OpenClaw R Stats

When to use

• User asks for statistical analysis, regression, hypothesis testing
• User asks to compare groups, test significance, find associations
• User mentions ANOVA, t-test, chi-square, correlation
• User asks for time series forecasting or trend analysis
• User uploads CSV and wants statistical insights
• User asks "is this significant?" or "what predicts X?"
• 用户用中文提到：回归、检验、预测、显著性、描述统计

What this skill does NOT do

• Do not claim causality from observational data. Use "associated with".
• Do not run large exploratory fishing without clear user intent.
• Do not silently ignore assumption violations.
• Do not execute arbitrary inline R code. Always use the wrapper script.
• Do not install packages during analysis. Installation is a separate step.
• Do not report only p-values. Always include effect sizes and CIs.

Pre-flight checks (mandatory before any analysis)

1. Confirm the dataset file exists and is readable.
2. Run schema inspection:
   bash {baseDir}/scripts/run-rstats.sh schema --data
3. Report to the user: row/column count, types, missing values, unique counts.
4. If missing data > 5%, warn and ask how to handle.
5. If sample size < 30, warn about small sample limitations.
6. Only then proceed to build the analysis spec.

Environment check

If first time or errors occur:
bash {baseDir}/scripts/run-rstats.sh doctor

If packages missing:
Rscript {baseDir}/scripts/install-core.R

Standard workflow

1. Determine the correct analysis type.
2. Inspect dataset schema and missingness.
3. Build a JSON analysis spec:
   {
   "dataset_path": "",
   "analysis_type": "",
   "outcome": "",
   "predictors": ["",""],
   "formula": "",
   "group_var": "",
   "hypothesis": "",
   "missing_strategy": "complete_case",
   "alpha": 0.05,
   "seed": 42,
   "output_dir": ""
   }
4. Save the spec as a .json file.
5. Run: bash {baseDir}/scripts/run-rstats.sh analyze --spec
6. Read summary.json and report.md from the output directory.
7. Present results: Summary → Statistics → Interpretation → Plots → Assumptions → Caveats.
8. Offer follow-up: diagnostics, alternative methods, export.

Analysis selection

User intent	analysis_type
Describe data / EDA	summary
Compare 2 groups (continuous)	ttest
Compare 2 groups (non-normal/small n)	wilcoxon
Compare 3+ groups (continuous, normal)	anova
Compare 3+ groups (non-normal/ordinal)	kruskal
Compare categorical variables	chisq
Categorical (small expected counts)	fisher
Paired categorical (before/after)	mcnemar
Repeated measures non-parametric	friedman
Association between 2 continuous vars	correlation
Predict continuous outcome	linear_regression
Predict binary outcome	logistic_regression
Predict count outcome	poisson_regression
Forecast time series	forecast_arima
Assess missing data patterns	missing_diagnostics
Impute missing values	multiple_imputation
Correct for multiple comparisons	p_adjust
Survival curves + median survival	kaplan_meier
Survival regression (HR)	cox_regression
Competing risks (Fine-Gray)	competing_risks
Time-dependent Cox model	cox_time_dependent
Restricted mean survival time	rmst
Odds ratio (case-control)	odds_ratio
Risk ratio + NNT (cohort/RCT)	risk_ratio
Incidence rate ratio (person-time)	incidence_rate
Stratified analysis (confounding)	mantel_haenszel
Number needed to treat/harm	nnt
Linear mixed model (random effects)	lmm
Generalized linear mixed model	glmm
GEE marginal model	gee
Intraclass correlation	icc
Propensity score matching	propensity_match
Propensity score weighting (IPW)	propensity_weight
Causal mediation analysis	mediation_analysis
Instrumental variable regression	iv_regression
Difference-in-differences	did
Regression discontinuity	rdd

Automatic method switching guardrails

• Normality doubtful AND n < 30 → prefer wilcoxon over ttest
• Variance equality doubtful → use Welch t-test (equal_var: false)
• Expected cell counts < 5 → prefer fisher over chisq
• Overdispersion in Poisson → warn, suggest negative binomial
• Residuals heteroscedastic → warn about robust SE

Reporting rules (non-negotiable)

Every analysis MUST include:

• Sample size (n) and missing data handling
• Method name and selection rationale
• Point estimates with confidence intervals
• Effect sizes (Cohen's d, η², R², OR, etc.)
• Assumption check results
• Warnings or limitations

Language rules:

• ✓ "associated with" / "evidence suggests" / "estimated effect"
• ✗ NEVER "causes" / "proves" / "definitively shows"

Output artifacts (every run produces all of these)

File	Contents
summary.json	Status, method, findings, warnings, artifact paths
schema.json	Column types, missingness, unique counts
report.md	Human-readable analysis report
session_info.txt	R version, packages, platform, timestamp
executed_spec.json	Copy of input spec for reproducibility
tables/*.csv	Coefficients, group stats, forecast values
figures/*.png	Diagnostic and result plots

Rules

• Never run ad-hoc inline R code when the wrapper script can be used.
• Never install packages during an analysis run.
• Never access the internet during analysis execution.
• Always set a random seed for reproducibility.
• Always save session info for every analysis.
• Support both English and Chinese (中文) queries.