# Protocol addendum R8: updated 2026 analysis by cancer site (colon, rectum, proximal and distal colon)

Version 1.0, written 30 October 2026, 18:05 EDT, **before any site-specific pooling of the 2026 cohort set has been run**. Filed as `PROTOKOLL_KOLON_R8.md`; a copy goes into the public package and the git history of larsms81/unit-1981 fixes its timestamp.

## What we already know (declared so the reader can judge us)
- CUP 2016/2017's colon estimate 1.22 (1.06–1.39) reproduces exactly, survives correction of documented errors (1.22; 1.07–1.38) and is significant in 100 % of the choice combinations available in 2016. CUP 2017 cited the significant colon result as part of the basis for "probable".
- In the comparative reanalysis, swapping the CUP 2016 colon cohorts for their later follow-ups (Islam 2019 for JPHC/Takayama, Bernstein 2015 for NHS/HPFS) gave 1.07 (0.94–1.20). That run used CUP's 2016 cohort set only.
- For colorectal cancer overall, the 20-cohort 2026 pool is 1.10 (1.04–1.15); cohorts published after 2015 give 1.13 (1.08–1.18).
- We have not yet looked at which colon or rectum values exist for the other 2026 cohorts, nor pooled them. This addendum is written before doing so.

## Aim
Apply the CUP 2016 method (R7) to colon, rectum, proximal colon and distal colon incidence, using the same 20-cohort set as the 2026 colorectal analysis, to answer: does the updated evidence support a site-specific association, and in particular the colon association CUP 2017 relied on?

## Rules
R8.1 **Cohort set.** Start from the 20 cohorts in the 2026 colorectal pool (one row per cohort, same cohort-overlap rules). A cohort enters a site analysis if, and only if, its publication (or its supplement) reports an estimate for that site for unprocessed red meat, or the closest available red-meat definition under the same hierarchy as R7 (unprocessed > partial/"red meat incl. hamburger" > undefined red meat; total red meat incl. processed is **not** used). No cohort is added that is not in the colorectal pool, except where a cohort's colorectal row came from one publication and its site rows from another publication of the same cohort version (e.g. Norat 2005 for EPIC sites if Papadimitriou 2022 has none) – that substitution must be listed in the inventory before pooling.
R8.2 **Row choice.** Main model, highest adjustment, not calibrated (calibrated as sensitivity, as in R7). Continuous per-unit estimates preferred; otherwise Greenland–Longnecker from categories with the documented dose rules (120 g/serving; printed medians; CUP midpoint rule), using the pipeline's documented covariance variant (K19) with textbook GL as sensitivity.
R8.3 **Colon from subsites.** If a cohort reports proximal and distal colon but not colon overall, the colon estimate is the fixed-effect combination of the two subsite estimates (they share no cases). This is an approximation (it ignores correlation through shared covariate estimation and treats the two as independent); it is applied identically to every such cohort and flagged in the table. Cohorts reporting colon overall use that row.
R8.4 **Sex strata.** Combined by fixed effect as in R7.
R8.5 **Pooling.** DerSimonian–Laird random effects; also REML, Hartung–Knapp, fixed; I², τ², Egger when k ≥ 10; leave-one-out.
R8.6 **Analyses, in this order, all reported whatever they show:** (a) colon; (b) rectum; (c) proximal colon; (d) distal colon; (e) test of difference colon vs rectum and proximal vs distal (ratio of pooled RRs with its CI, from the pooled log-RRs treated as independent); (f) subgroups as R7 (region, sex, publication period ≤2015 / >2015) where k ≥ 3; (g) sensitivity: cohorts with colon reported directly vs constructed from subsites; textbook GL; calibrated EPIC.
R8.7 **Pre-declared interpretation.** The heme/NOC and alkylation hypotheses predict a stronger association in the distal colon and rectum than the proximal colon. The CUP 2017 claim predicts a colon estimate significantly above 1. We will report whether each prediction is met. We commit to publishing the result in the overview article and the English update article regardless of direction, and to stating explicitly if the colon association is confirmed.
R8.8 **Reporting bias check.** List which of the 20 cohorts report site-specific results and which do not; state cases per site; note whether non-reporting cohorts are large or small. No imputation.
R8.9 **Control.** Independent recomputation (meta_check2 extension) and an independent read of the inventory against full texts before anything is published.

## Change log
- (none yet)
- 2026-10-30 18:20 EDT, after the inventory (kontroll/kolon2026/INVENTAR.md) and still before pooling: (1) EPIC sites come from Norat 2005 (Papadimitriou 2022 has no red-meat site analysis) – the R8.1 substitution; (2) UK Biobank proximal/distal from Bradbury 2020 is **not** the same cohort version as Knuppel 2020 and is used only as a sensitivity analysis; (3) EPIC right/left colon from Norat 2005 lacks per-category cases and uses a different anatomical split (splenic flexure excluded from left) – sensitivity only; (4) Islam 2019 site rows get their case counts from Suppl. S3/S4 (colon 2 552, rectum 1 127, proximal 1 200, distal 1 150) before use; (5) calibrated/regression-dilution-corrected continuous site rows (Knuppel, Bradbury, Norat) are sensitivity only, per R8.2; (6) 9 of 20 cohorts (41 % of colorectal cases, incl. Million Women and MEC) report no usable site estimate for unprocessed red meat – reported under R8.8, no imputation.
- 2026-11-01 (R100–R102, after K26): (1) The colorectal pool now has 19 cohorts: ATBC (Pietinen 1999) was a duplicate of ATBC inside Tammi 2024 and has been removed. Neither publication reports site rows, so no site pool changes; the R8.8 reporting table is now 8 of 19 cohorts without site data (18,131 cases, 40 %), against 9 of 20 (18,316, 41 %) when this addendum was written. (2) R8.2 says "not calibrated (calibrated as sensitivity, as in R7)". That parenthesis is inaccurate: R7 states no calibration rule, and the R7 colorectal main pool includes one calibrated row (Million Women, the only row Papier 2025 prints). For the site pools R8.2 applies as written: every site row is observed, and calibrated/regression-dilution-corrected rows are sensitivity only (change log 2026-10-30, item 5). Million Women has no site rows for unprocessed red meat, so the two rules do not meet in any site pool.