### Install and Run StatsPAI Examples (Released Package)

Source: https://github.com/brycewang-stanford/statspai/blob/main/examples/README.md

Install the released StatsPAI package and then run the example scripts. This is the standard way to run examples after the package has been installed.

```bash
python -m pip install statspai
python examples/card_iv.py
```

--------------------------------

### Install and Run StatsPAI Examples (Source Checkout)

Source: https://github.com/brycewang-stanford/statspai/blob/main/examples/README.md

Install StatsPAI in development mode with plotting dependencies and then run the example scripts. This is useful for development or when working directly from a source checkout.

```bash
python -m pip install -e ".[dev,plotting]"
python examples/card_iv.py
python examples/did_mpdta.py
python examples/rd_lee.py
python examples/synth_prop99.py
python examples/gmethods_timevarying.py
python examples/nhefs_whatif.py
```

--------------------------------

### Quickstart: Full Replication Pipeline

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/replication_workflow.md

This snippet demonstrates the core two-step process for creating a replication archive: first generating a draft from data and analysis, then packaging it into a submission-ready zip file. Ensure `statspai` and `pandas` are installed.

```python
import statspai as sp
import pandas as pd

df = pd.read_csv("training_panel.csv")

# 1. Question → estimate → draft
draft = sp.paper(df, "effect of trained on wage",
                 treatment="trained", y="wage",
                 fmt="qmd")  # Quarto-native output

# 2. Draft -> submission-oriented replication archive
sp.replication_pack(draft, "submission.zip",
                    code="analysis.py")
```

--------------------------------

### StatsPAI Quick Example

Source: https://github.com/brycewang-stanford/statspai/blob/main/README.md

A comprehensive example demonstrating various StatsPAI functionalities including estimation, post-estimation, table generation, and robustness workflows. Ensure StatsPAI is installed with necessary dependencies.

```python
import statspai as sp

# --- Estimation ---
card = sp.datasets.card_1995()           # Card (1995) returns-to-schooling (n=3010)
r1 = sp.regress("lwage ~ educ + exper", data=card, robust='hc1')
r2 = sp.ivreg("lwage ~ (educ ~ nearc4) + exper", data=card)

mp = sp.datasets.mpdta()                 # Callaway–Sant'Anna staggered DiD (n=2500)
r3 = sp.did(mp, y='lemp', treat='first_treat', time='year', id='countyreal')

lee = sp.datasets.lee_2008_senate()      # Lee (2008) sharp RD (n=6558)
r4 = sp.rdrobust(lee, y='voteshare_next', x='margin', c=0)

nsw = sp.datasets.nsw_dw()               # LaLonde / NSW-DW job training (n=2675)
r5 = sp.dml(nsw, y='re78', treat='treat',
            covariates=['age', 'education', 're74', 're75'])
r6 = sp.causal_forest("re78 ~ treat | age + education + re74 + re75", data=nsw)

# --- Post-estimation ---
sp.margins(r1, data=card)                 # Marginal effects
sp.test(r1, "educ = exper")               # Wald test
sp.oster_bounds(card, y='lwage', treat='educ', controls=['exper'])

# --- Tables (to Word / Excel / LaTeX) ---
sp.modelsummary(r1, r2, output='table2.docx')
sp.outreg2(r1, r2, r3, filename='results.xlsx')
sp.sumstats(card, vars=['lwage', 'educ', 'exper'], output='table1.docx')

# --- Robustness workflow ---
sp.spec_curve(card, y='lwage', x='educ',
              controls=[[], ['exper'], ['exper', 'black']],
              se_types=['nonrobust', 'hc1']).plot()

sp.robustness_report(card, formula="lwage ~ educ + exper",
                     x='educ',
                     winsor_levels=[0.01, 0.05]).plot()

sp.subgroup_analysis(card, formula="lwage ~ educ + exper",
                     x='educ',
                     by={'Region': 'south', 'Race': 'black'}).plot()
```

--------------------------------

### Setting up and Benchmarking GPU Acceleration on Colab

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/gpu_acceleration.md

This example demonstrates how to install StatsPAI with JAX and CUDA support, verify the JAX device, and benchmark CPU vs. GPU performance for bootstrap methods on Google Colab. It includes creating a sample dataset and timing both sequential CPU execution and GPU-accelerated vmap'd bootstrap.

```python
# In a Colab notebook with a GPU runtime selected
!pip install -q statspai jax[cuda12] jaxlib

import statspai as sp
print(sp.fast.jax_device_info())
# Expect: jax: <version>, default device: cuda

# Build a benchmark dataset
import numpy as np, pandas as pd
rng = np.random.default_rng(0)
n, n_firm = 1_000_000, 5_000
firm = rng.integers(0, n_firm, size=n)
fe = rng.normal(size=n_firm)[firm]
df = pd.DataFrame({
    "y": 0.5 * rng.normal(size=n) + fe,
    "x1": rng.normal(size=n),
    "x2": rng.normal(size=n),
    "firm": firm,
})

# Time CPU vs GPU
import time

t0 = time.perf_counter()
for _ in range(2_000):
    _ = sp.fast.feols("y ~ x1 + x2 | firm", df, vcov="hc1")
print(f"CPU sequential bootstrap (B=2000): {time.perf_counter() - t0:.1f}s")

t0 = time.perf_counter()
boot = sp.fast.feols_jax_bootstrap(
    "y ~ x1 + x2 | firm", df, n_boot=2_000, bootstrap="pairs",
    vmap_chunk_size=500,  # tune up for big-HBM GPUs
)
print(f"GPU vmap'd bootstrap (B=2000):     {time.perf_counter() - t0:.1f}s")
```

--------------------------------

### Install and Run Tests

Source: https://github.com/brycewang-stanford/statspai/blob/main/README.md

Clone the repository, install dependencies including development and plotting tools, and run tests.

```bash
git clone https://github.com/brycewang-stanford/statspai.git
cd statspai
pip install -e ".[dev,plotting,fixest]"
pytest
```

--------------------------------

### Install StatsPAI and Optional Backend

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/migration-from-r.md

Install the core StatsPAI library using pip. Optionally, install pyfixest for the HDFE backend to handle large-scale fixed-effects models.

```bash
pip install statspai
```

```bash
# Optional HDFE backend (for sp.feols / sp.fepois at scale):
pip install pyfixest
```

--------------------------------

### Install StatsPAI

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/getting-started.md

Install the core StatsPAI library using pip. Optional extras can be installed with specific dependencies.

```bash
pip install StatsPAI
```

```bash
pip install "StatsPAI[plotting]"   # matplotlib / seaborn / plotly figures
```

```bash
pip install "StatsPAI[bayes]"      # PyMC + ArviZ for Bayesian estimators
```

```bash
pip install "StatsPAI[tune]"       # Optuna for tuned meta-learners / Auto-CATE
```

```bash
pip install "StatsPAI[rd-cct]"     # rdrobust for exact CCT RD parity
```

```bash
pip install "StatsPAI[performance]"# JAX backend for fast feols / bootstrap
```

--------------------------------

### Set Up Python Development Environment

Source: https://github.com/brycewang-stanford/statspai/blob/main/CONTRIBUTING.md

Create a virtual environment, activate it, and install the project in development mode with necessary dev dependencies. Also, install pre-commit hooks for code quality checks.

```bash
# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install in development mode
pip install -e ".[dev]"

# Install pre-commit hooks
pre-commit install
```

--------------------------------

### Setup and Imports

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/mixtape_ch09_did.md

Import necessary libraries and set a random seed for reproducibility.

```python
import numpy as np
import pandas as pd
import statspai as sp

# Reproducible throughout the chapter:
RNG_SEED = 2026
```

--------------------------------

### Install StatsPAI from Repository Checkout

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/joss_reviewer_guide.md

Install StatsPAI in editable mode with development and plotting extras from a local repository checkout.

```bash
python -m pip install -e ".[dev,plotting]"
```

--------------------------------

### Install StatsPAI Core Package

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/joss_reviewer_guide.md

Use these commands to create a virtual environment, activate it, upgrade pip, and install the core StatsPAI package.

```bash
python -m venv .venv-joss
source .venv-joss/bin/activate
python -m pip install --upgrade pip
python -m pip install statspai
```

--------------------------------

### Install StatsPAI with Optional Dependencies

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/SKILL.md

Install StatsPAI with specific extras for full functionality. Use `[fixest,plotting]` for high-dim FE and plotting, `[neural]` for neural causal models, and `[text]` for text-as-treatment. A one-shot install for all features is also available.

```bash
pip install "statspai[fixest,plotting]"
```

```bash
pip install "statspai[fixest,plotting,neural,text]"
```

--------------------------------

### Check Examples Coverage

Source: https://github.com/brycewang-stanford/statspai/blob/main/plans/2026-06-21-statspai-hardening-month/README.md

Verifies the coverage of example scripts, ensuring no more than the maximum allowed missing examples. This is part of the full release or manual full-CI gate before publishing.

```bash
.venv/bin/python scripts/examples_coverage.py --check --max-missing 0
```

--------------------------------

### Check Example Execution

Source: https://github.com/brycewang-stanford/statspai/blob/main/plans/2026-06-21-statspai-hardening-month/README.md

Checks the execution of example scripts for failures. This is part of the full release or manual full-CI gate before publishing.

```bash
.venv/bin/python scripts/check_example_execution.py --quiet --max-failures 0
```

--------------------------------

### Install StatsPAI Package

Source: https://github.com/brycewang-stanford/statspai/blob/main/MIGRATION.md

Install the StatsPAI package using pip. This is the first step to begin using the new library.

```bash
pip install statspai
```

--------------------------------

### Install Dependencies for Parity Checks

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/joss_reviewer_guide.md

Installs the package with development and parity extras, which includes doubleml-for-py. Use this to set up the environment for parity testing.

```bash
python -m pip install -e ".[dev,parity]"
```

--------------------------------

### Required Package Installations

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/SKILL.md

Install `pyfixest` for `sp.feols`, `fepois`, `feglm`. Install `[neural]` for neural causal models and `[plotting]` for plotting functionalities.

```python
sp.feols(...) without installing pyfixest
```

--------------------------------

### Install StatsPAI with Optional Dependencies

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/index.md

Install the core StatsPAI package or with additional dependencies for plotting, specific estimators, or text analysis. Choose the installation that best suits your needs.

```bash
pip install statspai                       # core
```

```bash
pip install 'statspai[plotting]'           # matplotlib + seaborn
```

```bash
pip install 'statspai[fixest]'             # pyfixest HDFE
```

```bash
pip install 'statspai[deepiv]'             # PyTorch (Deep IV, TARNet)
```

```bash
pip install 'statspai[text]'               # sentence-transformers for sbert
```

--------------------------------

### Install Optional Dependency for R-parity

Source: https://github.com/brycewang-stanford/statspai/blob/main/MIGRATION.md

Shows the command to install the optional dependency `rdrobust` required for using `bwselect='cct'` in `sp.rdrobust`. This is a one-time installation.

```bash
pip install statspai[rd-cct]   # adds rdrobust>=1.3
```

--------------------------------

### Install Stata Packages

Source: https://github.com/brycewang-stanford/statspai/blob/main/benchmarks/hdfe/README.md

Install necessary Stata packages for running ppmlhdfe benchmarks. Requires Stata 17+.

```bash
ssc install ppmlhdfe, replace
ssc install reghdfe,  replace
ssc install ftools,   replace
```

--------------------------------

### Install Dependencies and Run Pytest

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/joss_validation_dossier.md

Installs development and plotting dependencies, then runs pytest on specified test files without coverage. Use this for a full repository checkout check.

```bash
python -m pip install -e ".[dev,plotting]"
python -m pytest tests/test_ols.py tests/test_did.py tests/test_registry.py -q --no-cov
```

--------------------------------

### Install StatsPAI Skill via Copy

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/README.md

Use this method to copy the skill directory to your Claude skills directory.

```bash
cp -r StatsPAI_full_data_analysis_skill ~/.claude/skills/StatsPAI_skill
```

--------------------------------

### Install StatsPAI with specific extras

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/faq.md

Install StatsPAI with optional dependencies for specific estimators like Bayesian methods, neural networks, or performance enhancements.

```bash
pip install "StatsPAI[bayes]"

```

```bash
pip install "StatsPAI[neural]"

```

```bash
pip install "StatsPAI[performance]"

```

--------------------------------

### Install StatsPAI Python Package (Recommended)

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/README.md

Installs the core statspai package along with dependencies for fixest and plotting, covering the default pipeline.

```bash
pip install "statspai[fixest,plotting]"
```

--------------------------------

### sp.examples

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/reference/smart.md

Provides examples for causal inference methods.

```APIDOC
## sp.examples

### Description
Provides examples for causal inference methods. This is a guide-friendly function used by agents and notebook workflows.

### Method Signature
```python
sp.examples(...)
```

### Parameters
(Details not provided in source, refer to `::: statspai.examples` for more information)
```

--------------------------------

### Get StatsPAI Version

Source: https://github.com/brycewang-stanford/statspai/blob/main/SECURITY.md

Use this command to retrieve the currently installed version of StatsPAI. This is required when reporting a vulnerability.

```python
python -c "import statspai; print(statspai.__version__)"
```

--------------------------------

### Get StatsPAI Version

Source: https://github.com/brycewang-stanford/statspai/blob/main/SUPPORT.md

Use this command to retrieve the currently installed version of the StatsPAI library. This is often required when reporting issues or seeking support.

```python
python -c "import statspai as sp; print(sp.__version__)"
```

--------------------------------

### Get R and clubSandwich Versions

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/superpowers/plans/2026-04-27-htz-clubsandwich-parity.md

Use these commands to retrieve the current R version and the installed version of the 'clubSandwich' package. This information is crucial for pinning fixture versions and auditing potential formula changes.

```bash
R --version | head -1
```

```bash
Rscript -e 'cat(as.character(packageVersion("clubSandwich")))'
```

--------------------------------

### Choosing an Entry Point for Matching

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/reference/matching.md

Demonstrates how to use the sp.match() dispatcher for default nearest-neighbor propensity-score matching and balancing-weight estimators. Also shows how to use standalone functions for overlap weights and balance diagnostics.

```python
import statspai as sp

# Default nearest-neighbour propensity-score matching.
r = sp.match(
    df,
    y="earnings",
    treat="training",
    covariates=["age", "education", "earnings_pre"],
)

# Balancing-weight estimators are available through method=...
r_ebal = sp.match(df, y="earnings", treat="training",
                  covariates=["age", "education"], method="ebalance")

# ...or as standalone functions with estimator-specific options.
w = sp.overlap_weights(df, treat="training",
                       covariates=["age", "education", "earnings_pre"])
diag = sp.balance_diagnostics(df, treat="training",
                              covariates=["age", "education"],
                              weights=w.weights)
```

--------------------------------

### Quick Start: Generate and Run Baselines

Source: https://github.com/brycewang-stanford/statspai/blob/main/benchmarks/hdfe/README.md

Generate small and medium datasets, then run the baselines for these datasets. This is the primary command for initial benchmarking.

```bash
cd benchmarks/hdfe

python3 datasets.py --configs small medium

python3 run_baseline.py --datasets small medium
```

--------------------------------

### Install StatsPAI

Source: https://github.com/brycewang-stanford/statspai/blob/main/README.md

Install the StatsPAI package using pip. Optional dependencies can be installed for extended functionality.

```bash
pip install statspai
```

```bash
pip install statspai[plotting]    # matplotlib, seaborn
```

```bash
pip install statspai[fixest]      # pyfixest for high-dimensional FE
```

```bash
pip install statspai[bayes]       # PyMC + ArviZ for Bayesian estimators
```

```bash
pip install statspai[tune]        # Optuna for tuned meta-learners / Auto-CATE
```

```bash
pip install statspai[rd-cct]      # rdrobust for exact CCT RD parity
```

```bash
pip install statspai[deepiv]      # PyTorch for DeepIV
```

```bash
pip install statspai[neural]      # PyTorch for TARNet/CFRNet/DragonNet
```

```bash
pip install statspai[text]        # sentence-transformers for sbert text embeddings
```

```bash
pip install statspai[performance] # JAX CPU backend for sp.fast.demean
```

--------------------------------

### Quick Start: Classic Synthetic Control Method

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/synth.md

Demonstrates loading a dataset and applying the classic synthetic control method using the `sp.synth()` dispatcher. Ensure the `statspai` library is imported.

```python
import statspai as sp

# Load the California Proposition 99 dataset (Abadie et al. 2010 style)
df = sp.synth.california_tobacco()

# Classic SCM (Abadie, Diamond & Hainmueller 2010)
res = sp.synth(
    df, outcome='cigsale', unit='state', time='year',
    treated_unit='California', treatment_time=1989,
    method='classic',
)
print(res.summary())
```

--------------------------------

### End-to-End Agent Workflow Example

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/agent_api.md

Illustrates a complete agent workflow: reading a CSV, identifying study design, pre-flighting an estimator, running the estimation, summarizing results, and identifying missing robustness checks.

```python
import statspai as sp
import pandas as pd

df = pd.read_csv("/path/to/dataset.csv")

# 1. Identify the study design.
design = sp.detect_design(df)
# {'design': 'panel', 'confidence': 1.0,
#  'identified': {'unit': 'firm_id', 'time': 'year'}, ...}

# 2. Pre-flight a candidate estimator before paying for it.
report = sp.preflight(df, 'did',
                      y='sales', treat='treated', time='year')
if report['verdict'] == 'FAIL':
    # The verdict carries structured failure info — agent can
    # decide whether to fix args, switch method, or stop.
    for c in report['checks']:
        if c['status'] == 'failed':
            print(f"  blocked by {c['name']}: {c['message']}")
    raise SystemExit
elif report['verdict'] == 'WARN':
    print("warnings present but proceeding")

# 3. Run the estimator. If it raises a structured StatsPAIError,
#    the MCP layer surfaces error_kind + alternative_functions.
result = sp.did(df, y='sales', treat='treated', time='year')

# 4. One-line dashboard summary for logs / multi-result loops.
print(result.brief())
# [Difference-in-Differences (2x2)]  estimand=ATT  est=0.412
# (se=0.087)  95% CI [0.241, 0.583]  ***  N=2,000

# 5. Reviewer checklist — which robustness checks are still
#    MISSING from the result's evidence base?
audit_card = sp.audit(result)
for c in audit_card['checks']:
    if c['status'] == 'missing' and c['importance'] == 'high':
        print(f"  follow-up: {c['suggest_function']}  ({c['name']})")
# follow-up: sp.pretrends_test  (parallel_trends)
# follow-up: sp.honest_did       (rambachan_roth)
```

--------------------------------

### Release Process for v1.8.0

Source: https://github.com/brycewang-stanford/statspai/blob/main/benchmarks/hdfe/AUDIT.md

Outlines the recommended steps for releasing version 1.8.0, including version bumping, tagging, and uploading to PyPI. This process is crucial for deploying the performance improvements.

```bash
git tag v1.8.0 && git push --tags && twine upload dist/*
```

--------------------------------

### Install Python Engines for Cross-Validation

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/cross_engine_validation.md

Install in-process Python engines using pip. Ensure these packages are installed to leverage their cross-validation capabilities.

```bash
pip install pyfixest linearmodels doubleml   # in-process Python engines
```

--------------------------------

### Install R Packages for Cross-Validation

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/cross_engine_validation.md

Install necessary R packages for cross-validation. This involves installing R itself and then using the install.packages function within an R session.

```r
# R cross-check: install R, then in R:  install.packages(c("fixest","jsonlite"))
```

--------------------------------

### End-to-End Econometric Workflow

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/data_mcp_ingestion.md

This example demonstrates a complete workflow: ingesting World Bank data, detecting the econometric design, recommending an estimator, and performing cross-engine validation using StatsPAI.

```python
import statspai as sp

# 1. (the agent calls the World Bank MCP and gets `payload`)
panel = sp.from_worldbank(payload, wide=True)

# 2. Let StatsPAI read the shape and suggest an estimator.
design = sp.detect_design(panel)
rec = sp.recommend(panel, y="life_expectancy", design=design.design)

# 3. Fit — and cross-validate the headline number across engines (see the
#    cross-engine validation guide).
cv = sp.cross_validate(
    panel, "feols",
    formula="life_expectancy ~ gdp_pc | iso3 + year", treatment="gdp_pc",
    engines=["statspai", "pyfixest", "R::fixest"],
)
print(cv.verdict)
```

--------------------------------

### Build and Check Package Distribution

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/joss_reviewer_guide.md

Installs necessary build tools, creates a source distribution and wheel, and then checks the integrity of the distribution files. This verifies that the package is ready for distribution.

```bash
python -m pip install build twine
```

```bash
python -m build
```

```bash
python -m twine check dist/*
```

--------------------------------

### Install Noto CJK Font on Fedora/RHEL

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/SKILL.md

Install the Noto CJK font package on Fedora or RHEL systems to resolve tofu character display issues in plots. This is a system-level font installation.

```bash
sudo dnf install google-noto-sans-cjk-fonts

```

--------------------------------

### Install Noto CJK Font on Debian/Ubuntu

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/SKILL.md

Install the Noto CJK font package on Debian or Ubuntu systems to resolve tofu character display issues in plots. This is a system-level font installation.

```bash
sudo apt install fonts-noto-cjk

```

--------------------------------

### sp.dml API Mapping Examples

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/sp_dml_vs_doubleml.md

Illustrates how to call sp.dml for different DoubleML models, showing the equivalent DoubleML classes and their parameters.

```python
sp.dml(model='plr', ml_g=..., ml_m=...)
```

```python
sp.dml(model='irm', ml_g=..., ml_m=...)
```

```python
sp.dml(model='pliv', instrument=..., ml_g=..., ml_m=..., ml_r=...)
```

```python
sp.dml(model='iivm', instrument=..., ml_g=..., ml_m=..., ml_r=...)
```

--------------------------------

### Compare All Synth Methods

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/synth.md

Run and compare all available synthetic control methods at once. Use the plot method to visualize counterfactuals and summary_table to get ATT and RMSPE. The synth_recommend function helps auto-select the best method.

```python
comp = sp.synth_compare(df, outcome='cigsale', unit='state', time='year',
                        treated_unit='California', treatment_time=1989)
comp.plot()                # overlay all counterfactuals
comp.summary_table()       # ATT, pre-RMSPE, p-value per method
best = sp.synth_recommend(comp)   # auto-pick best by pre-fit + robustness
```

--------------------------------

### Quick Start: Re-render Report Only

Source: https://github.com/brycewang-stanford/statspai/blob/main/benchmarks/hdfe/README.md

Re-render the benchmark report without re-running the benchmarks. Useful for updating the report after changes.

```bash
python3 run_baseline.py --datasets small medium --report-only
```

--------------------------------

### Exporting Full Session Bundle

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/SKILL.md

Create a comprehensive replication appendix that includes summary statistics, balance tables, regression tables, and prose in a single file.

```APIDOC
## Exporting Full Session Bundle

### Description
Creating a replication appendix that mixes summary statistics, balance tables, multiple regression tables, headings, and prose in a single file, equivalent to Stata 15 `collect`.

### API
- `c = sp.collect("Paper title", template="aer")`
- `c.add_heading("§1. Descriptives", level=...)
- `c.add_summary(df, vars=..., stats=..., labels=...)
- `c.add_balance(df, treatment=, variables=..., weights=..., test=...)
- `c.add_regression(M1, M2, ..., title="Table 2", **regtable_kwargs)
- `c.add_table(result)
- `c.add_text("Notes ...")`
- `c.save("paper.docx")` (auto-detect by extension; `.xlsx`/`.tex`/`.md`/`.html`/`.txt` all work)

### Hot kwargs for `add_regression`
`template`, `coef_labels`, `model_labels`, `panel_labels`, `dep_var_labels`, `stats`, `stars`, `add_rows`

### Example Usage
```python
c = sp.collect("My Replication Paper", template="qje")
c.add_heading("§1. Descriptive Statistics", level=1)
c.add_summary(df_descriptives, vars=["age", "income"])
c.add_heading("§2. Regression Analysis", level=1)
c.add_regression(M1, M2, title="Table 2", template="qje", drop=["Intercept"])
c.add_text("Standard errors are in parentheses.")
c.save("replication_appendix.docx")
```
```

--------------------------------

### Install Spatial Dependencies

Source: https://github.com/brycewang-stanford/statspai/blob/main/plans/2026-04-15-sp-01-s1.1-spatial-core.md

Add the 'spatial' extra to pyproject.toml to install necessary dependencies for spatial analysis. This includes geopandas, libpysal, shapely, and matplotlib.

```toml
[project.optional-dependencies]
spatial = ["geopandas>=0.14", "libpysal>=4.10", "shapely>=2.0", "matplotlib>=3.8"]

```

--------------------------------

### Build Documentation

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/joss_reviewer_guide.md

Builds the project's documentation using MkDocs with strict checking enabled. This ensures the documentation is well-formed and error-free.

```bash
python -m mkdocs build --strict
```

--------------------------------

### Install StatsPAI Python Package (Full Skill)

Source: https://github.com/brycewang-stanford/statspai/blob/main/StatsPAI_full_data_analysis_skill/README.md

Installs the complete statspai package, including dependencies for fixest, plotting, neural networks, and text analysis.

```bash
pip install "statspai[fixest,plotting,neural,text]"
```

--------------------------------

### Metalearner Estimator Example

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/choosing_ml_causal_estimator.md

Example of using the Metalearner estimator to estimate heterogeneous treatment effects (CATE). It reports the population ATE and provides per-unit CATEs.

```python
q = sp.causal_question(
    treatment='trained', outcome='wage',
    design='metalearner', estimand='CATE',
    covariates=['age', 'edu', 'exp'],
    data=df,
)
r = q.estimate(learner='dr')              # default
cate = r.underlying.model_info['cate']    # per-unit tau(x_i)
```

--------------------------------

### Wall-clock Time Comparison (pre-fix vs post-fix)

Source: https://github.com/brycewang-stanford/statspai/blob/main/benchmarks/hdfe/AUDIT.md

Table comparing the end-to-end wall-clock time for 'boottest' before and after the audit fixes, showing significant performance gains.

```text
| metric                            | pre-fix     | post-fix   |
|-----------------------------------|------------:|-----------:|
| `boottest` end-to-end             | ~3 s¹       | **~0.48 s** |
```

--------------------------------

### TMLE Estimator Example

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/choosing_ml_causal_estimator.md

Example of using the Targeted Maximum Likelihood Estimator (TMLE). This estimator is doubly robust and semiparametrically efficient, suitable for binary outcomes.

```python
q = sp.causal_question(
    treatment='trained', outcome='employed',  # binary outcome
    design='tmle',
    covariates=['age', 'edu', 'exp'],
    data=df,
)
r = q.estimate()
```

--------------------------------

### DML Estimator Example

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/choosing_ml_causal_estimator.md

Example of using the DML estimator for causal inference. This code automatically selects the appropriate sub-model based on the treatment and instrument variables.

```python
import statspai as sp
q = sp.causal_question(
    treatment='trained', outcome='wage',
    design='dml',
    covariates=['age', 'edu', 'exp', 'tenure', 'industry'],
    data=df,
)
r = q.estimate()  # auto-picks model='irm'
print(r.summary())
```

--------------------------------

### Initialize and Analyze a Social Network Graph

Source: https://github.com/brycewang-stanford/statspai/blob/main/docs/guides/social_network_analysis.md

Demonstrates initializing a graph from a built-in dataset and performing basic network analyses like summary and centrality.

```python
import statspai as sp

g = sp.karate_club()                 # Zachary (1977), 34 nodes / 78 edges
sp.network_summary(g)                # structural summary
sp.centrality(g, kind="all")         # per-node centrality table
sp.community_detection(g)            # Louvain partition
```

--------------------------------

### JAX Backend Raises Error When Unavailable

Source: https://github.com/brycewang-stanford/statspai/blob/main/benchmarks/hdfe/PHASE7_VERIFY.md

Tests that the JAX backend correctly raises a RuntimeError if JAX is not installed, providing an informative error message and installation hint.

```python
test_jax_when_unavailable_raises ... SKIPPED  (jax is installed locally)
```