ba

Bayesian Association — a unified probabilistic framework for exploring associations in categorical data, with dedicated support for QCA and Association Rule Mining.

To install: pip install ba

Claude Code skills

ba ships with two Claude Code skills in .claude/skills/:

• ba-user — For anyone using ba as a library: choosing priors, interpreting Bayes factors, writing QCA workflows, mining rules, handling small-sample warnings. Useful prompts that would trigger this skill automatically (if it's in scope):
  • "Analyze my DataFrame for pairwise associations with Bayesian credible intervals"
  • "What prior should I use for a dataset with only 15 cases?"
  • "Walk me through a QCA analysis with calibration and Boolean minimization"
• ba-contributor — For developers modifying ba's source: adding measures to the registry, extending the Bayesian or QCA layers, writing tests, understanding the ContingencyTable hierarchy and Pot bridge.

You can invoke a skill explicitly via the slash command:

/ba-user Help me set up a prior sensitivity analysis for this 2×2 table

To make the skill available in another project, symlink it into that project's .claude/skills/ directory:

# Make ba-user available in a specific project
mkdir -p /path/to/my-project/.claude/skills
ln -s /path/to/ba/.claude/skills/ba-user /path/to/my-project/.claude/skills/ba-user

# Or make it globally available (triggers across all projects)
ln -s /path/to/ba/.claude/skills/ba-user ~/.claude/skills/ba-user

See the skills documentation for more on skill scoping, and the Claude Code configuration guide for managing project vs. user-level settings.

Quick start

import ba

# Analyze all pairwise associations in a DataFrame
result = ba.analyze(df, outcome="Y")
result.summary()  # metrics + Bayesian CIs for every pair
result.top_pairs(5)  # strongest, most certain associations

Single-pair analysis

ct = ba.contingency_table(a=10, b=5, c=3, d=12)
ct.odds_ratio  # 8.0
ct.phi  # 0.471
ct.metrics(["lift", "phi", "cramers_v"])

# Bayesian posterior
post = ba.bayesian.posterior(ct, prior="jeffreys")
post.credible_interval["risk_difference"]  # (0.12, 0.71)
post.prob_gt(0.0, "risk_difference")  # 0.996

Per-tradition APIs

# Bayesian
ba.bayesian.posterior(ct, prior="uniform")
ba.bayesian.bayes_factor(ct)
ba.bayesian.sensitivity(ct, priors=["jeffreys", "uniform", "beta(2,2)"])

# QCA
binary_df = ba.qca.calibrate(df, {"age": 30, "illness": "any_present"})
tt = ba.qca.truth_table(binary_df, outcome="Y", conditions=["A", "B", "C"])
solution = ba.qca.minimize(tt)
ba.qca.necessity(binary_df, "Y", ["A", "B"])

# Association Rule Mining
rules = ba.rules.mine(df, min_support=0.1, outcome="Y")

Key features

• Categorical core, binary specialization. The core handles r×c contingency tables. Binary (2×2) is a specialization with extra metrics (OR, RR, phi, Yule's Q, QCA consistency/coverage).
• Uncertainty first. Every metric comes with Bayesian credible intervals. Bayes factors replace p-values as the default evidence measure.
• 22 built-in measures in an extensible registry — from support/lift to Cramér's V, mutual information, and Goodman-Kruskal gamma.
• QCA pipeline: calibrate → truth table → Quine-McCluskey minimization → necessity/sufficiency analysis, all with Bayesian CIs.
• Association rules with Bayesian augmentation, appearance constraints, and built-in brute-force mining (no mlxtend dependency required).
• Progressive disclosure. ba.analyze() for one-liners; ba.bayesian/ba.rules/ba.qca for per-tradition control; ContingencyTable/MeasureRegistry for full access.