Subgroup Analysis (bioSkills)

A Claude Code skill for analysing treatment effects across patient subgroups without over-claiming: interaction testing done properly, data-adaptive HTE methods where they are warranted, and a credibility framework applied to whatever comes out.


Type	Claude Skill
Supplier	GPTomics bioSkills (community OSS, MIT)
Availability	GA — part of the bioSkills collection
Pricing	Free / OSS (MIT) — `statsmodels`, `scipy`, `numpy`, `pandas`, `matplotlib`, `scikit-learn`, `econml` (Python) and `grf`, `policytree`, `causalToolbox`, `personalized`, `SIDES`/`rsides`, `stepp`, `gMCP`, `partykit`, `RBesT`, `brms` (R) are separately installed OSS.
Capabilities	Read/Write — Claude runs the skill’s workflow locally (Python and R), not as an MCP tool
Verified	works · 2026-08-13
Security	cleared · 2026-08-13 — GPTomics/bioSkills MIT, no external credentials

How to install

bioSkills is not an npm package — skills are plain markdown/code read directly by the agent. Clone the repo, then either run the installer for the whole category or copy the single skill directory.

Claude Code — clone and install via the bundled script:
```
git clone https://github.com/GPTomics/bioSkills
cd bioSkills
./install-claude.sh --categories "clinical-biostatistics"
```
The installer copies matching skills into ~/.claude/skills/ (default target). Use ./install-claude.sh --list to preview the skills first.
Claude Code / other agents — copy just this one skill:
```
cp -r bioSkills/clinical-biostatistics/subgroup-analysis ~/.claude/skills/
```
(run from inside your clone — the previous step left you in bioSkills/; otherwise replace bioSkills/ with the absolute path of your clone). Install the Python side on first use with pip install "statsmodels>=0.14" "scipy>=1.12" "numpy>=1.26" "pandas>=2.1" "matplotlib>=3.8" "scikit-learn>=1.4" econml; causal-forest work needs R with install.packages(c("grf", "policytree", "stepp")).

What it does

Routes the question to a method, then subjects the answer to a credibility check.

Scenario	Method
Pre-specified subgroup, confirmatory trial	Interaction term in a single model + graphical multiplicity (`gMCP`) — the EMA 2019 “assessment subgroup”
Stratified randomisation (site, region)	CMH, or logistic regression with strata as covariates; stratum-specific ORs in the forest plot
Many pre-specified subgroups	Bayesian shrinkage (Dixon–Simon, MAP, EXNEX) rather than independent stratum estimates
Exploratory discovery	STEPP, SIDES, causal forests, X- / R-learners — labelled as hypothesis-generating
Additive-scale interaction	RERI (> 0 synergism, = 0 no additive interaction, < 0 antagonism)

Numbers it holds you to:

Detecting an interaction needs roughly 4× the sample size of the main effect (Brookes 2004). This is why most subgroup analyses are underpowered by construction.
0.5σ standardised effect as the “noteworthy heterogeneity” benchmark (Dane 2019, EFSPI).
≤ 20% of total α for exploratory subgroups (EMA 2019).
Minimum n ≥ 5 per cell for CMH to avoid sparse-data bias; Breslow–Day is unreliable with fewer than 5 strata.
Bayesian shrinkage prior: tau ~ HalfNormal(0, 0.1) conservative, HalfNormal(0, 0.5) regulatory-tolerant, with a sensitivity sweep across 0.1–0.5 expected.
EXNEX mixture weights default 0.5 / 0.5, swept over 0.1–0.9.
Causal-forest honest-split fraction typically 0.5; tune with tune.parameters = "all".

Credibility is assessed against the Sun BMJ 2012 11-criterion checklist and the EMA 2019 subgroup guideline.

Primary use cases: pre-specifying a subgroup strategy in an SAP, assessing whether an observed subgroup effect is credible, exploratory HTE discovery.

Notes

A subgroup-specific p-value is not evidence of a differential effect — the interaction test is, and it is the comparison the skill insists on. Reporting “significant in men, not in women” without an interaction test is the classic failure this skill is built to prevent.

Ignoring the randomisation strata is a real cost, not a conservatism: Kahan–Morris 2012 show it biases standard errors upward and loses power, so stratified designs must be analysed as stratified.

Data-adaptive methods (causal forests, SIDES) find subgroups by searching, so their output is hypothesis-generating regardless of how the split is scored. The honest-splitting and cross-fitting machinery controls estimation bias within a discovered subgroup, not the selection itself.

Distributed as a SKILL.md (plus reference material) in the bioSkills collection — Claude executes the workflow locally rather than as an MCP server. The upstream skill front-matter name is bio-clinical-biostatistics-subgroup-analysis; if invoked as a namespaced plugin command it resolves under the bioSkills plugin, not as a bare /subgroup-analysis. Upstream directory: clinical-biostatistics/subgroup-analysis.

Related: Multiplicity and Graphical Procedures (defines the α budget this analysis spends), Bayesian Trials (the MAP/EXNEX shrinkage machinery), Power and Sample Size (the 4× interaction penalty), Trial Reporting, statsmodels, scikit-learn.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.