistical significance of the 2007 dose-response summary depended on two superseded reports that WCRF itself later replaced.

The corrected estimate has two caveats. First, removing the two reports leaves NHS and HPFS unrepresented, because Wei 2004's continuous estimate cannot be used: its Table III labels the unit "one additional serving per week", while its Methods (p435) say servings per day (database status disputed; correction log R15). Second, the remaining set is dominated by Bostick 1994, whose interval is implausibly narrow (D5). Both caveats cut in different directions, and neither can be resolved without the 2005 SLR or a usable Wei estimate.

The follow-up run, 1.02 (0.92–1.13), restores NHS/HPFS through Bernstein 2015 on the per-day scale. It is the most complete version of the 2007 analysis we can construct, but it should not be read as an updated meta-analysis: only one cohort pair is updated, and the result is driven by two rows (Bernstein 54.1 %, Bostick 32.3 %). Its value lies in showing that the cohorts whose early reports carried the published significance give a near-null per-day estimate in their later follow-up.

4.4 Limitations

• The 2005 SLR is unavailable. Per-study weights, case numbers and derivation rules are therefore unknown, and several rows can only be documented, not corrected.
• Partial page images. Only PDF pages 116–125 were available. Other parts of the report, such as Box 3.3 and the references, were read from the text version. Fig 4.3.4 has no values and was not used.
• One figure-read input. The per 100 g/day run uses one value read off the plot (Larsson's upper limit, ≈ 1.94). The calibration reproduces the two printed neighbours within 0.02, and the run reproduces two of the three published summaries. Nevertheless, the value is a measurement, not a printed number.
• Inferred unit. The per-day interpretation (§3.4) contradicts the report's printed and PDF-confirmed label. It rests on one arithmetic check, two plausibility arguments and the SLR 2010 label. The runs and the follow-up rescaling adopt it.
• AI extraction. The primary-paper database, the synthesis extraction and the controls were produced by AI agents. Every value used in a pooled run was independently verified against the printed source; the one disputed primary estimate (Wei 2004) is not used. Nevertheless, residual extraction error cannot be excluded, and all inputs are published for checking.
• Scope. The analysis concerns how the 2007 synthesis was computed. It does not address confounding, measurement error in dietary assessment, or causality.

5. Conclusions

The WCRF/AICR 2007 dose-response summary for red meat and colorectal cancer in cohort studies, 1.43 (1.05–1.94), is reproducible from the nine printed rows. Its unit is per time/day, not per time/week as printed. The rows, however, include one study split into two and represent NHS and HPFS by superseded, colon-only reports, although the report cites the later combined publication.

With these corrected, the estimate is 1.10 (0.89–1.36), and 1.02 (0.92–1.13) when NHS and HPFS are represented by their 2015 follow-up. The per 100 g/day summary exists in three versions. Two of them reproduce once the misprinted Larsson limit is read off the plot: the text's version by random effects and the figure's by fixed effect. The SLR 2010 version does not reproduce.

6. Data and code availability

All inputs, scripts and control reports are published alongside this article:

References

1. World Cancer Research Fund / American Institute for Cancer Research. Food, Nutrition, Physical Activity, and the Prevention of Cancer: a Global Perspective. Washington DC: AICR; 2007. (Section 4.3.5.1.1, pp120–121; Box 3.3, p53; chapter 4.3 references, p436.)
2. 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; 2010. (Section 2.5.1.3, pp155–177.)
3. World Cancer Research Fund / American Institute for Cancer Research. Continuous Update Project Report. Food, Nutrition, Physical Activity, and the Prevention of Colorectal Cancer. 2011.
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.
5. Larsson SC, Wolk A. Meat consumption and risk of colorectal cancer: a meta-analysis of prospective studies. Int J Cancer 2006;119:2657–64.
6. 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;323:1664–72.
7. Bostick RM, Potter JD, Kushi LH, Sellers TA, Steinmetz KA, et al. 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.
8. 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;54:2390–7.
9. Singh PN, Fraser GE. Dietary risk factors for colon cancer in a low-risk population. Am J Epidemiol 1998;148:761–74.
10. Chen J, Stampfer MJ, Hough HL, Garcia-Closas M, Willett WC, et al. A prospective study of N-acetyltransferase genotype, red meat intake, and risk of colorectal cancer. Cancer Res 1998;58:3307–11.
11. 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;13:383–93.
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;13:1509–14.
13. 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;108:433–42.
14. Pietinen P, Malila N, Virtanen M, Hartman TJ, Tangrea JA, et al. Diet and risk of colorectal cancer in a cohort of Finnish men. Cancer Causes Control 1999;10:387–96.
15. Norat T, Bingham S, Ferrari P, Slimani N, Jenab M, et al. Meat, fish, and colorectal cancer risk: the European Prospective Investigation into Cancer and Nutrition. J Natl Cancer Inst 2005;97:906–16.
16. 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;113:829–34.
17. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials 1986;7:177–88.
18. Bernstein AM, Song M, Zhang X, Pan A, Wang M, Fuchs CS, 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.
19. Higgins JPT, Thompson SG. Quantifying heterogeneity in a meta-analysis. Stat Med 2002;21:1539–58.
20. Greenland S, Longnecker MP. Methods for trend estimation from summarized dose-response data, with applications to meta-analysis. Am J Epidemiol 1992;135:1301–9.
21. Orsini N, Bellocco R, Greenland S. Generalized least squares for trend estimation of summarized dose-response data. Stata J 2006;6:40–57.
22. Phillips RL. Role of life-style and dietary habits in risk of cancer among Seventh-day Adventists. Cancer Res 1975;35:3513–22.
23. Chen K, Cai J, Liu XY, Ma XY, Yao KY, et al. Nested case-control study on the risk factors of colorectal cancer. World J Gastroenterol 2003;9:99–103.

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.