segInc

segInc is an experimental R package for estimating and comparing cancer incidence trends in CI5plus using segmented (piecewise) regression via the segmented R package, and selecting breakpoint models using a weighted BIC (wBIC) approach in Kim et al. (2023, J. Appl. Stat.).

This repository is also a methods case study intended to align (as closely as practical) with analyses using the NCI Joinpoint Regression Program, as used in Sung et al. (2025, The Lancet) for CI5plus-based trend analyses—while documenting where results may differ due to different optimisation and model-selection strategies.

Status: research / prototyping.

---

Case study scope

Cancer sites included

This case study focuses on:

• Colorectal
• Upper GI group:
  • oesophagus
  • stomach
  • pancreas
  • gallbladder
  • liver
• Breast
• Kidney

Practical note: CI5plus extracts may encode sites as ICD-10, ICD-O, or CI5plus site groupings depending on how you export. You’ll need a consistent site mapping layer in your preprocessing.

Age bands (primary + comparator)

Primary age bands:

• 15–39 (“early”)
• 40–49 (“early”)
• 50+ (“late”)

Comparator age bands to mirror the colorectal comparison in Sung et al. (2025):

• 15–49 (“early”)
• 50–74 (“late”)

---

---

---

Why segmented regression (and how it relates to Joinpoint)

The NCI Joinpoint software fits piecewise log-linear trends and typically selects joinpoints via permutation tests and related criteria.

log-linear ordinary least squares (OLS) regression, modelling the log-transformed incidence measure (log_est) as a linear function of calendar year of diagnosis

In contrast, this project explores:

• fitting log-linear ordinary least squared (OLS) regression, modelling the log-transformed incidence as a linear function of year of cancer diagnosis
• estimating breakpoints using segmented::segmented(),
• selecting among candidate numbers of breakpoints using a wBIC criterion in Kim et al. (2023).

These approaches often agree qualitatively, but can differ because of:

• different breakpoint search/initialisation strategies,
• different model-selection criteria (permutation tests vs information criteria),
• constraints (minimum segment length, max joinpoints),
• overdispersion handling.

Goal: make the R workflow transparent and reproducible, and quantify sensitivity of breakpoint placement and trend estimates.

---

Data: CI5plus

CI5plus is produced by the International Agency for Research on Cancer (IARC).

1. Obtain CI5plus data through the appropriate channels.
2. Do not commit restricted CI5plus data to this repository unless you have explicit permission.
3. Ensure you comply with CI5plus terms of use and citation guidance.

Expected analysis-ready structure (minimum)

Your derived dataset should be aggregatable to something like:

• year (integer)
• age or age_group (needed to create the study age bands)
• cases (incident counts)
• py (person-years / population-at-risk measure)
• site (or site code)
• optional: sex, registry, country, quality_flags, etc.

---

Installation

# install.packages("devtools")
devtools::install_github("cmclean5/segInc")

References

• (Sung et al 2025) Colorectal cancer incidence trends in younger versus older adults: an analysis of population-based cancer registry data. The Lancet Oncology, Volume 26, Issue 1, 51 - 63
• (Kim, et, al. 2023) Data-driven choice of a model selection method in joinpoint regression. JOURNAL OF APPLIED STATISTICS. 2023, VOL. 50, NO. 9, 1992–2013. https://doi.org/10.1080/02664763.2022.2063265