← Tilbake til oversikten (norsk)

Reproducing the cohort stratum of Norat et al. (2002) on red meat and colorectal cancer: the published estimate depends on one study entry that the paper did not print

A re-analysis under a pre-specified protocol

Abstract

Background. Norat et al. (2002) published a dose–response meta-analysis of meat and colorectal cancer covering case–control and cohort studies. Its red-meat cohort stratum gave a highest-versus-lowest relative risk (RR) of 1.27 (95 % CI 1.11–1.45; 9 studies) and an RR per 120 g/day of 1.22 (1.05–1.41; 9 studies). We tried to reproduce both.

Methods. Under a pre-specified protocol, we pooled the per-study values Norat printed (Appendix III; Fig 2) by DerSimonian–Laird random effects and checked each against the primary paper. Deviations were classified and corrected one class at a time and together; later follow-ups were substituted. AI agents extracted; independent AI controllers checked.

Results. The eight Fig 2 cohorts with a printed confidence interval pool to 1.38 (1.16–1.64), which does not reproduce the published estimate. The ninth, Thun 1992, is plotted with an interval that neither Norat nor Thun prints; two independent readings put it at about 1.14 (0.93–1.39). Adding it gives 1.27 (1.11–1.45), matching the published estimate; it carries 43.5 % of the weight. With the calibrated high-resolution readings of Thun (0.91–1.38) and Singh (1.49) instead, the pool is 1.28 (1.12–1.46), within 0.01 of the published value. Thun prints only crude "fatty meats" estimates without intervals, so this entry is Norat's own, undocumented construction. Fig 2 also plots Singh 1998 at 1.49, not the tabulated 1.41. The seven Appendix III values we could check match the primary papers. The per-120 g estimate could not be reproduced (no slopes or study list printed). Substituting ATBC's unprocessed row and removing the two mortality cohorts gives 1.27 (1.02–1.59); adding later follow-ups of the two Harvard cohorts gives 1.01 (0.88–1.16).

Conclusions. The printed cohort highest-versus-lowest estimate is correct arithmetic, but it can be verified only by digitising a figure, and its largest input cannot be traced to the primary paper. The dose–response estimate cannot be verified. The corrected and follow-up results depend on the red-meat definition and on which publication of each cohort is used.

1. Introduction

Norat et al. (2002) [1] pooled 34 case–control and 14 cohort studies published in English between 1973 and 1999 (p. 243). For red meat, all designs combined gave an RR of 1.35 (1.21–1.51) for the highest versus lowest category (Table I, p. 244). The figure per 120 g/day was 1.24 (1.08–1.41) (Table II, p. 244). The cohort stratum was homogeneous. For highest versus lowest it gave 1.27 (1.11–1.45), n = 9, p-heterogeneity 0.45. Per 120 g/day it gave 1.22 (1.05–1.41), n = 9, p-heterogeneity 0.17 (Tables I–II, p. 244).

The paper shaped the later syntheses in three ways:

• Larsson & Wolk (2006). Larsson & Wolk [2] checked Norat's reference list, and they took Norat as the source for the 120 g portion size used in their dose–response analysis (pp. 2657, 2660).
• CUP systematic literature review (2010). The review [3] quotes two of Norat's estimates. It gives the cohort-only dose–response estimate, 1.22 (1.05–1.41), "nine studies", in section 2.5.1 ("Published meta-analysis"). It also quotes the all-designs fresh-red-meat estimate, 1.19 (0.91–1.55), 8 studies, in section 2.5.1.3.
• IARC Monograph 114. The monograph lists Norat et al. (2002) among the meta-analyses before 2010 that found "a moderate, but significant, increase" (section 2.2.3, p. 142) [4].

In a text search of the 2007 WCRF/AICR Second Expert Report we found no citation of the 2002 meta-analysis.

The study list behind the cohort estimates is not given in the text. The per-study values sit in a forest plot (Fig 2, p. 245) that prints no numbers, and in a dose–response figure (Fig 4, p. 246) that prints only pooled lines. This paper asks whether the two cohort estimates can be rebuilt from what Norat printed, and how far each depends on extraction choices.

2. Methods

2.1 Protocol

The re-analysis follows a protocol fixed before any run (REANALYSE_PROTOKOLL.md, version 1.0, 30 October 2026).

Scope. It covers cohort and nested case–control studies, red meat as each synthesis defined it, and colorectal, colon or rectal cancer incidence. Norat pooled case–control studies, so only the cohort stratum is reproduced here (protocol §2, §7).

Reproduction rules. - R1: use the synthesis's own model and increment. - R2: use exactly the studies it lists. - R3: use its own per-study values where printed. Reproduction succeeds when the pooled point estimate and both confidence limits match to two decimals. - R4: where values are not printed, derive them from the primary paper with the synthesis's dose rules.

Deviation classes (R5). exposure, dose, cases, model, duplicate, transcription and inclusion.

Corrected and follow-up runs (R6). Corrected runs apply one class at a time, then all classes. A further run substitutes later follow-ups. Leave-one-out is reported for every run.

2.2 Norat's stated method

Norat describes the method as follows (p. 242):

• Effect measure. The overall effect size was "the average of the logarithm of the observed relative risks … associated to the highest versus the lowest level of consumption, as reported in the papers", with inverse-variance weights.
• Model. "A random effect meta-analysis was performed in situations where heterogeneity was present."
• Eligibility for pooling. "Only studies reporting RR estimates with confidence intervals or quantitative information allowing their computation were included."
• Dose–response. Greenland–Longnecker trend estimation [5] was used, with "120 g as the approximate average 'portion size' of meat and of red meat".
• Category doses. Each category's dose was "the mid-point of the logarithm of the boundaries". An open upper category took the width of the closest interval. An open lower category started at zero.
• Inclusion criteria. Studies could have "incidence or mortality as the endpoint".
• Definition of red meat. The definition varied, and both versions were accepted: "Red meat was sometimes defined as the intake of beef, veal, pork, mutton and lamb consumed fresh, whereas in others, processed red meat was also included as part of the red meat group."

2.3 Data

Norat's extraction. We transcribed the per-study values Norat printed for each cohort from Appendix III (pp. 255–256) into synthesis_extraction.csv (18 rows for this synthesis). Fig 2 (p. 245) was rendered at 220 dpi to identify which cohorts are in the red-meat highest-versus-lowest stratum. It names nine:

Willett 90, Chen 92, Thun 92, Giovannucci 94, Bostick 94, Kato 97, Singh 98, Hsing 98, Pietinen 99

"Chen 92" is Chen 1998 (Appendix III, ref. 57). The figure prints no per-study RRs, confidence intervals or weights.

Primary-paper values. These come from the project database (estimates.csv). Each row in it was verified against the full text by a controller who had not seen the extraction. Hsing 1998 is not in the database. Thun 1992 has been coded from the PDF (work/thun1992) and independently controlled (K13, 0 transcription errors); its 40 rows are in estimates.csv.

2.4 Reproduction rule applied here

R3 could not be met strictly, because Norat printed no per-study values for the pooled estimate. We therefore pooled the Appendix III value for the highest versus lowest category of each named cohort.

Thun 1992 has no confidence interval in Appendix III. The appendix prints separate values for men and women (1.21 and 1.05) without intervals. Fig 2 plots a single Thun entry with an interval. The independent controller, K11, read that entry from the rendered page as about 1.14 (0.93–1.39). K11 calibrated the axis on the tick marks and graded the reading low-confidence (K11 report §1a).

The author later supplied a high-resolution scan of p. 245 (fig_in/norat_fig2_hires_from_author.png). It was read on a linear axis calibrated on the eight printed tick marks (140.3 pixels per RR unit; correction log R63). That reading gives Thun 1.14 (0.91–1.38), confirming K11. The runs keep K11's value, 1.14 (0.93–1.39), which reproduces the published estimate exactly.

We report the pool both without and with this figure-read value. For the per-120 g estimate no reproduction was possible, so none was attempted (§3.5).

2.5 Corrected and follow-up runs

Exposure. We substituted the unprocessed red-meat row where the primary paper separates it. This is protocol §2: "in corrected runs, unprocessed red meat where the paper separates it". Where the paper does not separate it, the deviation is documented and the row is kept.

Inclusion. We removed mortality-only cohorts (protocol §2).

All classes. Both corrections were applied together.

Follow-ups. We substituted later publications of the same cohorts, as coded in scripts/reanalyses.py: - NHS (Willett 1990) by Bernstein et al. (2015), women, >1 serving/day versus the lowest category, unprocessed red meat (Suppl. Table S1). - HPFS (Giovannucci 1994) by Bernstein et al. (2015), men (Suppl. Table S2) [6]. - IWHS (Bostick 1994) was not substituted. Its later publication, Sellers et al. (1998) [7], reports red meat only within family-history strata (Table 3, p. 362), so it has no whole-cohort estimate.

2.6 Software and statistics

All runs come from one script, scripts/reanalyses.py, which calls scripts/pool.py.

• Pooling. DerSimonian–Laird random effects [8] on log RR, with the standard error taken as (ln upper − ln lower)/(2 × 1.96). Fixed-effect estimates, Cochran's Q, I² and τ² are also reported.
• Duplicates. The script refuses two rows from the same cohort.
• Independent check. The pooled results were recomputed by K11 with a separate script written without reference to pool.py (kontroll/meta_check.py).

2.7 AI-assisted extraction and control

All extraction, coding and control were done by AI agents (Anthropic Claude Opus) under the author's direction. A codebook (KODEBOK.md) and fixed vocabularies were used. Every database row was checked against the full text by an independent agent instance. The synthesis extraction was checked by a separate controller (K11) against Norat's printed pages. Every change is recorded in a correction log (rettelogg.md, entries R1–R80). No number in this paper was typed by hand from memory: each comes from the run files kontroll/reanalyser/norat2002.md/.json or from the cited source page.

3. Results

3.1 Norat's per-study inputs against the primary papers

Table 1. Red-meat cohort highest-versus-lowest inputs: Norat 2002 Appendix III compared with the primary papers.

Transcription. Seven of the nine values can be checked against a primary paper in the database, and Appendix III reproduces all seven exactly (run: "HvL – the same contrasts read from the primary papers (database; Thun, Hsing not in database)"). K11 found no transcription error in the 17 Norat rows it checked; the eighteenth row, the figure-read Thun entry, is K11's own reading (R59).

Thun 1992. Appendix III's two Thun values, 1.21 and 1.05, are Thun's quintile-5 estimates for "fatty meats" (Table 4, p. 1495). Thun prints them without confidence intervals, as crude estimates from conditional logistic regression matched only on age, race and sex. The paper contains no multivariable red-meat estimate; its text says only that no associations were seen with red meat. Thun's quintiles are of a weekly frequency index, with no gram values, although Appendix III labels the rows "Red meat (g/day)". The 2,757 deaths that Appendix III gives are all colon-cancer deaths in the cohort before exclusions; the analysis used 1,150 cases (Thun 1992; work/thun1992/NOTES.md). The single Thun entry plotted with an interval in Fig 2 therefore corresponds to nothing that Thun prints.

Appendix III printing errors. The appendix has several errors. None of them affects the nine top-category values: - Willett's 84–105 g row is printed "0.25 (0.73–2.13)". - Giovannucci's RR column is offset by one row in the text layer. - Chen's lowest category is printed "≤5".

Fig 2 against Appendix III. K11 first digitised the nine Fig 2 lines from the 220-dpi rendering; seven agreed with Appendix III to ±0.04, and Singh sat nearer 1.50 (K11 report §1a, low confidence). The calibrated high-resolution reading (R63) settles both points: - Seven of the eight printed Appendix III values are reproduced within ±0.01. Examples: Willett 1.77 (1.09–2.89), Chen 1.16 (0.67–2.03), Giovannucci 1.71 (1.14–2.55), Bostick 1.04 (0.61–1.76) and Kato 1.23 (0.67–2.23). The figure is therefore an accurate plot of the values Norat used. - Singh 1998 is plotted at 1.49 (0.94–2.35), not at the 1.41 (0.90–2.21) printed in Appendix III and in Singh's Table 3. Norat thus pooled a different Singh estimate from the one he tabulated; we have not identified which. - The choice of Singh value barely matters: with the plotted value instead of the tabulated one, the pool is 1.27 (1.12–1.46). The point estimate is unchanged and each confidence limit moves by 0.01 (our recomputation; R63).

3.2 Pooled estimates

Table 2. Published, reproduced, corrected and follow-up estimates (DerSimonian–Laird random effects).

The eight printed values. Pooled, they give 1.38 (1.16–1.64) (run: "HvL – Appendix III values for the 9 cohorts named in Fig 2 (8 with CIs)"). This fails R3 by 0.11 on the point estimate.

Adding Thun. With the figure-read Thun entry added, the pool is 1.27 (1.11–1.45), with Q = 7.34 and I² 0 % (run: "HvL – the same 8 + Thun 1992 read off Fig 2 (pixel reading by controller K11, not printed)"). This is the published estimate to two decimals. Norat's p-heterogeneity, 0.45, is not printed to a precision that allows a check of Q. Because τ² is zero, the fixed-effect estimate is identical, so the result does not depend on whether Norat used fixed or random effects in this stratum.

Weights. In the reproduced run the Thun entry carries 43.5 % of the weight. The next largest is Giovannucci 1994 at 11.1 %.

3.3 Deviations

The deviation list is in norat2002.md ("Deviations").

1. Pietinen 1999 (ATBC): exposure. Norat's row, 203 vs 79 g/day, is Pietinen's "total red meat", which includes processed meat (Pietinen Table 5, p. 393; Chan et al. 2011, Table S1 [18]). Pietinen also prints beef, pork and lamb alone: Q4 (median 99 g/day) vs Q1 gives 0.80 (0.50–1.20). Substituted in the exposure and all-classes runs.
2. Hsing 1998 (Lutheran Brotherhood): inclusion. Fatal colorectal cancer only; Appendix III records deaths (p. 256). Removed in the inclusion and all-classes runs.
3. Thun 1992 (CPS-II): inclusion. Fatal colon cancer only (1,150 matched cases in the analysis), and the value is figure-read. Removed in the inclusion and all-classes runs.
4. Chen 1998 (PHS): inclusion. A nested case–control study placed in the cohort stratum. Documented only; retained.
5. Giovannucci 1994 (HPFS): exposure. "Red meat" includes processed meat (Table 2, p. 2392). The database holds no unprocessed alternative. Documented only.
6. Willett 1990 (NHS): exposure. "Red meat" includes processed meat. (Willett Table 4, p. 1668.) No unprocessed alternative. Documented only.
7. Thun 1992 (CPS-II): transcription. Thun prints the quintile-5 "fatty meats" estimates, 1.21 (men) and 1.05 (women), without confidence intervals (Table 4, p. 1495). The single entry with an interval in Norat's Fig 2, ≈1.14 (0.91–1.38), is therefore Norat's own construction from inputs the paper does not print. No alternative exists; documented only. (Thun is removed in the corrected runs in any case, under deviation 3.)

There is no dose, cases, model or duplicate deviation in the highest-versus-lowest stratum. The Singh discrepancy between Fig 2 and Appendix III (§3.1) is an inconsistency within Norat's paper, not a deviation from the primary paper, and is not a run-file entry.

3.4 Corrected and follow-up runs

Exposure correction. Substituting ATBC's unprocessed row moves the estimate to 1.26 (1.06–1.50). Heterogeneity rises to I² 27 %, and the fixed-effect estimate is 1.23 (1.08–1.41).

Inclusion correction. Removing the two mortality cohorts gives 1.35 (1.13–1.62). This is the same seven-study set as the primary-paper run.

Both corrections. Together they give 1.27 (1.02–1.59), I² 34 %, with a fixed-effect estimate of 1.28 (1.07–1.54) (run: "HvL (Norat's 9-cohort set) – corrected: all classes").

Follow-ups. NHS and HPFS were replaced by their later publication on top of the corrections; IWHS keeps Norat's Bostick 1994 row. The estimate becomes 1.01 (0.88–1.16), I² 0 % (run: "HvL – corrected + follow-ups"). The two Bernstein 2015 rows are 0.96 (0.76–1.21) for NHS and 0.98 (0.75–1.27) for HPFS; together they carry 62.1 % of the weight.

[Figure 1. Forest plots for every run listed in Table 2. See the HTML report, kontroll/reanalyser/html/index.html, panel forest-norat2002.]

3.5 Leave-one-out

Reproduced run. Only one omission moves the estimate by more than 0.05: dropping Thun gives 1.38 (1.16–1.64). The remaining leave-one-out estimates range from 1.22 (1.06–1.41) without Giovannucci to 1.29 (1.12–1.48) without Bostick or without Pietinen.

All-classes corrected run. Five of the seven leave-one-out intervals include 1.00: - without Willett: 1.21 (0.96–1.53) - without Chen: 1.29 (0.99–1.67) - without Giovannucci: 1.19 (0.94–1.51) - without Kato: 1.28 (0.99–1.65) - without Singh: 1.25 (0.96–1.63)

The two that exclude 1.00 are without Bostick, 1.31 (1.02–1.69), and without Pietinen, 1.41 (1.16–1.72).

Follow-up run. Every leave-one-out estimate lies between 0.98 (0.85–1.13), without Singh, and 1.04, and every interval includes 1.00.

3.6 Dose–response per 120 g/day

The cohort dose–response estimate of 1.22 (1.05–1.41) could not be reproduced: - The text names neither the nine cohort studies in Table II nor their slopes. - Fig 4 (p. 246) shows pooled lines only. - Appendix III prints category values, but not the doses Norat assigned under its log-midpoint rule. It prints one continuous estimate, Goldbohm 1994, per 5 g/day. That estimate is a primary-study value; Goldbohm is not in the red-meat highest-versus-lowest set.

The paper is also inconsistent about the number of studies. The text says the dose–response analysis included "Eighteen studies (9 case-control and 9 cohort)" (p. 244), while Table II gives n = 17 in total and n = 8 for case–control studies.

Norat states the inter-quantile mean ranges: 116 g/day for the three cohorts reporting g/day, and 101 g/day for the cohorts whose intake was rescaled (pp. 244–245). Without the per-study inputs, R4 (derivation from the primary papers) would mean guessing which nine cohorts were used. We did not do this.

4. Discussion

4.1 What the reproduction shows

The published estimate is correct arithmetic. It was not computed from the Appendix III values alone. The difference between 1.38 and 1.27 is explained by one study: Thun 1992. Norat's own rule allowed studies with "quantitative information allowing [the] computation" of a confidence interval (p. 242). So including Thun with a derived interval is consistent with the stated method.

What cannot be reconstructed is how the interval was derived. The plotted value, 1.14, lies between Thun's two sex-specific estimates (1.21 and 1.05) and is plausibly a combination of them. A confidence interval for such crude estimates can be computed from the numbers of cases and controls in each quintile. Thun does not print these for the food groups, so Norat must have had information beyond the published paper (for example from the authors) or used an approximation. Norat does not say which, and the entry's input is crude, while Norat's stated rule was to take the estimate with "the greatest degree of controlling for confounders" (p. 242). Since Thun offers no adjusted red-meat estimate, using the crude one is defensible, but it should have been stated.

The estimate cannot be verified from what was printed. It can be verified only by measuring pixel positions on a printed figure. That entry holds 43.5 % of the weight, so the published estimate rests largely on a number the paper never states. This is a reporting limitation, not an arithmetic error.

Transcription is clean, with one internal inconsistency. Every Appendix III top-category value we could check matches its source, and Fig 2 plots seven of them accurately. Singh 1998 is the exception: the figure, and so the pool, uses a value (1.49) that differs from the one tabulated (1.41). The effect on the pooled estimate is negligible, but a reader of the appendix cannot know that the pooled input differs.

4.2 Errors or defensible choices?

None of the deviations listed in §3.3 is a breach of Norat's own stated rules, but one is undocumented: the Thun entry (deviation 7; §4.1). The others are as follows:

• Mortality cohorts. Norat's criteria explicitly admitted "incidence or mortality" endpoints (p. 242). Removing Thun and Hsing applies the scope of our protocol (which follows CUP practice), not Norat's. The inclusion label in the run file should be read in that sense.
• Red-meat definition. Norat states that red meat was defined in two ways in the literature, and kept both (p. 242). It then tested the difference as a subgroup across all designs: with processed meat included, 1.49 (1.26–1.77); fresh only, 1.28 (1.11–1.47) (Table I, p. 244; text p. 245).
• For ATBC, Pietinen reports both definitions. Norat took the one including processed meat without stating a rule for that choice. This is a choice, not a transcription error.
• Willett 1990 and Giovannucci 1994 offer no unprocessed alternative, so their "red meat" rows are what the papers report.
• Chen 1998. Placing this nested case–control study in the cohort stratum is a classification decision. It is common in this literature (Larsson & Wolk did the same, with a sensitivity analysis).

4.3 What the corrected and follow-up runs mean

Corrected runs. Under our protocol's definitions the point estimate barely moves: 1.27 in the published analysis and 1.27 after all corrections. The precision falls, from 1.11–1.45 to 1.02–1.59, and the result is no longer robust to the omission of single studies (§3.5).

Follow-up run. Replacing the 1990 and 1994 Harvard reports with their later follow-up gives 1.01 (0.88–1.16). The follow-up rows are not like-for-like: - Bernstein 2015's "unprocessed" red meat includes hamburger, which is coded partial under the codebook's hamburger rule (KODEBOK §4.2b). - The categories (>1 serving/day) differ from the original quintiles. - Only two of the seven cohorts are updated; the Iowa cohort has no later whole-cohort estimate (§2.5).

The run is a sensitivity analysis of what the same cohorts later reported. It is not a re-estimate of what Norat should have found with the literature available to 1999.

4.4 Limitations

1. The dose–response analysis could not be reproduced. This is the estimate that was carried into the CUP systematic literature review [3].
2. The highest-versus-lowest reproduction depends on one figure-read value. It is not printed by Norat and cannot be derived from Thun 1992. Two independent readings, one from a 220-dpi rendering (K11) and one calibrated on a high-resolution scan supplied by the author (R63), agree to within 0.02 on each limit. With the high-resolution readings the pool is 1.27–1.28 (1.11–1.12 to 1.46). The reproduction therefore holds to within 0.01, but the exact two-decimal match depends on K11's reading and the tabulated Singh value. It remains a reproduction from a figure.
3. Primary-paper checks for the mortality cohorts are incomplete. Hsing 1998 is not in the database; its value is Norat's transcription. Thun 1992 has been coded from the PDF and controlled (K13); the paper prints no confidence intervals for its meat rows.
4. OCR. Appendix III exists only as a text layer with column misalignments. Values were resolved against the rendered page.
5. AI extraction. All extraction was done by AI agents. The independent control and the correction log are published so that every value can be checked by a human reader.

5. Data and code availability

The following files are published with this paper:

• Database: database/estimates.csv, studies.csv, cohorts.csv, derived_per100g.csv, synthesis_extraction.csv (earlier versions _v1, _v2), and the codebook KODEBOK.md.
• Scripts: scripts/reanalyses.py, scripts/pool.py, kontroll/meta_check.py, kontroll/scripts/pool_synth.py.
• Results: kontroll/reanalyser/norat2002.md and norat2002.json.
• Figure and Thun coding: fig_in/norat_fig2_hires_from_author.png (high-resolution scan of p. 245), work/thun1992/ (Thun 1992 rows and notes; control K13 complete (0 errors)).
• Control reports: kontroll/K11/K11_rapport.md and meta_check_output.txt, kontroll/synthesis_extraction_NOTES.md, kontroll/synthesis_figures_NOTES.md, kontroll/rettelogg.md.
• Protocol: REANALYSE_PROTOKOLL.md.
• HTML report with forest plots: kontroll/reanalyser/html/index.html.

References

1. Norat T, Lukanova A, Ferrari P, Riboli E. Meat consumption and colorectal cancer risk: dose-response meta-analysis of epidemiological studies. Int J Cancer 2002;98:241–56.
2. Larsson SC, Wolk A. Meat consumption and risk of colorectal cancer: a meta-analysis of prospective studies. Int J Cancer 2006;119:2657–64.
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. London: Imperial College London; 2010.
4. International Agency for Research on Cancer. Red Meat and Processed Meat. IARC Monographs on the Evaluation of Carcinogenic Risks to Humans, Vol. 114. Lyon: IARC; 2018.
5. 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.
6. 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.
7. Sellers TA, Bazyk AE, Bostick RM, Kushi LH, Olson JE, Anderson KE, et al. Diet and risk of colon cancer in a large prospective study of older women: an analysis stratified on family history (Iowa, United States). Cancer Causes Control 1998;9:357–67.
8. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials 1986;7:177–88.
9. 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.
10. Chen J, Stampfer MJ, Hough HL, Garcia-Closas M, Willett WC, Hennekens CH, et al. A prospective study of N-acetyltransferase genotype, red meat intake, and risk of colorectal cancer. Cancer Res 1998;58:3307–11.
11. Thun MJ, Calle EE, Namboodiri MM, Flanders WD, Coates RJ, Byers T, et al. Risk factors for fatal colon cancer in a large prospective study. J Natl Cancer Inst 1992;84:1491–500.
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;54:2390–7.
13. Bostick RM, Potter JD, Kushi LH, Sellers TA, Steinmetz KA, McKenzie DR, 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.
14. Kato I, Akhmedkhanov A, Koenig K, Toniolo PG, Shore RE, Riboli E. Prospective study of diet and female colorectal cancer: the New York University Women's Health Study. Nutr Cancer 1997;28:276–81.
15. Singh PN, Fraser GE. Dietary risk factors for colon cancer in a low-risk population. Am J Epidemiol 1998;148:761–74.
16. Hsing AW, McLaughlin JK, Chow WH, Schuman LM, Co Chien HT, Gridley G, et al. Risk factors for colorectal cancer in a prospective study among U.S. white men. Int J Cancer 1998;77:549–53.
17. Pietinen P, Malila N, Virtanen M, Hartman TJ, Tangrea JA, Albanes D, et al. Diet and risk of colorectal cancer in a cohort of Finnish men. Cancer Causes Control 1999;10:387–96.
18. Chan DSM, Lau R, Aune D, Vieira R, Greenwood DC, Kampman E, et al. Red and processed meat and colorectal cancer incidence: meta-analysis of prospective studies. PLoS ONE 2011;6(6):e20456.

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.