Statistical Research Pipeline

This plugin is the classical-statistics half of the human–AI research lab. It decomposes the statistical-research execution layer — outcome-type detection, assumption checking, test selection, effect-size estimation, extended modeling, SEM, manuscript drafting — into thirty sub-skills covering fourteen outcome / model families. It is, to my knowledge, the most comprehensive open-source statistical-research pipeline currently published as installable Claude skills.

Why thirty sub-skills, not three

Real statistical research is not one workflow. The right test for a binary outcome is not the right test for survival data, which is not the right test for repeated measures, which is not the right test for a designed experiment, which is not the right test for a meta-analysis. Each outcome family has its own assumption checks, its own primary tests, its own effect-size conventions, its own extended-modeling battery.

Most “AI for statistics” tools collapse this distinction. They run regression on whatever you point at them. The result is a sentence in your manuscript that survives peer review until a thoughtful reviewer asks, “why did you use a Poisson model on your zero-inflated counts?” This plugin is structured against that failure mode: outcome type is detected first, the testing skill matched to that type is run, and the analyzing skill for that type extends the analysis with the model battery the field expects.

What’s covered

• 11 outcome families for classical hypothesis testing: continuous, binary, ordinal, nominal, count, survival, repeated measures, time series, multivariate, designed experiments, meta-analysis
• 3 SEM families for latent-variable modeling: confirmatory factor analysis (CFA), full structural equation models, growth and change models (LGCM, latent change scores)
• 2 orchestration pipelines: application papers (data → manuscript) and methodology papers (idea → benchmark → manuscript)

Each outcome family has paired testing and analyzing sub-skills. Testing handles assumption checks + primary tests + effect sizes (steps 01–03 of a typical analysis). Analyzing continues from where testing stopped, running the extended model battery + interpretability + manuscript-section drafting (steps 04–08).

Who it’s for

• Biostatisticians in clinical research, pharmacovigilance, or real-world evidence work — every outcome family in this plugin maps to standard regulatory analysis types
• Clinical scientists and trial designers who run their own statistical analyses or need to validate a CRO’s
• Quantitative psychologists, epidemiologists, public-health scientists working with longitudinal cohorts, repeated measures, latent-variable models
• Computational biologists and behavior-research PhDs writing analysis-heavy methods sections
• Career-switchers entering biostatistics from adjacent fields, who need an installable reference for what good practice looks like across the full inventory of methods

What this plugin will not do

The same boundary applies as the AI plugin: it structures execution, not judgment. It cannot decide which assumption violations matter for your specific scientific context, whether a marginal effect is substantively meaningful, whether your sample size justifies the model you’re fitting, or whether the analysis you’ve completed is the analysis the journal will accept. Those are judgment calls. They stay yours.

The plugin will, however, surface the diagnostic that should make you pause — a Box’s M test that fails, a Hosmer-Lemeshow with a low p-value, a failed proportional-hazards check. It puts the question in front of you. You decide what to do.

How to install

Clone the repo from VeraSuperHub/stat-research-pipeline and copy the vera-stat-research-skillset/ directory contents into your ~/.claude/skills/ folder. Or use the bundled vera-stat-research.plugin file if you’re managing skills through a plugin manager.

Pair with the AI Research Pipeline

This plugin and the AI Research Pipeline are designed to be installed together. Same interface conventions, same manuscript-output formats, complementary methodological territory. The classical-statistics plugin is the right tool when you have a hypothesis test or estimation problem; the AI plugin is the right tool when you have a prediction or modeling problem. Most papers in the life sciences end up using both — and having both available as installable plugins means the workflow doesn’t context-switch in the middle of a project.