statspai-skill

name: StatsPAI_skill description: Use when the user asks to run a full empirical / causal analysis in Python — by default in the style of an applied economics paper (AER / QJE / JPE / ReStud / AEJ) with DID / RD / IV / SCM / DML / matching, written-out estimating equation + identifying assumption, Table 1 / Table 2 / event-study figure / robustness gauntlet — OR in epidemiology / public health style (target-trial emulation, IPTW + g-formula + TMLE triplet, Mendelian randomization, KM/AFT survival, E-value sensitivity, STROBE/TRIPOD reporting) — OR in ML causal inference style (DML, S/T/X/R/DR meta-learners, causal forest, Dragonnet/TARNet/CEVAE, BCF, CATE distribution, policy learning, conformal causal, fairness audit, causal discovery). Also covers exporting multi-column regression tables to Word / Excel / LaTeX (Stata outreg2 / esttab / R modelsummary equivalent) and bundling an entire replication appendix into one .docx / .xlsx / .tex file. Triggers on keywords "StatsPAI", "statspai", "AER empirical analysis", "applied micro pipeline", "Table 1 balance", "event study", "first-stage F", "Oster bound", "honest_did", "spec_curve", "callaway_santanna", "dragonnet", "text as treatment", "outreg2 in Python", "regression table to Word/Excel", "sp.regtable", "sp.collect", "sp.paper_tables", "sp.feols", "summary_col", "modelsummary", "AER style table", "QJE style table", "epidemiology pipeline", "target trial emulation", "g-formula", "IPTW", "TMLE", "Mendelian randomization", "STROBE", "TRIPOD", "公共健康", "流行病学", "DML", "double machine learning", "causal forest", "meta-learner", "CATE", "conformal causal", "policy learning", "因果机器学习", "ML causal". triggers: - causal inference in python - applied microeconomics pipeline - AER empirical analysis - QJE style robustness - DID IV RD SCM - callaway_santanna - synthetic control - double machine learning - causal forest - event study plot - first stage F-statistic - Oster bound - honest_did - spec_curve - estimand-first DSL - LLM-assisted DAG discovery - text as treatment - export regression table to Word - export regression table to Excel - regression table docx - regression table xlsx - outreg2 in Python - summary_col equivalent - modelsummary equivalent - AER house style table - QJE house style table - journal template regression - Stata collect equivalent - replication bundle - sp.regtable - sp.collect - sp.paper_tables - sp.feols - sp.cite - high-dim fixed effects - two-way clustering - StatsPAI - statspai - fmt auto regression table - magnitude-adaptive coefficient formatting - mixed magnitude coefficients - sumstats by_labels - Control Treated auto labels - epidemiology pipeline - public health causal inference - target trial emulation - g-formula - IPTW marginal structural model - TMLE doubly robust - HAL-TMLE - Mendelian randomization - MR-Egger weighted median - STROBE TRIPOD reporting - E-value sensitivity - Kaplan-Meier AFT survival - 流行病学 - 公共健康 - ML causal inference - double machine learning DML - meta-learner S T X R DR - causal forest GRF - Dragonnet TARNet CEVAE - Bayesian causal forest BCF - CATE distribution - policy tree - off-policy evaluation - conformal causal prediction - fairness audit - causal discovery PC NOTEARS - 因果机器学习

StatsPAI: Agent-Native Causal Inference & AER-Style Empirical Workflow

StatsPAI is a validation-tiered Python package for causal inference and applied econometrics: one import statspai as sp, 1,100+ registered functions behind a self-describing API, and mature estimator result objects that commonly export to LaTeX / Word / Excel / BibTeX.

This skill drives StatsPAI through the canonical pipeline of an applied AER empirical paper. Each step emits a paper-ready artifact (Table 1, event-study figure, Table 2 main results, robustness panel, replication stamp).

• Source: https://github.com/brycewang-stanford/StatsPAI
• Install: pip install "statspai[fixest,plotting]" (API surface re-validated against statspai 1.19.0 — every sp.* reference, signature, and result-object attribute claim in this skill is checked by validate_api_claims.py in this folder). The bare pip install statspai is not enough for the default pipeline — see the dependency matrix below.
• Paper: JOSS submission under review; JSS materials in Paper-JSS/README.md and docs/jss_source_audit_dossier.md

Install the right extras or the documented calls will raise ImportError. Several core functions live behind optional dependency groups (verified from pyproject.toml):

You use…	Needs extra	Install	Symptom if missing
sp.feols / sp.fepois / sp.feglm (high-dim FE — the default for any y ~ x | fe regression)	fixest (pyfixest)	pip install "statspai[fixest]"	ImportError: pyfixest is required …
Any figure (sp.coefplot, sp.binscatter, event-study/RD/SCM plots, .plot())	plotting (matplotlib/seaborn)	pip install "statspai[plotting]"	ImportError on first plot
sp.dragonnet / sp.tarnet / sp.cfrnet / sp.cevae (neural causal)	neural (torch)	pip install "statspai[neural]"	ImportError: PyTorch is required …
sp.causal_text.* (text-as-treatment)	text (sentence-transformers)	pip install "statspai[text]"	ImportError on embed

A one-shot install covering the whole skill: pip install "statspai[fixest,plotting,neural,text]". sp.regtable / sp.collect / Word+Excel+LaTeX export, sp.regress, IV, RD, DID (callaway_santanna), matching, DML, meta-learners, causal forest, BCF, TMLE, and the epi stack work on the base install.

Verified skeleton (copy, then swap in your columns)

This minimal pipeline runs start-to-finish against statspai 1.19.0 (every call below was executed). It is the golden path — adapt column names / design, keep the call shapes and the unpack-then-save figure idiom. The full playbook (§−1 → §8) expands each step.

import numpy as np, pandas as pd, statspai as sp
# df has: wage, training(0/1), worker_id, firm_id, year, first_treat_year, age, edu, tenure, ...

# §1 Table 1 → Word/Excel/LaTeX
mc = sp.mean_comparison(df, ["age","edu","tenure"], group="training", test="ttest",
                        title="Table 1. Summary statistics")
mc.to_word("tables/table1.docx"); mc.to_excel("tables/table1.xlsx")

# §2 Estimand-first plan (freeze BEFORE estimating)
q = sp.causal_question(treatment="training", outcome="wage", data=df, estimand="ATT",
                       design="did", time_structure="panel", time="year", id="worker_id",
                       covariates=["age","edu","tenure"])
plan = q.identify(); print(plan.summary())

# §3 Identification figure — from a CS/SA result (NOT event_study()); plotters return (fig, ax)
cs = sp.callaway_santanna(df, y="wage", g="first_treat_year", t="year", i="worker_id", x=["age","edu"])
fig, ax = sp.enhanced_event_study_plot(cs, shade_pre=True); fig.savefig("figures/fig2a.png", dpi=300)

# §4 Main table — mix sp.regress (no FE) + sp.feols (HDFE, needs statspai[fixest]) in ONE regtable
M1 = sp.regress("wage ~ training", df, cluster="firm_id")
M2 = sp.feols("wage ~ training + age + edu + tenure | industry + year", df, vcov={"CRV1":"firm_id"})
rt = sp.regtable(M1, M2, template="aer", coef_labels={"training":"Job training"},
                 model_labels=["(1) OLS","(2) FE"], stats=["N","R2","Cluster","FE"],
                 title="Table 2. Effect of training on wages")
rt.to_word("tables/table2.docx"); rt.to_excel("tables/table2.xlsx")
open("tables/table2.tex","w").write(rt.to_latex())

# §5 Heterogeneity — per-row CATE at result.model_info["cate"] (there is NO .cate_estimates)
ml = sp.metalearner(df, y="wage", treat="training", covariates=["age","edu","tenure"], learner="dr")
fig, ax = sp.cate_plot(ml, kind="hist"); fig.savefig("figures/fig4.png", dpi=300)

# §7 Robustness — Oster + E-value + honest-DID sensitivity figure
sp.oster_bounds(data=df, y="wage", treat="training", controls=["age","edu","tenure"], r_max=1.3)
sp.evalue(estimate=M2.params["training"], ci=tuple(M2.conf_int().loc["training"]), measure="RR")
fig, ax = sp.sensitivity_plot(sp.honest_did(cs, method="smoothness"),
                              original_estimate=cs.estimate, original_ci=cs.ci)
fig.savefig("figures/fig6.png", dpi=300)

# §8 One-file replication bundle (Word/Excel/LaTeX/Markdown from one source)
c = sp.collect("Replication", template="aer")
c.add_summary(df, vars=["wage","age","edu","tenure"], stats=["mean","sd","n"], title="Table 1")
c.add_regression(M1, M2, model_labels=["(1)","(2)"], stats=["N","R2"], title="Table 2")
for ext in ("docx","xlsx","tex","md"): c.save(f"replication/paper.{ext}")

Epi (§A) and ML-causal (§B) reuse this exact scaffolding — only the §4 estimator stack changes (TMLE/g-formula/MR for epi; DML/meta-learner/causal-forest for ML), and every estimator still returns a result that drops into sp.regtable / sp.collect.

Why for Agents

1. Self-describing: sp.list_functions() / sp.describe_function(name) / sp.function_schema(name) — registered symbols are discoverable without doc lookup.
2. Structured results: mature estimators return result objects with methods such as .summary(), .plot(), .diagnostics, .to_latex(), .to_word(), .cite() when supported.
3. One import, full pipeline: data contract → Table 1 → estimand-first DSL → identification graphs → main table → heterogeneity → mechanisms → robustness → replication package.
4. Estimand-first: sp.causal_question(...).identify() forces the "DID vs RD vs IV?" decision before estimation, with the identifying assumption written down — the way a referee expects to read it.

SkillOpt-derived operating loop (read before the playbook)

SkillOpt's useful lesson for this skill is procedural, not cosmetic: a skill is a bounded decision policy that should improve from rollout evidence while preserving verified behavior. Treat every StatsPAI request as a mini rollout:

1. Route the mode first: choose Default/AER, Mode A/epi, Mode B/ML-causal, or a narrow export-only path from the user's words. Do not run the full paper pipeline when the request is only "make Table 1" or "export this regression".
2. Freeze the contract before estimating: name y, treatment/exposure, unit/time ids, estimand, design, required artifacts, and install extras. If any field is missing, infer only when the column names make the choice obvious; otherwise produce a short blocking checklist instead of hallucinating columns.
3. Start from the smallest verified call shape: prefer the skeleton and the relevant section-specific snippet over ad hoc API guesses. For an unfamiliar function, call sp.describe_function(name) / sp.function_schema(name) before writing code.
4. Widen one block at a time: data contract → plan → diagnostic figure → main estimate → robustness/export. After each block, read warnings and object attributes before passing the result downstream.
5. Gate the answer on artifacts, not intentions: final responses should list the files produced, the identifying assumption, the estimator class, and any failed or skipped gate. Never claim "paper-ready" if Word/Excel/LaTeX exports or required diagnostics were not actually generated.
6. Turn failures into bounded corrections: if a call raises, fix the smallest wrong rule (signature, result type, optional extra, plot return shape) and continue from the last verified artifact. Do not rewrite the pipeline wholesale.

Acceptance gates by request type

Maintenance rule for future skill edits

When improving this skill itself, follow a SkillOpt-style accept rule: propose a small add/delete/replace edit, then accept it only if it helps a concrete failure case and does not regress the verified skeleton, export cookbook, or Common Mistakes table. Use EVALS.md as the held-out gate set for future skill edits. Keep reusable fixes near the earliest section where an agent will need them; keep rare API traps in Common Mistakes.

The AER-style empirical pipeline

The skill mirrors the canonical sections of an applied AER / QJE / AEJ paper. Each step below is one paper section and one set of artifacts on disk.

Paper section               Step  StatsPAI moves
─────────────────────────── ───── ────────────────────────────────────────────────
Pre-Analysis Plan           −1    sp.power.* + freeze IdentificationPlan to disk
§1. Data                     0    data_contract + sample-construction log (footnote 4)
§1.1 Descriptives (Table 1)  1    sp.sumstats · sp.balance_table · sp.describe
§2. Empirical Strategy       2    write equation + identifying assumption + sp.causal_question
   (LLM-DAG addendum)        2.5  sp.llm_dag_propose · validate · constrained
§3. Identification graphics  3    event-study · first-stage F · McCrary · love plot
§4. Main Results (Table 2)   4    progressive controls + FE  (sp.regtable / sp.causal)
§5. Heterogeneity (Table 3)  5    sp.subgroup_analysis · sp.continuous_did · CATE
§6. Mechanisms               6    sp.mediation · sp.decompose
§7. Robustness gauntlet      7    placebo · Oster · honest_did · E-value · 2-way / Conley SE · spec_curve
§8. Replication package      8    .to_latex() · .plot() · reproducibility stamp

All code blocks below share one running example (training → wage, with worker_id / firm_id / year / age / edu / tenure) purely for readability. Column names, population, estimand, and design values are illustrative — substitute the user's actual columns and research question. Only sp.* function names and argument shapes are normative.

Three domain modes (default = AER econ; alternates = epi & ML-causal)

The default playbook above is AER-style applied econometrics — the AEA convention: written-out estimating equation, identifying assumption table, design horse-race, full robustness gauntlet. The skill also ships two parallel sub-pipelines for the other two big causal-inference traditions, each reusing the same export stack (sp.regtable / sp.collect / sp.paper_tables) and result objects:

How to invoke a non-default mode (Claude / agent picks this up from the user's wording):

The three modes share the same export stack, the same CausalResult interface, and the same sp.causal_question(...).identify() estimand-first DSL — switching modes only changes which Step 4 estimators you reach for, not the surrounding scaffolding. If you only want descriptive stats / Table 1 / a balance check, the AER sp.sumstats / sp.mean_comparison / sp.collect calls work in all three modes.

Paper-ready figure & table inventory (what to produce by section)

A modern AER paper has 5–7 figures and 3–5 main tables + an appendix robustness table. Every step below should leave at least one numbered artifact on disk. Default file names assume parallel .tex / .docx / .xlsx exports (the agent should produce all three so co-authors can edit in Word / Excel and the build system can use LaTeX):

Every CausalResult and OLS model can be passed straight into sp.regtable(...), sp.coefplot(...), and sp.collect(). Don't hand-roll LaTeX, and don't render Word/Excel from pandas — the export functions apply book-tab borders, AER-style stars, and the right SE label automatically.

Export cookbook — Word / Excel / LaTeX in one line

StatsPAI's export stack is the agent-native equivalent of Stata's outreg2 / esttab / collect and R's modelsummary / gtsummary. Three tiers, picked by scope of what you're exporting:

Journal templates (apply the right SE label, star levels, and notes automatically):

sp.list_journal_templates()
# → ('aer', 'qje', 'econometrica', 'restat', 'jf', 'aeja', 'jpe', 'restud')

rt = sp.regtable(M1, M2, M3, template="qje")    # QJE styling; default = full coef list (incl. intercept)
rt.to_word("table2_qje.docx")
# Opt-in filters:
#   • drop the constant only:    sp.regtable(M1, M2, M3, template="qje", drop=["Intercept"])
#   • focal-coefficient only:    sp.regtable(M1, M2, M3, template="qje", keep=["x"])

sp.get_journal_template("aer")                                 # inspect a preset
# → {'label': 'American Economic Review', 'star_levels': (0.1, 0.05, 0.01),
#    'se_label': 'Standard errors', 'stats': ('N', 'R-squared'),
#    'notes_default': ('Standard errors in parentheses.', '*** p<0.01, ** p<0.05, * p<0.10.'),
#    'font_name': 'Times New Roman'}    # note: tuples, not lists

Inline citations in prose (drop a coefficient straight into a sentence):

sp.cite(M3, "training")                  # → "1.239*** (0.153)"
sp.cite(M3, "training", output="latex")  # → "1.239^{***}~(0.153)"  (wrap in $...$ yourself)

Naming gotcha: sp.regtable(..., output="docx") is invalid — the enum is {"text", "latex", "tex", "html", "markdown", "md", "qmd", "quarto", "word", "excel"}. Use output="word" / "excel", or — simpler — drop output= and call .to_word(filename) / .to_excel(filename) on the result.

Notebook setup — CJK fonts + retina DPI

Run once at the top of every analysis script / notebook, before any matplotlib-backed plot (sp.regtable.to_* exporters do not need this — only .savefig / sp.coefplot / sp.binscatter / sp.cate_plot / etc.). Two failures it fixes in one shot:

1. CJK labels render as ▢▢▢ tofu — the matplotlib default DejaVu Sans carries no Chinese / Japanese / Korean glyphs, so ax.set_title("教育回报") silently degrades into squares.
2. Plots look fuzzy on hi-DPI displays — matplotlib's default figure.dpi=100 is half the density of a Retina / 4K screen.

Drop-in snippet

import matplotlib as mpl
import matplotlib.pyplot as plt

def setup_plot(retina: bool = True) -> None:
    """One-shot matplotlib boilerplate: CJK font fallback + retina DPI.

    Idempotent — safe to call multiple times. Call BEFORE any plotting.
    """
    # 1. CJK font fallback chain — covers macOS / Windows / Linux in one list.
    #    matplotlib uses the first available font; later names are fallbacks,
    #    so listing all three platforms is harmless on any single host.
    mpl.rcParams["font.sans-serif"] = [
        "PingFang SC", "Heiti SC", "Hiragino Sans GB",   # macOS
        "Microsoft YaHei", "SimHei", "SimSun",           # Windows
        "Noto Sans CJK SC", "Source Han Sans SC",        # Linux / Adobe
        "WenQuanYi Micro Hei",                           # Linux fallback
        "Arial Unicode MS",                              # universal fallback
        "DejaVu Sans",                                   # last-resort Latin
    ]
    mpl.rcParams["axes.unicode_minus"] = False          # 修复中文字体下负号渲染成 □

    # 2. Retina-grade DPI. figure.dpi controls on-screen / inline rendering;
    #    savefig.dpi controls .png exports. Set both — they are independent.
    if retina:
        mpl.rcParams["figure.dpi"]  = 144   # 2× default — sharp on Retina/HiDPI
        mpl.rcParams["savefig.dpi"] = 300   # manuscript/export PNG (AER house norm)
        # Jupyter inline retina backend (no-op outside IPython):
        try:
            from IPython import get_ipython
            ipy = get_ipython()
            if ipy is not None:
                ipy.run_line_magic("config", "InlineBackend.figure_format = 'retina'")
        except Exception:
            pass

setup_plot()                                            # call once at the top

Smoke test (5 seconds, run once after setup_plot())

fig, ax = plt.subplots(figsize=(4, 2.5))
ax.plot([0, 1, 2], [-1, 0, 1])
ax.set_title("中文标题测试 — Card (1995) 教育回报")
ax.set_xlabel("受教育年数 (years)")
fig.tight_layout()
fig.savefig("figures/_font_smoke_test.png", dpi=300)    # delete after verifying

If the saved PNG shows Chinese characters cleanly and the y-axis tick -1 is a real minus sign (not a square), the setup is good. Otherwise see troubleshooting below.

Saving figures — the (fig, ax) idiom (READ THIS)

Every StatsPAI plotter and every result .plot() returns a (fig, ax) tuple — NOT a bare Figure. So sp.parallel_trends_plot(...).savefig(...) raises AttributeError: 'tuple' object has no attribute 'savefig'. Always unpack, then save the figure:

fig, ax = sp.parallel_trends_plot(df, y="wage", time="year", treat="training", treat_time=2015)
fig.savefig("figures/fig1.png", dpi=300)

Two exceptions to memorize:

• sp.binscatter(...) returns a 3-tuple (fig, ax, binned_df) — fig, ax, _ = sp.binscatter(...).
• sp.kaplan_meier(...).plot() returns a bare Axes (it is a KMResult, not a CausalResult) — save via ax = km.plot(); ax.figure.savefig(...).

This applies uniformly to coefplot, binscatter, rdplot, rddensity().plot(), bacon_plot, enhanced_event_study_plot, did_summary_plot/ggdid/group_time_plot, synthdid_plot, cate_plot, cate_group_plot, dose_response().plot(), sensitivity_plot, match().plot(), synth().plot(), and a generic result.plot(). The code blocks below all use the unpack-then-save form.

Troubleshooting

Persist as project default (optional)

Drop the same rcParams into a project-level matplotlibrc next to pyproject.toml so co-authors and CI runners pick it up without calling setup_plot():

# matplotlibrc — committed to the repo
font.sans-serif: PingFang SC, Heiti SC, Microsoft YaHei, SimHei, Noto Sans CJK SC, Arial Unicode MS, DejaVu Sans
axes.unicode_minus: False
figure.dpi: 144
savefig.dpi: 300

The setup_plot() function above is the in-script fallback when a project matplotlibrc is not present.

Step −1 — Pre-Analysis Plan (pre-data; AEA RCT Registry style)

sp.power(design, n=..., effect_size=..., power_target=...) is a unified dispatcher — leave one argument None to solve for it (sample size, MDE, or power). Convenience wrappers: sp.power_rct, sp.power_did, sp.power_rd, sp.power_iv, sp.power_cluster_rct, sp.power_ols.

# Always go through the dispatcher when you want auto-solve. The
# `sp.power_<design>` wrappers (power_rct / power_did / power_rd /
# power_iv / power_cluster_rct / power_ols) accept *only* the design's
# native arguments — they will NOT solve for power_target / n / effect
# unless you go via `sp.power(design, ..., power_target=...)`.

sp.power("rct", effect_size=0.3, power_target=0.80)                  # → PowerResult(n=349, power=0.80)
sp.power("did", n=200, effect_size=0.15, power_target=0.80,
         n_periods=4, n_treated_periods=2)                            # DID: solves MDE / n / power
sp.power("cluster_rct", cluster_size=50, icc=0.05,
         effect_size=0.2, power_target=0.80)                          # Cluster RCT: solves n_clusters
# Roth (2022) pre-trends power is a POST-estimation diagnostic — it needs an estimated
# event-study result, so run it in §3 once you have `es = sp.event_study(...)`:
#   sp.pretrends_power(es)

Persist the PowerResult next to data_contract.json and empirical_strategy.md — a referee will ask whether the design was powered before data collection, not after.

Step 0 — Sample construction & data contract (Section "Data")

An AER §1 Data section has three jobs: (a) describe sources, (b) document every sample restriction (the "footnote 4" sample log), (c) lock the panel structure. StatsPAI assumes an analysis-ready DataFrame — do ETL (imputation, type coercion, merges, transforms) in pandas first, then run the 5-check contract.

0.1 Sample-construction log (footnote 4)

sample_log = []
df0 = df_raw.copy();                                       sample_log.append(("0. raw",                len(df0)))
df1 = df0.dropna(subset=["wage"]);                          sample_log.append(("1. drop missing wage",  len(df1)))
df2 = df1[df1["age"].between(18, 65)];                      sample_log.append(("2. drop age outside 18-65", len(df2)))
df3 = df2[df2["industry"].isin(MANUF_CODES)];               sample_log.append(("3. keep manufacturing", len(df3)))
df  = df3
import json; json.dump(sample_log, open("artifacts/sample_construction.json", "w"), indent=2)

Paste this log verbatim as footnote 4 of your paper. AER reviewers use it to reconstruct the analysis sample.

0.2 Five-check data contract (go / no-go gate)

import pandas as pd, numpy as np, statspai as sp

def data_contract(df, *, y, treatment, id=None, time=None, covariates=()):
    """Return a go/no-go dict. Stop the pipeline if any required check fails."""
    keys = [y, treatment] + ([id, time] if id and time else []) + list(covariates)
    c = {
        "n_obs":       len(df),                                           # 1. shape
        "dtypes":      df[keys].dtypes.astype(str).to_dict(),             # 2. dtypes on keys
        "n_missing":   df[keys].isna().sum().to_dict(),                   # 3. missing pattern
        "n_dupes_on_keys": 0,
        "panel_balanced":  None,
        "cohort_sizes":    None,
    }

    if id and time:
        c["n_dupes_on_keys"] = int(df.duplicated([id, time]).sum())       # 4. duplicate (id,time)
        balanced = sp.balance_panel(df, entity=id, time=time)              # 5. panel balance
        c["panel_balanced"]        = len(balanced) == len(df)
        c["n_dropped_by_balance"]  = len(df) - len(balanced)

        if "first_treat_year" in df.columns:                               # staggered cohorts
            c["cohort_sizes"] = (
                df.drop_duplicates(id).groupby("first_treat_year").size().to_dict()
            )

    c["y_range"]          = (float(df[y].min()), float(df[y].max()))
    c["treatment_share"]  = float(df[treatment].mean())

    # Missingness mechanism hint (Rubin): compare covariate means between
    # rows missing-on-y vs observed. Any p < 0.05 ⇒ NOT MCAR → use MI / IPW,
    # not listwise deletion.
    from scipy import stats
    miss_y = df[y].isna()
    c["mcar_hint"] = "likely MCAR (listwise OK)"
    if miss_y.any() and (~miss_y).any():
        for cov in covariates:
            if df[cov].dtype.kind in "fi":
                _, p = stats.ttest_ind(df.loc[miss_y, cov].dropna(),
                                        df.loc[~miss_y, cov].dropna(),
                                        equal_var=False)
                if p < 0.05:
                    c["mcar_hint"] = f"NOT MCAR (y-miss differs on {cov}, p={p:.3f}) → use MI / IPW"
                    break
    return c

contract = data_contract(df, y="wage", treatment="training",
                         id="worker_id", time="year",
                         covariates=["age", "edu", "tenure"])

assert contract["n_dupes_on_keys"] == 0, "duplicate (id, time) — fix before panel methods"
assert all(v == 0 for v in contract["n_missing"].values()), \
       f"NaNs on keys: {contract['n_missing']}"

If any assertion fires, stop and fix it in pandas — StatsPAI estimators silently drop NaN rows, the most common source of "mysterious sample-size shrinkage" bugs. Persist:

import json; json.dump(contract, open("artifacts/data_contract.json", "w"), indent=2, default=str)

Step 1 — Descriptive statistics (Table 1)

The signature AER Table 1 has three column blocks plus a difference column:

| | (1) Full | (2) Treated | (3) Control | (4) Δ (t-test) |

The Imbens–Rubin rule of thumb: a normalized difference |Δ| / √((s²₁+s²₀)/2) > 0.25 flags substantive imbalance and should trigger matching / reweighting before you trust an OLS comparison.

# Quick text/LaTeX preview (use sumstats `output=` for a string-only render).
# When `by=` is binary 0/1 and you don't pass `by_labels=`, sumstats auto-fills
# the panel headers as **Control / Treated** so the academic Table 1 reads
# correctly out of the box. For non-0/1 codings or different wording, pass
# `by_labels={0:"Untrained", 1:"Trained"}` (or `{"A":"Control","B":"Treated"}`).
print(sp.sumstats(df, vars=["wage","edu","exp","tenure","age"],
                  by="training", output="text"))

# AER-style balance table → Word + Excel + LaTeX in three lines.
# `mean_comparison` returns a MeanComparisonResult that exposes the full
# export chain (.to_word / .to_excel / .to_latex / .to_markdown / .to_html).
mc = sp.mean_comparison(df,
                        ["age","edu","tenure","firm_size"],
                        group="training",
                        test="ttest",
                        title="Table 1. Summary statistics by treatment status")
mc.to_word ("tables/table1_summary.docx")     # editable in Word
mc.to_excel("tables/table1_summary.xlsx")     # editable in Excel
open("tables/table1_summary.tex", "w").write(mc.to_latex())
sp.describe(df).to_markdown("references/codebook.md")              # auto-codebook

1.1 Multi-panel Table 1 (AER convention)

Group rows into Panel A: Outcomes, Panel B: Treatment intensity, Panel C: Controls, Panel D: Sample composition. The cleanest path is to push each panel into a sp.collect() bundle — one .save("file.docx") call then writes the whole multi-panel Table 1 with AER book-tab borders, in Word and Excel and LaTeX from one source.

panels = {
    "A. Outcomes":             ["wage", "log_wage", "weeks_employed"],
    "B. Treatment":            ["training", "training_hours"],
    "C. Demographic controls": ["age", "edu", "female", "married"],
    "D. Labor market":         ["tenure", "firm_size", "industry_id"],
}

c1 = sp.collect("Table 1. Summary statistics", template="aer")
for label, vs in panels.items():
    c1.add_heading(f"Panel {label}", level=2)
    c1.add_summary(df, vars=vs, stats=["mean", "sd", "n"])
c1.save("tables/table1_summary.docx")          # editable Word, AER book-tab borders
c1.save("tables/table1_summary.xlsx")          # one sheet per panel (heading drives the sheet name)
c1.save("tables/table1_summary.tex")           # multi-panel LaTeX

# Plain-text alternative (no Collection): one `sp.sumstats` per panel, concat strings.
# Useful when you only need the .tex preview without a binary export.
import io; buf = io.StringIO()
for label, vs in panels.items():
    buf.write(f"\n% Panel {label}\n")
    buf.write(sp.sumstats(df, vars=vs, by="training",
                          stats=["mean", "sd", "n"], output="latex"))
open("tables/table1_summary_flat.tex", "w").write(buf.getvalue())

1.2 Figure 1 — raw trends / treatment rollout

For DID / event-study designs, the first figure of an applied paper is almost always either (a) raw treated-vs-control means over time, or (b) the staggered rollout heat-strip showing which units are treated when. Both are one-liners:

# (a) Raw trends with vertical line at treatment start (DID Figure 1 style)
fig, ax = sp.parallel_trends_plot(df, y="wage", time="year", treat="training",
                                  treat_time=2015, ci=True,
                                  labels={"treated":"Trained", "control":"Untrained"})
fig.savefig("figures/fig1a_raw_trends.png", dpi=300)

# (b) Treatment rollout heatmap (staggered DID convention; Goodman-Bacon-friendly)
fig, ax = sp.treatment_rollout_plot(df, time="year", treat="training", id="worker_id",
                                    sort_by="first_treat_year",
                                    title="Figure 1. Treatment timing")
fig.savefig("figures/fig1b_rollout.png", dpi=300)

For matching designs, also produce a love plot of standardized differences pre/post matching (Step 3.4).

Step 2 — Empirical strategy (Section "Identification")

This is the heart of an AER paper. Before any code, write down the equation explicitly and state the identifying assumption. Vague identification language is the single most common reason a referee rejects an applied paper.

2.1 Equation × identifying assumption table

2.2 Design picker

When design="auto" is too opaque, use this decision tree:

                 ┌─ running var + cutoff ───────────────── RDD   (sp.rdrobust)
                 │
                 ├─ exogenous instrument Z ─────────────── IV    (sp.ivreg, sp.dml)
data + question ─┤
                 ├─ pre/post × treat/control ─┬ 2 periods  ── 2×2 DID (sp.did)
                 │                            └ staggered  ── CS / SA  (sp.callaway_santanna)
                 │
                 ├─ 1 treated unit + donor pool + long pre ── SCM   (sp.synth, sp.sdid)
                 │
                 ├─ high-dim X, selection-on-observables ── DML / Causal Forest
                 │
                 └─ none of the above ──────────────────── matching + E-value (sp.match, sp.evalue)

2.3 Estimand-first DSL = pre-registration

sp.causal_question declares the five-tuple (population, treatment, outcome, estimand, design) and .identify() picks the estimator with its assumptions written down. Treat the IdentificationPlan as your pre-registration artifact — freeze it before running q.estimate() so the analysis plan is a dated document, not a post-hoc rationalization.

q = sp.causal_question(
    treatment="training", outcome="wage", data=df,
    population="manufacturing workers, 2010–2020",
    estimand="ATT",
    design="auto",                 # 'auto' | 'did' | 'event_study' | 'regression_discontinuity'
                                   # | 'iv' | 'rct' | 'selection_on_observables'
                                   # | 'synthetic_control' | 'natural_experiment'
                                   # | 'policy_shock' | 'longitudinal_observational'
    time_structure="panel", time="year", id="worker_id",
    covariates=["age", "edu", "tenure"],
)
plan = q.identify()                # IdentificationPlan: estimator + assumptions + fallbacks
print(plan.summary())              # human-readable Methods paragraph
print(plan.identification_story)   # narrative of why this estimator identifies the estimand

# FREEZE the plan to disk BEFORE estimating — this is your pre-registration.
# `q` (CausalQuestion) carries the question (population / treatment / outcome).
# `plan` (IdentificationPlan) carries the strategy (estimator / story /
# assumptions / fallbacks / warnings). The estimating equation is *your*
# job to write down — paste it from the §2.1 table that matches plan.estimator.
from pathlib import Path
bullets = lambda xs: "\n".join(f"- {x}" for x in xs) if xs else "- (none)"
Path("artifacts/empirical_strategy.md").write_text(
    f"# Empirical Strategy (pre-registration)\n\n"
    f"**Population**: {q.population}\n"
    f"**Treatment**: `{q.treatment}`    **Outcome**: `{q.outcome}`\n"
    f"**Estimand**: {plan.estimand}\n"
    f"**Estimator**: `sp.{plan.estimator}`\n\n"
    f"## Estimating equation (paste from §2.1 row matching `{plan.estimator}`)\n"
    f"```\n<paste here>\n```\n\n"
    f"## Identification story\n{plan.identification_story}\n\n"
    f"## Identifying assumptions (must defend in §2)\n{bullets(plan.assumptions)}\n\n"
    f"## Auto-flagged warnings\n{bullets(plan.warnings)}\n\n"
    f"## Fallback estimators (Step 7 robustness)\n{bullets(plan.fallback_estimators)}\n"
)
# Machine-readable sidecar (full question, replayable):
Path("artifacts/causal_question.yaml").write_text(q.to_yaml())

result = q.estimate()              # run only after the plan is committed to disk / git

2.5 (Optional) LLM-assisted DAG addendum

Useful when the user wants an explicit DAG to defend in §2 or §7. Pipe the discovered DAG into sp.causal(..., dag=...).

proposal   = sp.llm_dag_propose(
    variables=df.columns.tolist(),
    domain="labor economics: training, wages, tenure",
    client=my_llm_client,                          # .complete(prompt) -> str; None = heuristic
)
validation = sp.llm_dag_validate(proposal, df, alpha=0.05)   # (dag, data) positional
print(validation.edge_evidence)

discovered = sp.llm_dag_constrained(
    df,
    descriptions={"wage": "monthly wage USD", "training": "0/1 program"},
    oracle=my_llm_client.suggest_edges,            # optional; falls back to plain PC
    max_iter=3,
)
# The result is an LLMConstrainedDAGResult — it has NO `.dag` attribute. Get a DAG with
# `.to_dag()` (or inspect `.final_edges`). Pass into Step 4 as:
#   sp.causal(..., dag=discovered.to_dag())

Step 3 — Identification graphics (Section "Identification, graphical evidence")

AER convention: the identification figure precedes the regression table. The reader should see graphical evidence that PT holds / first stage is strong / RD jumps cleanly before you ask them to trust your point estimate.

3.1 Event-study plot + numerical pre-trends test (DID identification)

Pre-period coefficients ≈ 0 (with the −1 reference period normalized to zero) is the visual evidence for parallel trends. Pair the figure with a numerical pre-trends test so reviewers don't have to eyeball it.

# --- The event-study FIGURE comes from a Callaway–Sant'Anna (or sun_abraham)
#     result, NOT from sp.event_study(). The figure plotters
#     (enhanced_event_study_plot / cs.plot() / ggdid / group_time_plot) consume a
#     CS/SA result; feeding them sp.event_study() output raises KeyError('att').
#     Use `x=` for covariates (NOT `covariates=` — that kwarg does not exist on CS).
cs = sp.callaway_santanna(df, y="wage", g="first_treat_year",
                          t="year", i="worker_id",
                          x=["age", "edu"])

# Figure 2a — dynamic ATT / event-study coefficient plot. Plotters return (fig, ax).
fig, ax = sp.enhanced_event_study_plot(
    cs, shade_pre=True,
    title="Figure 2a. Event-study coefficients (95% CI; ref. period = −1)")
fig.savefig("figures/fig2a_event_study.png", dpi=300)
# (equivalently: `fig, ax = cs.plot()` or `fig, ax = sp.ggdid(cs)` /
#  `fig, ax = sp.group_time_plot(cs)` — all consume the CS result and return (fig, ax).)

# Numerical pre-trends test (Roth 2022 power) for the table footnote. THIS is what
# sp.event_study() is for — the coefficient/pre-trend numerics, not the figure.
es = sp.event_study(df, y="wage", treat_time="first_treat_year",
                    time="year", unit="worker_id",
                    window=(-4, 4), ref_period=-1,
                    covariates=["age", "edu"])
print(sp.pretrends_summary(es))                       # F-stat, p-value, max-PT bound
# es.model_info["pretrend_test"] holds the same numbers machine-readably.

# Bacon decomposition figure for staggered DID (Figure 2a-bis)
bd = sp.bacon_decomposition(df, y="wage", treat="training",
                            time="year", id="worker_id")
fig, ax = sp.bacon_plot(bd, title="Figure 2a-bis. Goodman-Bacon weights")
fig.savefig("figures/fig2a2_bacon.png", dpi=300)

# Borusyak–Jaravel–Spiess joint pre-trends test — needs the CS/SA result
# AND the underlying panel (NOT the event_study() output):
sp.bjs_pretrend_joint(cs, df, y="wage", group="first_treat_year",
                      time="year", first_treat="first_treat_year",
                      controls=["age", "edu"])

3.2 First-stage F-statistic + scatter (IV identification)

Rule of thumb: first-stage F ≥ 10 for OLS-style inference; F ≥ 23 for AR-equivalent inference (Stock–Yogo / Lee 2022).

iv = sp.ivreg("wage ~ (training ~ Z1 + Z2) + age + edu", df, cluster="firm_id")
print(iv.summary())                                    # reports first-stage F (Cragg–Donald / KP)
fig, ax, _binned = sp.binscatter(df, y="training", x="Z1",   # binscatter → 3-tuple
                                 controls=["age", "edu"],
                                 n_bins=20, ci=True)
fig.savefig("figures/fig_first_stage.png", dpi=300)

3.3 RD: McCrary density + canonical RD plot + binscatter

The signature RD figure is sp.rdplot (CCT-style binned scatter with local-polynomial fit on each side), paired with the McCrary manipulation test. Together they answer: (a) is there a visual jump? (b) is the density continuous at the cutoff?

# Figure 2b — canonical RD plot (binned means + local poly fit on each side)
fig, ax = sp.rdplot(df, y="y", x="running_var", c=0,
                    p=4, kernel="triangular", binselect="esmv",
                    shade_ci=True, ci_level=0.95)
fig.savefig("figures/fig2b_rdplot.png", dpi=300)

# Figure 2b-bis — McCrary density (manipulation test). .plot() → (fig, ax)
fig, ax = sp.rddensity(df, x="running_var", c=0).plot()
fig.savefig("figures/fig2b2_mccrary.png", dpi=300)

# Optional: covariate-adjusted binscatter (continuity in covariates is also testable)
fig, ax, _ = sp.binscatter(df, y="age", x="running_var", n_bins=40, ci=True)
fig.savefig("figures/fig2b3_cov_binscatter.png", dpi=300)

3.4 Matching: love plot (standardized differences)

m = sp.match(df, y="wage", treat="training",
             covariates=["age", "edu", "tenure"], method="nearest")
fig, ax = m.plot()                                    # |std diff| pre vs post; target |Δ|<0.1
fig.savefig("figures/fig2c_love_plot.png", dpi=300)

3.5 SCM: synthetic-control trajectory + gap plot

For synthetic-control designs the canonical Figure 2 is the treated-vs-synthetic time-series with treatment time annotated. synthdid_plot does this in one line.

sc = sp.synth(df, outcome="y", unit="unit", time="time",
              treated_unit=1, treatment_time=2000)
fig, ax = sc.plot()                                   # treated vs synthetic + gap
fig.savefig("figures/fig2d_synth_trajectory.png", dpi=300)
sd = sp.sdid(df, outcome="y", unit="unit", time="time",
             treated_unit=1, treatment_time=2000)
fig, ax = sp.synthdid_plot(sd, title="Figure 2d. Synthetic DID")
fig.savefig("figures/fig2d2_sdid.png", dpi=300)

3.6 Generic pre-flight (identification-independent)

sp.diagnose(df, y="wage", x=["age", "edu", "tenure"])  # leverage, overlap, missing

Identification-specific checks (PT for DID, weak-IV F, density for RD, common support for matching) are also auto-run inside sp.causal(...) in Step 4 — don't duplicate the numerics here, but DO produce the figures: a referee scans the figures first.

Step 4 — Main results (multi-regression tables, AER style)

This is the densest section of an applied paper. A modern AER §4 typically contains 2–3 multi-regression tables and one coefficient plot:

• Table 2 (main): progressive controls, 4–6 columns
• Table 2-bis (design horse race): same coefficient under OLS / 2SLS / DID / DML
• Table 2-ter (multi-outcome): same treatment, several outcomes side-by-side
• Figure 3 (coefplot): visual summary of β̂ and 95% CI across specs

Estimator routing (memorize this — getting it wrong silently produces nonsense):

• No FE → sp.regress("y ~ x1 + x2", df, cluster="firm_id")
• High-dim FE → sp.feols("y ~ x1 + x2 | fe1 + fe2", df, vcov={"CRV1":"firm_id"})
• Two-way cluster → sp.feols(..., vcov={"CRV1":"firm_id+year"})
• 2SLS / IV → sp.ivreg("y ~ (x ~ z) + controls", df, cluster=...)
• DID / event-study → sp.callaway_santanna(...) / sp.sun_abraham(...)

Never write sp.regress("y ~ x | firm_id") — sp.regress does not parse | and silently treats x | firm_id as a single variable name. Use sp.feols for any formula containing |.

sp.regtable(*models, ...) is the workhorse. Useful kwargs:

keep              : list of coef names to display (e.g. ["training"])
drop              : list of coef names to suppress (controls)
model_labels      : column labels   ["(1) Baseline", "(2) +Demog", ...]
dep_var_labels    : dep-var-row labels (for multi-outcome tables)
panel_labels      : panel-A / panel-B layout for stacked tables
coef_labels       : pretty-print names for coefficients
stars             : "aer" → * 0.10 ** 0.05 *** 0.01  (or "default", "none")
stats             : footer rows ["N","R2","Cluster","FE","DV mean", ...]
output            : "latex" | "html" | "markdown" | "text"
filename          : path to write the table

4.1 Pattern A — Progressive controls (the canonical Table 2)

Stable β̂ across columns ⇒ less concern that selection on observables is driving the estimate (Oster 2019 selection-stability logic; quantified in Step 7.5). sp.regtable(*models) is the StatsPAI equivalent of Stata outreg2 / esttab and R modelsummary::msummary / summary_col — it consolidates N models into ONE table with one column per model.

# RULE: pure OLS → sp.regress; high-dim FE absorption → sp.feols
# (sp.regress does NOT parse `|` as FE — it's a thin OLS wrapper. Use
# `sp.feols("y ~ x | fe1 + fe2", df, vcov={"CRV1":"firm_id"})` for FE.)
M1 = sp.regress("wage ~ training",                                  df, cluster="firm_id")
M2 = sp.regress("wage ~ training + age + edu",                      df, cluster="firm_id")
M3 = sp.regress("wage ~ training + age + edu + tenure + firm_size", df, cluster="firm_id")
M4 = sp.feols  ("wage ~ training + age + edu + tenure + firm_size | region + industry + year",
                df, vcov={"CRV1": "firm_id"})
M5 = sp.feols  ("wage ~ training + age + edu + tenure + firm_size | worker_id + year",
                df, vcov={"CRV1": "firm_id"})

# Consolidate 5 models into ONE table (= Stata `outreg2 [M1..M5] using ..., replace`).
# **Default = show ALL coefficients verbatim — controls AND the intercept**
# (AER convention; readers verify the full spec). Pass NO `keep=`/`drop=` and
# `regtable` will surface every estimated parameter. Add `drop=["Intercept"]`
# only if you want to suppress the constant for paper aesthetics; add
# `keep=[focal]` only when a focal-coefficient-only table is intentional.
rt = sp.regtable(M1, M2, M3, M4, M5,
                 template="aer",                  # auto-applies SE label, star levels, font
                 coef_labels={"training": "Job training"},
                 model_labels=["(1) Baseline", "(2) +Demog.", "(3) +Labor-mkt",
                               "(4) Region×Ind. FE", "(5) Worker FE"],
                 stats=["N", "R2", "Cluster", "FE", "DV mean"],
                 title="Table 2. Effect of training on wages")
# Variants (all opt-in — the default above is preferred):
#   • drop intercept only:    sp.regtable(..., drop=["Intercept"])
#   • focal-coefficient only: sp.regtable(..., keep=["training"])
#   • mixed-magnitude table:  sp.regtable(..., fmt="auto")
#       Use whenever a single table mixes dollar-magnitude coefficients
#       (e.g. earnings ≈ 1500) with elasticity-magnitude coefficients
#       (e.g. log-earnings ≈ 0.09). The default fmt="%.3f" pads the dollar
#       side; a fixed fmt="%.0f" rounds the elasticity side to "0" while
#       significance stars survive — the silent LaLonde-style precision
#       trap. fmt="auto" picks per-value precision: thousands separator
#       for |β|≥1000, integer for ≥100, 1 dp for ≥10, 2 dp for ≥1, 3 dp
#       below — so neither magnitude is killed.

# Export to ALL THREE in three lines — Word for co-authors, Excel for editors, LaTeX for build:
rt.to_word ("tables/table2_main.docx")
rt.to_excel("tables/table2_main.xlsx")
open("tables/table2_main.tex", "w").write(rt.to_latex())

4.2 Pattern B — Design horse race (Table 2-bis)

Show the same coefficient of interest under multiple identification strategies. This is the AER credibility move: convergent evidence across designs each making different identifying assumptions.

ols  = sp.feols  ("wage ~ training + age + edu + tenure | industry + year",
                   df, vcov={"CRV1": "firm_id"})                                          # OLS + 2-way FE
ivr  = sp.ivreg("wage ~ (training ~ Z1 + Z2) + age + edu + tenure",
                 df, cluster="firm_id")                                                    # 2SLS
did  = sp.callaway_santanna(df, y="wage", g="first_treat_year",
                             t="year", i="worker_id",
                             x=["age","edu","tenure"])                                     # CS-DID (kwarg is x=)
dml  = sp.dml(df, y="wage", treat="training",
               covariates=["age","edu","tenure","firm_size"], model="plr")                 # DML
mtch = sp.match(df, y="wage", treat="training",
                 covariates=["age","edu","tenure"], method="nearest")                      # PSM

rt = sp.regtable(ols, ivr, did, dml, mtch,
                 template="aer",
                 coef_labels={"training": "Job training (β̂)"},
                 model_labels=["(1) OLS+FE", "(2) 2SLS", "(3) CS-DID",
                               "(4) DML-PLR", "(5) PSM"],
                 stats=["Estimator", "Identifying assumption",
                        "N", "R2 / Pseudo-R2", "Cluster"],
                 title="Table 2-bis. Convergent evidence across designs")
rt.to_word ("tables/table2b_design_race.docx")
rt.to_excel("tables/table2b_design_race.xlsx")
open("tables/table2b_design_race.tex", "w").write(rt.to_latex())

4.3 Pattern C — Multi-outcome table (same X, several Y's)

A single treatment, several outcomes. Use dep_var_labels so each column carries the Y name.

ys = ["wage", "log_wage", "weeks_employed", "left_firm", "promoted"]
multi_y = [sp.feols(f"{y} ~ training + age + edu + tenure | industry + year",
                     df, vcov={"CRV1": "firm_id"})
           for y in ys]

rt = sp.regtable(*multi_y,
                 template="aer",
                 dep_var_labels=ys,                    # column header: dep var
                 model_labels=["(1)","(2)","(3)","(4)","(5)"],
                 stats=["N","R2","DV mean","Cluster"],
                 title="Table 2-ter. Effect of training on multiple outcomes")
rt.to_word ("tables/table2c_multi_outcome.docx")
rt.to_excel("tables/table2c_multi_outcome.xlsx")
open("tables/table2c_multi_outcome.tex", "w").write(rt.to_latex())

4.4 Pattern D — Stacked Panel A / Panel B table

Same model family, two horizons (short-run / long-run) or two samples (pre-2015 / post-2015) stacked vertically. Use panel_labels.

panelA = [sp.feols("wage_t1 ~ training + X | industry + year",  df, vcov={"CRV1":"firm_id"}),
          sp.feols("wage_t1 ~ training + X | worker_id + year", df, vcov={"CRV1":"firm_id"})]
panelB = [sp.feols("wage_t5 ~ training + X | industry + year",  df, vcov={"CRV1":"firm_id"}),
          sp.feols("wage_t5 ~ training + X | worker_id + year", df, vcov={"CRV1":"firm_id"})]

rt = sp.regtable(*panelA, *panelB,
                 template="aer",
                 panel_labels=["Panel A. Short-run (1 year)",
                               "Panel A. Short-run (1 year)",
                               "Panel B. Long-run (5 years)",
                               "Panel B. Long-run (5 years)"],
                 model_labels=["(1) Industry FE","(2) Worker FE"]*2,
                 stats=["N","R2"],
                 title="Table 2-quater. Short- vs long-run effects")
rt.to_word ("tables/table2d_horizons.docx")
rt.to_excel("tables/table2d_horizons.xlsx")
open("tables/table2d_horizons.tex", "w").write(rt.to_latex())

4.5 Pattern E — IV reporting triplet (first-stage / reduced-form / 2SLS)

The textbook AER IV table presents the first stage, the reduced form, and the 2SLS in three columns so the reader can verify Wald-ratio = RF / FS.

Trap: sp.ivreg does not absorb | fe and does not parse C(fe) — it silently drops a | industry + year term (identical β̂ with or without it), so a 2SLS column written that way would not control for the FE the first-stage/reduced-form columns absorb. Keep the IV triplet on the same low-dim control set in all three columns; to control for fixed effects in a 2SLS, pre-build dummy columns in pandas and add them explicitly, or partial the FE out first.

fs = sp.feols("training ~ Z + age + edu", df, vcov={"CRV1":"firm_id"})   # 1st stage
rf = sp.feols("wage     ~ Z + age + edu", df, vcov={"CRV1":"firm_id"})   # reduced form
iv = sp.ivreg("wage ~ (training ~ Z) + age + edu", df, cluster="firm_id")  # 2SLS (same controls)

rt = sp.regtable(fs, rf, iv,
                 template="aer",
                 keep=["Z", "training"],               # IV triplet is intentionally focal:
                                                       # show only Z + endog so the reader can
                                                       # eyeball Wald-ratio = RF / FS. For the
                                                       # full coef list, drop the kwarg entirely.
                 dep_var_labels=["training", "wage", "wage"],
                 model_labels=["(1) First stage", "(2) Reduced form", "(3) 2SLS"],
                 stats=["First-stage F", "N", "R2", "Cluster"],
                 title="Table 2-quinto. IV reporting triplet")
rt.to_word ("tables/table2e_iv_triplet.docx")
rt.to_excel("tables/table2e_iv_triplet.xlsx")
open("tables/table2e_iv_triplet.tex", "w").write(rt.to_latex())

4.6 Pattern F — Causal-design main via sp.causal(...)

For DID / IV / RD / SCM mains, the sp.causal(...) orchestrator returns a CausalResult plus diagnostics and an automatic robustness preview. Pipe .result into regtable:

w = sp.causal(df, y="wage", treatment="training",
              id="worker_id", time="year", design="did",
              covariates=["age", "edu", "tenure"],
              dag=discovered.to_dag())             # optional (LLMConstrainedDAGResult.to_dag())
print(w.diagnostics)                               # PT verdict + warnings
print(w.recommendation)                            # which estimator + why
print(w.result.summary())                          # point estimate + cluster-robust SE + CI
print(w.robustness_findings)                       # automated robustness battery preview

4.7 Figure 3 — coefficient plot of the main table

Replace one of the wall-of-numbers tables with a coefplot in the body, push the table to the appendix. Modern AER papers increasingly do this.

fig, ax = sp.coefplot(M1, M2, M3, M4, M5,
                      model_names=["(1)","(2)","(3)","(4)","(5)"],
                      variables=["training"],
                      title="Figure 3. β̂ on training across specifications (95% CI)",
                      alpha=0.05)
fig.savefig("figures/fig3_coefplot.png", dpi=300)

Reporting checklist for the Table 2 footnote (AER house style)

• Standard-error cluster level (and whether it's two-way / Conley)
• Fixed-effects absorbed — regtable auto-adds one footer row per FE name (e.g. Industry FE: Yes / Year FE: Yes / Worker_id FE: No) whenever any column comes from sp.feols(... | fe1 + fe2 ...). Don't hand-roll these rows.
• Sample size and number of clusters
• Estimator (OLS / 2SLS / CS-DID / SCM / DML)
• Stars convention * 0.10 ** 0.05 *** 0.01
• Mean of dependent variable in the estimation sample (so β̂ can be read as a % of the base rate)

Step 5 — Heterogeneity (Table 3 + Figure 4)

The AER §5 Heterogeneity combines (a) a subgroup regression table with one column per subgroup (binary moderators + interaction terms), and (b) a CATE / dose-response figure for continuous moderators. Both should appear; they answer different questions.

5.1 Pattern G — Subgroup regtable (Table 3)

One column per subgroup, with the same specification re-run on each slice. Clean, easy to read, expected by referees.

slices = {
    "(1) All":        df,
    "(2) Female":     df[df["female"] == 1],
    "(3) Male":       df[df["female"] == 0],
    "(4) Low skill":  df[df["skill_quartile"].isin([1, 2])],
    "(5) High skill": df[df["skill_quartile"].isin([3, 4])],
    "(6) Small firm": df[df["firm_size"] < 100],
    "(7) Large firm": df[df["firm_size"] >= 100],
}
gmodels = [sp.feols("wage ~ training + age + edu + tenure | industry + year",
                     d, vcov={"CRV1": "firm_id"}) for d in slices.values()]

rt = sp.regtable(*gmodels,
                 template="aer",
                 coef_labels={"training": "Training"},
                 model_labels=list(slices),
                 stats=["N","R2","DV mean"],
                 title="Table 3. Heterogeneous effects of training")
rt.to_word ("tables/table3_heterogeneity.docx")
rt.to_excel("tables/table3_heterogeneity.xlsx")
open("tables/table3_heterogeneity.tex", "w").write(rt.to_latex())

5.2 Interaction-form heterogeneity (alternative Table 3)

Test moderation formally with interaction terms — referees often ask whether the gap between subgroups is statistically significant, which requires the interaction p-value.

H1 = sp.feols("wage ~ training*female + age + edu + tenure | industry + year",
              df, vcov={"CRV1": "firm_id"})
H2 = sp.feols("wage ~ training*C(skill_quartile) + age + edu + tenure | industry + year",
              df, vcov={"CRV1": "firm_id"})
H3 = sp.feols("wage ~ training*log_firm_size + age + edu + tenure | industry + year",
              df, vcov={"CRV1": "firm_id"})

rt = sp.regtable(H1, H2, H3,
                 template="aer",
                 keep=["training", "training:female", # interaction-form heterogeneity
                       "training:C(skill_quartile)[T.2]",   # is intentionally focal:
                       "training:C(skill_quartile)[T.3]",   # only the main effect + interactions
                       "training:C(skill_quartile)[T.4]",   # are reported. Drop this kwarg
                       "training:log_firm_size"],           # entirely to show full controls.
                 model_labels=["(1) ×Female", "(2) ×Skill quartile", "(3) ×log(Firm size)"],
                 stats=["N","R2"],
                 title="Table 3-bis. Interaction-form heterogeneity")
rt.to_word ("tables/table3b_interactions.docx")
rt.to_excel("tables/table3b_interactions.xlsx")
open("tables/table3b_interactions.tex", "w").write(rt.to_latex())

5.3 Figure 4 — dose-response (continuous treatment)

dr = sp.dose_response(df, y="wage", treat="training_hours",
                      covariates=["age","edu","tenure","firm_size"],
                      n_dose_points=20)
fig, ax = dr.plot(title="Figure 4a. Dose-response: training hours → wage")
fig.savefig("figures/fig4a_dose_response.png", dpi=300)

# DID-flavored continuous treatment (de Chaisemartin–D'Haultfœuille):
fig, ax = sp.continuous_did(df, y="wage", dose="training_hours",
                            time="year", id="worker_id").plot()
fig.savefig("figures/fig4a2_continuous_did.png", dpi=300)

5.4 Figure 4-bis — CATE distribution (DR-Learner / causal forest)

The CATE plotters read per-row conditional effects out of the result's model_info["cate"] array. There is no .cate_estimates attribute — the raw per-row CATE vector lives at ml.model_info["cate"] (an ndarray of length n), and summary stats at model_info["cate_mean"] / cate_q25 / cate_q75 / .... sp.causal_forest returns a summary result that does not populate model_info["cate"], so for the CATE histogram and grouped bar chart use a meta-learner (or any DR-/X-/R-learner) and pass its result to the plotters.

ml = sp.metalearner(df, y="wage", treat="training",
                    covariates=["age","edu","tenure","firm_size"], learner="dr")

# Raw per-row CATE vector (if you need the numbers, not just the figure):
cate_i = ml.model_info["cate"]                        # ndarray, length n  (NOT ml.cate_estimates)

fig, ax = sp.cate_plot(ml, kind="hist",
                       title="Figure 4b. Distribution of conditional ATE")
fig.savefig("figures/fig4b_cate_hist.png", dpi=300)

# CATE by group bar chart: first compute the group-level table, THEN plot it.
# `cate_group_plot` takes a DataFrame (from cate_by_group), not the result object.
g = sp.cate_by_group(ml, df, by="skill_quartile", n_groups=4)
fig, ax = sp.cate_group_plot(g, title="Figure 4c. CATE by skill quartile")
fig.savefig("figures/fig4c_cate_by_group.png", dpi=300)

# Tabular summary for the appendix
print(sp.cate_summary(ml))
print(g)                                              # group-level CATE table

5.5 Subgroup-analysis dispatcher (one-liner)

sp.subgroup_analysis(df, formula="wage ~ training + age + edu + tenure",
                     x="training",
                     by={"gender": "female", "skill": "skill_quartile"},
                     robust="hc1")                 # quick subgroup β̂ table (HC1 by default; no cluster arg)

For continuous moderators or many subgroups, prefer:

• sp.continuous_did(...) — dose-response under DID
• sp.metalearner(..., learner="dr") + sp.cate_plot / sp.cate_by_group — DR-Learner CATE (recommended for plotting)
• sp.causal_forest(formula="wage ~ training | X", data=df) — CATE summary only (does not populate model_info["cate"]; use a meta-learner for per-row CATEs)

Step 6 — Mechanisms / channels

sp.mediation(df, y="wage", d="training", m="hours_worked",
             X=["age", "edu", "tenure"])           # ACME / ADE / total effect
sp.decompose(...)                                   # Oaxaca-Blinder / RIF / FFL / KOB

Step 7 — Robustness gauntlet (the AER referee gauntlet)

The seven canonical robustness blocks of an applied paper. A modern AER paper expects most of these in the body or appendix — assemble a Table A1-style robustness panel from the outputs.

7.1 Placebo tests

sp.rdplacebo(df, y="y", x="running_var", c=0,
             placebo_cutoffs=[-2, -1, 1, 2])                      # RD: fake cutoffs
sp.synth_time_placebo(df, outcome="y", unit="unit", time="time",
                      treated_unit=1, treatment_time=2000,
                      n_placebo_times=10)                          # SCM in-time placebo
sp.synthdid_placebo(...)                                           # SDID placebo
# For DID: re-run with a fake treat year before actual treatment and confirm β̂ ≈ 0.

7.2 Alternative samples

result_no_outliers = sp.causal(df.query("wage < wage.quantile(0.99)"), ...)
result_drop_early  = sp.causal(df.query("first_treat_year > 2008"),  ...)
result_balanced    = sp.causal(sp.balance_panel(df, entity="worker_id", time="year"), ...)

7.3 Alternative specifications (spec curve)

sp.spec_curve(df, y="wage", x="training",
              controls=[["age"], ["age", "edu"], ["age", "edu", "tenure"]],
              subsets={"all": None, "manuf": df["industry"].eq("manufacturing")})

7.4 Alternative standard errors

Cluster-level choice is itself a robustness check — show the result is not driven by an over-narrow cluster.

# For statsmodels-backed sp.regress / sp.ivreg results:
sp.twoway_cluster(M3, df, cluster1="firm_id", cluster2="year")     # two-way clustering
sp.conley(M3, df, lat="lat", lon="lon",
          dist_cutoff=100, kernel="uniform")                        # spatial HAC (Conley 1999)

# For pyfixest-backed sp.feols results, set 2-way cluster directly in `vcov`:
sp.feols("y ~ x | firm_id + year", df,
         vcov={"CRV1": "firm_id+year"})                              # 2-way: firm × year

7.5 Oster (2019) selection bound

"How big would unobserved selection have to be for β to flip sign / vanish?" The Oster δ tells you whether the bound on selection on unobservables, relative to selection on observables, has to exceed an implausible value to overturn the result.

sp.oster_bounds(data=df, y="wage", treat="training",
                controls=["age", "edu", "tenure"],
                r_max=1.3)                          # β* assuming δ=1, R̃²=1.3·R²
# `oster_delta` uses x_base / x_controls (NOT treat= / controls=):
sp.oster_delta(data=df, y="wage",
               x_base=["training"],                 # treatment(s) of interest
               x_controls=["age", "edu", "tenure"], # observed controls
               r_max=1.3)                           # δ for which β=0

7.6 Honest DID — Rambachan–Roth (2023) PT sensitivity

honest_did only consumes a CS / SA / did_multiplegt event-study result (or aggte(result, type='dynamic')). Pass the cs object built in §3.1, not a generic OLS/FE main-table result:

sp.honest_did(cs, method="smoothness")              # bound β under bounded PT violation

7.7 E-value & unified sensitivity (unmeasured confounding)

sp.evalue(estimate=result.params["training"],       # E-value takes point + CI, NOT result
          ci=tuple(result.conf_int().loc["training"]),
          measure="RR")
sp.unified_sensitivity(result, r2_treated=0.05,
                       r2_controlled=0.10,
                       include_oster=True)          # Cinelli-Hazlett + Oster combined
sp.sensitivity_dashboard(result)                    # one-page sensitivity figure

7.8 RD-specific bandwidth / kernel sensitivity

sp.rdbwsensitivity(df, y="y", x="running_var", c=0,
                    bw_grid=[0.5, 1.0, 1.5, 2.0])   # is β̂ stable across bandwidths?

7.9 TWFE diagnostic (staggered DID)

Goodman-Bacon decomposition flags when the TWFE estimate is contaminated by forbidden 2×2's (already-treated as control).

sp.bacon_decomposition(df, y="y", treat="training",
                       time="year", id="worker_id")

7.10 Sequential confounder blocks (Oster-style robustness table)

blocks = {
    "M1 base":           [],
    "M2 +demographics":  ["age", "edu"],
    "M3 +labor-market":  ["age", "edu", "tenure", "firm_size"],
    "M4 +psychosocial":  ["age", "edu", "tenure", "firm_size", "motivation"],
}
models = [sp.regress(f"wage ~ training + {' + '.join(c) or '1'}",
                     df, cluster="firm_id")
          for c in blocks.values()]
rt = sp.regtable(*models,
                 template="aer",
                 model_labels=list(blocks),
                 title="Table 7. Selection-stability across confounder blocks")
rt.to_word ("tables/table_robust_blocks.docx")
rt.to_excel("tables/table_robust_blocks.xlsx")
open("tables/table_robust_blocks.tex", "w").write(rt.to_latex())

7.11 Pattern H — Robustness master table (Table A1, one row per check)

The canonical AER appendix Table A1 stacks every robustness specification next to the baseline so reviewers see at a glance that β̂ survives. sp.regtable accepts any mix of EconometricResults / CausalResult, so build the list dynamically:

baseline = sp.feols("wage ~ training + age + edu + tenure | industry + year",
                     df, vcov={"CRV1": "firm_id"})

rob = {
    "(1) Baseline":            baseline,
    "(2) Drop top 1% wage":    sp.feols("wage ~ training + age + edu + tenure | industry + year",
                                        df.query("wage < wage.quantile(0.99)"),
                                        vcov={"CRV1": "firm_id"}),
    "(3) Balanced panel":      sp.feols("wage ~ training + age + edu + tenure | industry + year",
                                        sp.balance_panel(df, entity="worker_id", time="year"),
                                        vcov={"CRV1": "firm_id"}),
    "(4) Drop early cohorts":  sp.feols("wage ~ training + age + edu + tenure | industry + year",
                                        df.query("first_treat_year > 2008"),
                                        vcov={"CRV1": "firm_id"}),
    "(5) Worker FE":           sp.feols("wage ~ training + age + edu + tenure | worker_id + year",
                                        df, vcov={"CRV1": "firm_id"}),
    "(6) 2-way cluster":       sp.feols("wage ~ training + age + edu + tenure | industry + year",
                                        df, vcov={"CRV1": "firm_id+year"}),  # 2-way: firm × year
    # sp.conley needs a STATSMODELS-backed result (sp.regress/sp.ivreg) — it raises
    # KeyError on a pyfixest feols result. Re-fit the spec via sp.regress for this row.
    "(7) Conley spatial SE":   sp.conley(sp.regress("wage ~ training + age + edu + tenure",
                                                    df, cluster="firm_id"),
                                         df, lat="lat", lon="lon", dist_cutoff=100),
    "(8) Log outcome":         sp.feols("log_wage ~ training + age + edu + tenure | industry + year",
                                        df, vcov={"CRV1": "firm_id"}),
    "(9) IHS outcome":         sp.feols("ihs_wage ~ training + age + edu + tenure | industry + year",
                                        df, vcov={"CRV1": "firm_id"}),
    "(10) PSM-weighted":       sp.match(df, y="wage", treat="training",
                                         covariates=["age","edu","tenure","firm_size"],
                                         method="nearest"),
    "(11) Entropy balance":    sp.ebalance(df, y="wage", treat="training",
                                            covariates=["age","edu","tenure","firm_size"]),
    "(12) DML-PLR":            sp.dml(df, y="wage", treat="training",
                                       covariates=["age","edu","tenure","firm_size"], model="plr"),
}

# Robustness master = AER Table A1 — readers MUST see every coefficient
# across every spec to verify nothing is hiding behind `keep=`. Default to
# the full coef table (intercept included); only switch to
# `keep=["training"]` if a referee has explicitly asked for a focal-only
# summary, or add `drop=["Intercept"]` if you want the constant suppressed.
rt = sp.regtable(*rob.values(),
                 template="aer",
                 coef_labels={"training": "Training (β̂)"},
                 model_labels=list(rob),
                 stats=["N", "R2", "Cluster", "FE"],
                 title="Table A1. Robustness of the main estimate")
rt.to_word ("tables/tableA1_robustness.docx")
rt.to_excel("tables/tableA1_robustness.xlsx")
open("tables/tableA1_robustness.tex", "w").write(rt.to_latex())

# Equivalent one-shot via the paper-format multi-panel API — produces a
# single .docx / .xlsx that you can hand a co-author, with main + robustness
# (+ heterogeneity / placebo if you have them) auto-laid-out per AER style:
sp.paper_tables(main=[M1, M2, M3, M4, M5],
                robustness=list(rob.values()),
                template="aer",
                coef_labels={"training": "Training"},
                model_labels_main=["(1)","(2)","(3)","(4)","(5)"],
                model_labels_robustness=list(rob),
                # paper_tables only accepts `keep=`, not `drop=`. Omit both to
                # show every coefficient (AER convention). Pass `keep=["training"]`
                # only when a focal-only summary is desired.
                ).to_docx("tables/paper_tables.docx")

7.12 Figure 5 — coefficient forest plot of all robustness specs

A single visual summary that an AER referee can parse in 5 seconds: every β̂ and 95% CI on one axis. Confirms the estimate is not knife-edge.

fig, ax = sp.coefplot(*rob.values(),
                      model_names=list(rob),
                      variables=["training"],
                      title="Figure 5. β̂ on training across robustness specifications",
                      alpha=0.05)
fig.savefig("figures/fig5_robustness_forest.png", dpi=300)

7.13 Figure 5-bis — spec curve

The Simonsohn et al. (2020) specification curve plots β̂ across every combination of {controls × subsamples × outcome transforms × SE types}. Useful when you want to head off "what about specification X?" referee letters.

# se_types accepts only: 'nonrobust', 'hc1' (alias 'robust'), 'cluster' (needs cluster_var).
# y_transforms is a DICT {name: callable} — NOT a list of strings.
sc = sp.spec_curve(df, y="wage", x="training",
                   controls=[["age"], ["age","edu"], ["age","edu","tenure"],
                             ["age","edu","tenure","firm_size"]],
                   se_types=["nonrobust", "robust", "cluster"],   # 'cluster' uses cluster_var below
                   y_transforms={"level": lambda s: s,
                                 "log":   np.log,
                                 "ihs":   np.arcsinh},
                   subsets={"all": None,
                            "manuf":  df["industry"].eq("manufacturing"),
                            "no99":   df["wage"] < df["wage"].quantile(0.99)},
                   cluster_var="firm_id")
fig, ax = sc.plot(title="Figure 5-bis. Specification curve")
fig.savefig("figures/fig5b_spec_curve.png", dpi=300)

7.14 Figure 6 + sensitivity dashboard

sp.unified_sensitivity(...) and sp.sensitivity_dashboard(...) both return a text/numeric SensitivityDashboard (Cinelli–Hazlett + Oster + Rosenbaum + E-value). It is not a figure — it has .summary() and numeric attributes (.e_value_point, .e_value_ci, .oster, .rosenbaum, .sensemakr, .breakdown), no .plot() / .savefig() / .results. The sensitivity figure (sp.sensitivity_plot) is a Rambachan–Roth honest-DID plot and consumes the DataFrame returned by sp.honest_did(...) (columns M / ci_lower / ci_upper / rejects_zero).

# (a) Numeric sensitivity dashboard — print the summary, read the attributes.
dash = sp.unified_sensitivity(baseline, r2_treated=0.05, r2_controlled=0.10,
                              include_oster=True)
print(dash.summary())                                 # one-page text dashboard
print(dash.e_value_point, dash.e_value_ci)            # numeric fields for the §7 prose
# sp.sensitivity_dashboard(baseline).summary() is the auto-dimensioned variant.

# (b) Sensitivity FIGURE (honest-DID PT sensitivity) — needs a CS/SA event-study
#     result (`cs` from §3.1) and its honest_did() DataFrame.
sens_df = sp.honest_did(cs, method="smoothness")      # → DataFrame (M, ci_lower, ci_upper, ...)
fig, ax = sp.sensitivity_plot(sens_df,
                              original_estimate=cs.estimate,
                              original_ci=cs.ci,
                              title="Figure 6. Sensitivity to PT violations (Rambachan–Roth)")
fig.savefig("figures/fig6_sensitivity.png", dpi=300)

7.15 One-stop robustness reporter

sp.diagnose_result(result)                          # PT / weak-IV / overlap / leverage verdict
sp.robustness_report(df, formula="wage ~ training + age + edu",
                     x="training", cluster_var="firm_id")
sp.estat(result, test="all")                        # Stata-style postestimation battery

Step 8 — Replication package

The agent's job at §8 is to produce a single artifact a co-author can open in Word, Excel, or LaTeX without further StatsPAI calls. There are three packaging tiers, picked by what you need to ship:

8.1 Per-result export (one estimator → one Word/Excel file)

result.to_docx("tables/main_result.docx",
               title="Table 2. Main result")          # CausalResult → .docx
result.to_latex(caption="Main result", label="tab:main")
fig, ax = result.plot()                               # publication-quality figure → (fig, ax)
fig.savefig("figures/main.png", dpi=300)
print(sp.cite(result, "training"))                    # → "1.239*** (0.153)"  ← inline citation

8.2 Per-table export (already covered in Steps 4 / 5 / 7)

Every sp.regtable(*models) returns a RegtableResult with .to_word() / .to_excel() / .to_latex() / .to_markdown() / .to_html(). Use these in §4–§7 so that by the time you reach §8 the tables/ folder already has parallel .docx / .xlsx / .tex for every numbered table.

8.3 Multi-panel paper-format (Tier 2 — Tables 2 + 3 + A1 + A2 in one file)

sp.paper_tables(
    main          = [M1, M2, M3, M4, M5],            # → "Table 2. Main results"
    heterogeneity = [g_full, g_fem, g_male],         # → "Table 3. Heterogeneity"
    robustness    = list(rob.values()),              # → "Table A1. Robustness"
    placebo       = [pb1, pb2],                      # → "Table A2. Placebo tests"
    template      = "aer",
    coef_labels   = {"training": "Training"},
    keep          = ["training"],
).to_docx("replication/paper_tables.docx")           # → 4 panels in one .docx
# .to_xlsx(...) writes one sheet per panel; .to_latex(...) one .tex with section breaks.

8.4 Full session bundle (Tier 3 — the Stata collect equivalent)

The single most efficient §8 deliverable: descriptives + balance + main + heterogeneity + robustness + prose in one Word file. sp.collect() is the agent-native counterpart of Stata 15's collect and R's gtsave.

c = sp.collect("Effect of Training on Wages — Replication", template="aer")

c.add_heading("§1. Descriptive statistics", level=1)
c.add_summary(df, vars=["wage","age","edu","tenure"],
              stats=["mean","sd","n"],
              title="Table 1. Summary statistics")
c.add_balance(df, treatment="training",
              variables=["age","edu","tenure","firm_size"],
              title="Table 1b. Balance by treatment")

c.add_heading("§4. Main results",        level=1)
c.add_regression(M1, M2, M3, M4, M5,
                 model_labels=["(1)","(2)","(3)","(4)","(5)"],
                 stats=["N","R2","Cluster","FE"],
                 title="Table 2. Effect of training on wages")

c.add_heading("§5. Heterogeneity",       level=1)
c.add_regression(*gmodels,
                 model_labels=list(slices),
                 title="Table 3. Heterogeneous effects")

c.add_heading("§7. Robustness",          level=1)
c.add_regression(*rob.values(),
                 model_labels=list(rob),
                 title="Table A1. Robustness")

c.add_text(
    "Standard errors clustered at the firm level. *** p<0.01, ** p<0.05, * p<0.10. "
    "Sample restrictions and full variable definitions are documented in "
    "artifacts/sample_construction.json and artifacts/data_contract.json.",
    title="Notes",
)

# One artifact, three formats — auto-detected from the path extension:
c.save("replication/paper.docx")   # editable Word, page-break between tables
c.save("replication/paper.xlsx")   # one sheet per add_*() item
c.save("replication/paper.tex")    # multi-section LaTeX
c.save("replication/paper.md")     # GitHub-flavoured Markdown for the README

Inspect the bundle before saving:

print(c)                # → <Collection title='...' template='aer' items=8 kinds=['heading','summary','balance',...]>
print(c.list())         # DataFrame with name / kind / title for every item

8.5 Reproducibility stamp

A CausalResult (from DID / CS / IV-causal / DML / TMLE / …) exposes .estimate (scalar), .ci (tuple), .estimand, and .n_obs — it has no .conf_int(), and .data_info's key is "nobs", not "n_obs". An EconometricResults (regress / feols / ivreg) instead exposes .params[name] and .conf_int().loc[name] — use that branch for an OLS/FE main result.

import json
ci = result.ci                                    # CausalResult: (lo, hi) tuple
json.dump({
    "statspai":          sp.__version__,
    "seed":              42,
    "n_obs":             int(result.n_obs),
    "estimand":          result.estimand,
    "estimate":          float(result.estimate),
    "ci95":              [float(ci[0]), float(ci[1])],
    # Econometric (feols/regress) main result instead:
    #   "estimate": float(M.params["training"]),
    #   "ci95":     list(M.conf_int().loc["training"]),
    "pre_registration":  "artifacts/empirical_strategy.md",
    "data_contract":     "artifacts/data_contract.json",
    "sample_log":        "artifacts/sample_construction.json",
    "paper_bundle":      "replication/paper.docx",
}, open("artifacts/result.json", "w"), indent=2)

For full-draft generation (abstract + methods + results + bibliography), see sp.paper(result, ...) — out of scope for this skill; call it only when the user explicitly asks for a paper draft.

Regtable cookbook (one-page recipe index)

sp.regtable(*models, ...) is the single primitive behind every multi-regression table in an AER paper. The eight patterns above map to:

Default sp.regtable settings for AER house style — and the export pipeline (produce .docx + .xlsx + .tex from the same RegtableResult):

rt = sp.regtable(*models,
                 template="aer",                  # journal preset: aer/qje/econometrica/restat/jf/aeja/jpe/restud
                 # AER convention: pass NEITHER `keep=` NOR `drop=` —
                 # `regtable` will then surface every estimated parameter
                 # (controls AND the intercept). Add `drop=["Intercept"]`
                 # only if you want the constant suppressed; add
                 # `keep=[focal]` only for an intentional focal-only table.
                 coef_labels={"training": "Training"},
                 model_labels=[...],              # column labels
                 stats=["N", "R2", "Cluster", "FE", "DV mean"],
                 title="Table N. ...")

# One-call exports — never hand-roll Word/Excel from pandas:
rt.to_word ("tables/tableN.docx")                  # editable Word, AER book-tab borders
rt.to_excel("tables/tableN.xlsx")                  # editable Excel, one sheet
open("tables/tableN.tex", "w").write(rt.to_latex()) # LaTeX for the build
print(rt.to_text())                                 # quick terminal preview

For pyfixest-style native output, sp.etable(*models, ...) is the alternative; for stacking many tables in one .docx, use sp.paper_tables(...) (Tier 2) or sp.collect() (Tier 3) — see Step 8.

Figure factory (the 12 standard AER figures)

Every plotting function above accepts ax= so panels can be combined with matplotlib subplots, and returns a (fig, ax) tuple — unpack it and call fig.savefig(path, dpi=300) for publication output (see "Saving figures — the (fig, ax) idiom" above). sp.binscatter returns (fig, ax, binned_df); sp.kaplan_meier(...).plot() returns a bare Axes (use ax.figure.savefig(...)).

§A. Epidemiology / public health pipeline (Mode A)

Convention: STROBE (observational) / TRIPOD-AI (prediction) reporting. A modern epidemiology reference design is target-trial emulation (Hernán & Robins): write the protocol of the hypothetical RCT first, then emulate it with observational data using a doubly-robust estimator. Outcomes are commonly risk differences, risk ratios, hazard ratios, or restricted mean survival time, not just OLS coefficients. The skill mirrors the AER 8-section flow but swaps the Step-4 estimator stack and adds survival/MR-specific reporting rows.

Running example: statin_initiation → 5-yr_MACE in an EHR cohort (patient_id / index_date / age / sex / ldl_baseline / comorbidity_index / followup_days / event). The exposure is time-varying, confounders are time-varying, and competing-risk censoring matters — the canonical setting where naïve OLS / Cox-with-baseline-adjustment is biased.

A.0 Cohort construction & target-trial protocol

import statspai as sp

# Eligibility, treatment-strategy, time-zero, follow-up, outcome — written down BEFORE estimation.
# NOTE the two validated enum fields:
#   • assignment    ∈ {"randomization", "observational emulation"}
#   • causal_contrast ∈ {"ITT", "per-protocol", "as-treated", "observational-analogue"}
# (free-text in these two fields raises ValueError). Put the prose description in `notes=`.
protocol = sp.target_trial.TargetTrialProtocol(
    eligibility           = "adults 40-75, LDL ≥ 130, no prior MI/stroke, no statin in 12mo washout",
    treatment_strategies  = ["initiate statin within 30d of index", "no statin within 30d"],
    assignment            = "observational emulation",       # enum — not free text
    time_zero             = "index_date (first eligible cardiology visit)",
    followup_end          = "first MACE / death / disenrollment / index_date + 5yr",
    outcome               = "first MACE (composite: MI, stroke, cardiovascular death)",
    causal_contrast       = "per-protocol",                  # enum — not free text
    analysis_plan         = "IPTW-MSM + g-formula + TMLE triplet; report all three with CIs",
    baseline_covariates   = ["age","sex","ldl_baseline","comorbidity_index","smoker"],
    time_varying_covariates = ["ldl_current"],
    notes                 = "emulate randomization via IPTW + g-formula; 5-yr risk difference",
)
# Signature: target_trial_emulate(protocol, data, outcome_col, treatment_col,
#                                 time_zero_filter=None, weights=None) -> TargetTrialResult.
# Eligibility is applied as `data.query(protocol.eligibility)` UNLESS you pass a
# `time_zero_filter` callable (which then defines the eligible/time-zero rows and
# lets `eligibility` stay human-readable prose). Use the callable for non-query-able rules:
cohort_res = sp.target_trial_emulate(
    protocol, df, outcome_col="mace", treatment_col="statin_initiation",
    time_zero_filter=lambda d: d["age"].between(40, 75) & (d["ldl_baseline"] >= 130),
)
cohort = df   # downstream estimators run on the eligible analysis frame you constructed

A.1 Table 1 — baseline characteristics by exposure

# Same sumstats stack as AER mode; binary 0/1 by= auto-renders Control/Treated.
mc = sp.mean_comparison(cohort, ["age","sex","ldl_baseline","comorbidity_index","smoker"],
                        group="statin_initiation", test="ttest",
                        title="Table 1. Baseline characteristics by statin initiation")
mc.to_word ("tables/table1_epi.docx")
mc.to_excel("tables/table1_epi.xlsx")

A.2 Identification — DAG, propensity overlap, KM curves

# 2.1 DAG (manual or LLM-assisted). sp.dag(spec) parses an edge STRING
# ("A -> B; C -> B"); build edges with the string spec or chained .add_edge(parent, child)
# (singular — there is no .add_edges). Back-door sets come from .adjustment_sets(exposure,
# outcome) (PLURAL, positional) and return a LIST of valid sets.
dag = sp.dag(
    "age -> ldl_baseline; age -> statin_initiation; "
    "ldl_baseline -> statin_initiation; ldl_baseline -> mace; "
    "comorbidity_index -> statin_initiation; comorbidity_index -> mace; "
    "statin_initiation -> mace"
)
adj = dag.adjustment_sets("statin_initiation", "mace")    # list of back-door sets, e.g. [{...}]

# 2.2 Propensity-score overlap (positivity check; epi convention before any IPW)
# Returns a pd.Series of fitted PS — draw mirrored histograms by exposure.
ps = sp.propensity_score(cohort, treatment="statin_initiation",
                          covariates=["age","sex","ldl_baseline","comorbidity_index","smoker"],
                          method="logit")
import matplotlib.pyplot as plt
fig, ax = plt.subplots(figsize=(6,4))
ax.hist(ps[cohort["statin_initiation"]==1], bins=40, alpha=0.5, label="Treated")
ax.hist(ps[cohort["statin_initiation"]==0], bins=40, alpha=0.5, label="Control")
ax.set_xlabel("Estimated propensity score"); ax.legend()
fig.savefig("figures/figA1_ps_overlap.png", dpi=300)

# 2.3 Crude KM curves by exposure (descriptive identification graphic).
# KMResult.plot() returns a bare Axes (NOT a (fig, ax) tuple) — save via ax.figure.
km = sp.kaplan_meier(cohort, duration="followup_days", event="mace", group="statin_initiation")
ax = km.plot()
ax.figure.savefig("figures/figA2_km.png", dpi=300)

A.3 Main estimate — IPTW · g-formula · TMLE triplet (the modern epi standard)

Report all three in one regtable so the reader sees convergent doubly-robust evidence — this is the epi equivalent of the AER design horse race:

# (1) IPTW marginal structural model
iptw = sp.msm(cohort, y="mace", treat="statin_initiation",
              id="patient_id", time="month",
              time_varying=["ldl_current","comorbidity_index"],
              baseline=["age","sex"])

# (2) Parametric g-formula (g-computation). `sp.gformula` is a MODULE, not a function.
# For a point-treatment g-formula use the top-level `sp.g_computation`:
gcomp = sp.g_computation(cohort, y="mace", treat="statin_initiation",
                         covariates=["age","sex","ldl_baseline","comorbidity_index","smoker"])
# For a TIME-VARYING treatment/confounder g-formula (the Robins setting) use the
# Monte-Carlo g-formula in the module:
#   sp.gformula.gformula_mc(cohort, treatment_cols=["statin_t1","statin_t2",...],
#                           confounder_cols=[["ldl_t1"],["ldl_t2"],...],
#                           outcome_col="mace", strategy=(1,1,1), control_strategy=(0,0,0),
#                           id_col="patient_id", time_col="month")

# (3) TMLE -- doubly robust targeted learning estimator.
# Pass an sklearn-style library list for nuisance learners; statspai stacks them
# internally via SuperLearner. Keep `outcome_library` and `propensity_library`
# explicit so the reviewer can see your nuisance choices.
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import GradientBoostingClassifier, RandomForestClassifier
sl_lib = [LogisticRegression(max_iter=1000),
          GradientBoostingClassifier(),
          RandomForestClassifier()]
tmle = sp.tmle(cohort, y="mace", treat="statin_initiation",
               covariates=["age","sex","ldl_baseline","comorbidity_index","smoker"],
               outcome_library=sl_lib, propensity_library=sl_lib)

# (3-bis) HAL-TMLE if you want a fully nonparametric variant.
# `variant=` only accepts "delta" (the default; "projection" is NotImplemented).
hal = sp.hal_tmle(cohort, y="mace", treat="statin_initiation",
                  covariates=["age","sex","ldl_baseline","comorbidity_index","smoker"],
                  variant="delta")

# Convergent-evidence table — risk difference at 5 years
rt = sp.regtable(iptw, gcomp, tmle, hal,
                 model_labels=["(1) IPTW-MSM","(2) g-formula","(3) TMLE","(4) HAL-TMLE"],
                 stats=["N","Effect type","Risk diff. (RD)","Risk ratio (RR)"],
                 title="Table 2. Effect of statin initiation on 5-yr MACE — convergent estimators")
rt.to_word ("tables/table2_epi.docx"); rt.to_excel("tables/table2_epi.xlsx")

A.4 Survival outcomes — KM / AFT / restricted mean

import pandas as pd

# AFT formula LHS is "duration + event" (NOT R-style Surv(time, event)).
aft = sp.aft("followup_days + mace ~ statin_initiation + age + sex + ldl_baseline",
             cohort, family="weibull")
print(aft.summary())                                  # AFTResult exposes .summary() (text)

# AFTResult exposes `.params` (a pd.Series) + `.std_errors`, so it drops STRAIGHT into
# sp.regtable → SEs, stars, and one-line Word/Excel/LaTeX export via the RegtableResult.
# (AFTResult itself still has no `.to_word`/`.to_latex`/`.conf_int` — go through regtable.)
aft_tbl = sp.regtable(aft, model_labels=["Weibull AFT"],
                      title="Table 3. Survival (accelerated failure time)")
aft_tbl.to_word("tables/table3_survival.docx"); aft_tbl.to_excel("tables/table3_survival.xlsx")
# Read N / events / AIC straight off the result for the table footer:
print(f"N={aft.n}, events={aft.n_events}, family={aft.family}, AIC={aft.aic:.1f}")
# Manual fallback only if you need a custom layout (regtable is preferred):
#   pd.DataFrame({"coef": aft.beta, "se": aft.se}, index=aft.var_names)

# For a CAUSAL survival estimand (risk/RMST contrast under unconfoundedness), use the
# doubly-robust longitudinal-TMLE survival estimator instead of a raw AFT:
#   sp.ltmle_survival(cohort, ...)   # returns an LTMLESurvivalResult

A.5 Mendelian randomization (genetic IV — when relevant)

import pandas as pd

# Standard MR triple: IVW → Egger → weighted median, on summary statistics.
# Each mr_* returns a DICT (keys: estimate, se, ci_lower, ci_upper, p_value, ...) —
# NOT a result object, so it does NOT go into sp.regtable. Assemble a DataFrame instead.
ivw    = sp.mr_ivw   (beta_exposure, beta_outcome, se_exposure, se_outcome)
egger  = sp.mr_egger (beta_exposure, beta_outcome, se_exposure, se_outcome)   # 'intercept'(_p) = pleiotropy test
median = sp.mr_median(beta_exposure, beta_outcome, se_exposure, se_outcome, penalized=True)

mr_table = pd.DataFrame(
    {"IVW": ivw, "MR-Egger": egger, "Weighted median": median}
).T[["estimate", "se", "ci_lower", "ci_upper", "p_value"]]
mr_table.to_excel("tables/table4_mr.xlsx")        # or .to_latex() / .to_markdown()
print(mr_table)
# Egger intercept ≠ 0 (egger["intercept_p"] < 0.05) flags directional pleiotropy.

A.6 Robustness — E-value, bounds, principal stratification

# E-value: minimum strength of unmeasured confounding to explain away the result.
# CausalResult exposes `.estimate` and `.ci` (there is NO `.point_estimate`).
ev = sp.evalue(estimate=tmle.estimate, ci=tmle.ci, measure="RR")
# → "E-value 1.84; CI E-value 1.42" (a confounder must be ~2x associated with both
#   exposure and outcome to nullify the effect — interpret in your domain)

# Manski / Lee bounds — `sp.bounds` is a MODULE; call the specific estimator:
bds  = sp.bounds.manski_bounds(cohort, y="mace", treat="statin_initiation")
# selection bias (truncation-by-death / attrition): sp.bounds.lee_bounds(..., selection="observed")

# Principal stratification — BOTH `strata` and `instrument` must be BINARY (0/1) columns
# that already exist in the frame (build them first; they are NOT created for you).
cohort["high_density_zip"] = (cohort["zip_pharmacy_density"] > 0).astype(int)   # binary instrument
cohort["adherent"]         = (cohort["adherence_score"] > 0.8).astype(int)      # 0/1 stratum indicator
ps_strat = sp.principal_strat(cohort, y="mace", treat="statin_initiation",
                              instrument="high_density_zip",
                              strata="adherent")

A.7 Reporting checklist (epi-specific footer for notes=)

When producing the Table-2 footer, include — in addition to the AER stars/SE language:

• Cohort size, person-years of follow-up, event count
• Adjustment set (variables in the back-door set, not just "controls")
• Positivity diagnostic (PS truncation rule, % of cohort with extreme weights)
• E-value for the main effect and its CI bound
• For survival: proportional-hazards check (Schoenfeld residuals p-value) or "PH violated, RMST reported instead"
• STROBE checklist completion (cite as a supplementary file)

Output path stays identical: every estimator above returns a CausalResult and slots straight into sp.regtable(...) / sp.collect(...) / sp.paper_tables(...). Doubly-robust estimators (TMLE, HAL-TMLE, AIPW) are preferred over single-robust IPTW or g-formula alone — report all three for transparency, but treat TMLE as the primary.

§B. ML causal inference pipeline (Mode B)

Convention: estimand-first, doubly-robust, ML-nuisance-learned, with CATE distribution + policy value as first-class outputs (not just a single ATE). The skill mirrors the AER skeleton but the Step-4 estimator stack is DML + meta-learners + causal forest + neural-causal + BCF, and Step-5 always reports a CATE distribution. Uncertainty is quantified by conformal prediction (sp.conformal_causal), not just normal-approximation SE.

Running example: a marketing uplift study — treatment = personalized_offer, outcome = revenue_30d, with 80+ covariates including text features (prior_browsing_text).

B.0 Prep + nuisance super-learner

import statspai as sp

# 0.1 Train/holdout split — DML uses cross-fitting internally, but holdout is for policy eval.
# statspai doesn't expose its own splitter; use sklearn directly.
from sklearn.model_selection import train_test_split
train, holdout = train_test_split(df, test_size=0.2, stratify=df["treatment"], random_state=42)

# 0.2 Nuisance learners. IMPORTANT: `sp.dml` / `sp.metalearner` do NOT accept a
# `sp.super_learner(...)` object — pass a scikit-learn estimator OBJECT, or (for `dml`
# only) a string alias from {'gbm','rf','lasso','ridge','linear','xgb','lgbm'}.
from sklearn.linear_model import LogisticRegression, LassoCV
from sklearn.ensemble import (GradientBoostingRegressor, GradientBoostingClassifier,
                              RandomForestRegressor, RandomForestClassifier)
g_outcome = GradientBoostingRegressor()       # nuisance E[Y|X]  (a sklearn estimator)
g_treat   = GradientBoostingClassifier()      # nuisance E[D|X]  (propensity)

# `sp.super_learner` is a separate STANDALONE stacked predictor (it returns a fitted
# SuperLearner with .predict / .predict_proba) — use it for a reward model in OPE (B.4)
# or for your own predictions, NOT as the nuisance argument to dml/metalearner.
sl_reward = sp.super_learner(X=train[X_cols].values, y=train["revenue_30d"].values,
                             library=[LassoCV(), GradientBoostingRegressor(), RandomForestRegressor()],
                             n_folds=5, task="regression")

B.1 Estimand & DAG learning (Step 2 + 2.5 in ML key)

# estimand is an UPPERCASE enum: 'ATE'|'ATT'|'ATU'|'LATE'|'CATE'|'ITT'. The strategy
# is set via design=/estimand= on causal_question; q.identify() takes NO arguments.
q = sp.causal_question(treatment="treatment", outcome="revenue_30d", data=train,
                       population="marketed users", estimand="ATE",
                       design="selection_on_observables", covariates=X_cols)
plan = q.identify()

# DAG learning (when domain DAG isn't given)
proposed = sp.llm_dag_propose(variables=X_cols + ["treatment","revenue_30d"],
                              domain="e-commerce uplift")
constrained = sp.pc_algorithm(train[X_cols + ["treatment","revenue_30d"]],
                              variables=X_cols + ["treatment","revenue_30d"], alpha=0.05)
validated = sp.llm_dag_validate(proposed, train, alpha=0.05)   # (dag, data) positional
# Alternative learners: sp.notears(...), sp.causal_discovery(..., method="ges")

B.2 Estimator stack — DML / meta-learner / GRF / neural / Bayesian

# (1) DML — Chernozhukov double machine learning.
# Nuisance kwargs are `model_y` (outcome) and `model_d` (treatment) — NOT ml_g/ml_m.
# Each takes a sklearn estimator OR a string alias ('gbm'/'rf'/'lasso'/'xgb'/...).
dml = sp.dml(train, y="revenue_30d", d="treatment", X=X_cols,
             model="plr",                    # plr / irm / iv / pliv
             model_y=g_outcome, model_d=g_treat, n_folds=5)

# (2) Meta-learners — S / T / X / R / DR. outcome_model/propensity_model take sklearn
# estimator OBJECTS (not strings); omit them for sensible defaults.
ml_dr = sp.metalearner(train, y="revenue_30d", treat="treatment", covariates=X_cols,
                       learner="dr",         # 's' / 't' / 'x' / 'r' / 'dr'
                       outcome_model=GradientBoostingRegressor(),
                       propensity_model=GradientBoostingClassifier())

# (3) Causal forest (GRF / honest splits)
cf = sp.causal_forest("revenue_30d ~ treatment | " + " + ".join(X_cols),
                       train, n_estimators=2000, honest=True)

# (4) Neural causal — Dragonnet / TARNet / CEVAE. REQUIRES torch: pip install statspai[neural].
# Omit this block (and the neural columns below) if torch is not installed.
dn   = sp.dragonnet(train, y="revenue_30d", treat="treatment", covariates=X_cols,
                    repr_layers=(200,100), head_layers=(100,))
tar  = sp.tarnet  (train, y="revenue_30d", treat="treatment", covariates=X_cols)

# (5) Bayesian causal forest (full posterior over CATE)
bcf  = sp.bcf(train, y="revenue_30d", treat="treatment", covariates=X_cols,
              n_trees_mu=200, n_trees_tau=50)

# (6) Panel matrix completion (when units × periods)
mc   = sp.matrix_completion(panel_df, y="revenue", d="treatment", unit="user_id", time="week")

# Convergent evidence table — same regtable / collect stack. CausalResult AND CausalForest
# both flow into regtable; drop `dn` if you skipped the neural block.
rt = sp.regtable(dml, ml_dr, cf, dn, bcf,
                 model_labels=["(1) DML-PLR","(2) DR-Learner","(3) Causal forest",
                               "(4) Dragonnet","(5) BCF"],
                 stats=["N","ATE","CATE 5–95% range","Cross-fit folds","Nuisance R²"],
                 title="Table 2. ATE — ML estimator horse race")
rt.to_word ("tables/table2_ml.docx"); rt.to_excel("tables/table2_ml.xlsx")

B.3 CATE distribution & subgroup view (the ML-causal headline)

# 3.1 Per-row CATE. Plotters return (fig, ax). The raw per-row CATE vector lives at
# ml_dr.model_info["cate"] (an ndarray) — there is NO .cate_estimates attribute.
fig, ax = sp.cate_plot(ml_dr, kind="hist",
                       title="Figure B1. CATE distribution — DR-Learner")
fig.savefig("figures/figB1_cate_dist.png", dpi=300)

# 3.2 CATE by group (skill quartiles, gender, channel, …)
g = sp.cate_by_group(ml_dr, train, by="customer_value_quartile", n_groups=4)
fig, ax = sp.cate_group_plot(g, title="Figure B2. CATE by customer-value quartile")
fig.savefig("figures/figB2_cate_group.png", dpi=300)

# 3.3 Causal-forest local effects. CausalForest has no .local_effects(); get the per-row
# CATE vector with cf.effect(X) (ndarray) and plot it yourself.
import numpy as np, matplotlib.pyplot as plt
tau = cf.effect(train[X_cols].values)                # per-row CATE, length n
fig, ax = plt.subplots(figsize=(7, 4))
ax.hist(tau, bins=40); ax.set_xlabel("Causal-forest CATE"); ax.set_title("Figure B3. CF local effects")
fig.savefig("figures/figB3_local.png", dpi=300)

B.4 Policy learning + off-policy evaluation

import numpy as np

# 4.1 Learn an interpretable policy tree from CATE estimates. The result is dict-like
# with .plot_tree() (NOT .plot()), .summary(), .to_latex(), .to_excel(). plot_tree → (fig, ax).
pol_tree = sp.policy_tree(train, y="revenue_30d", d="treatment", X=X_cols, max_depth=3)
fig, ax = pol_tree.plot_tree()
fig.savefig("figures/figB4_policy.png", dpi=300)

# 4.2 Safe policy under a cost constraint. `state` and `action` must each be a SINGLE
# DISCRETE column name (not a list of feature columns). Encode the state into one
# discrete segment column first if you have many features.
train = train.assign(segment=train["customer_value_quartile"])      # one discrete state col
safe = sp.offline_safe_policy(train, state="segment", action="treatment",
                              reward="revenue_30d", cost="offer_cost", cost_threshold=2.50)

# 4.3 Off-policy evaluation on holdout — IPS / DR / SNIPS.
# sp.ope exposes ips / direct_method / doubly_robust / snips / switch_dr. CRITICAL shapes:
#   pi_b, pi_e are (n, K) probability matrices over K actions (one-hot for deterministic);
#   reward_model is a CALLABLE reward_model(X, a) -> length-n predicted reward.
from sklearn.ensemble import GradientBoostingClassifier, GradientBoostingRegressor
g_t = GradientBoostingClassifier().fit(train[X_cols], train["treatment"])
g_r = GradientBoostingRegressor().fit(
    np.column_stack([train[X_cols].values, train["treatment"].values]), train["revenue_30d"])

X_test = holdout[X_cols].values
A_test = holdout["treatment"].to_numpy(int)
R_test = holdout["revenue_30d"].to_numpy(float)
p1     = g_t.predict_proba(X_test)[:, 1]
pi_b   = np.column_stack([1 - p1, p1])                              # (n, 2) behavior policy
a_e    = (g_r.predict(np.column_stack([X_test, np.ones(len(X_test))]))    # treat-if-uplift>0
          > g_r.predict(np.column_stack([X_test, np.zeros(len(X_test))]))).astype(int)
pi_e   = np.column_stack([1 - a_e, a_e]).astype(float)             # (n, 2) one-hot eval policy
reward_model = lambda X, a: g_r.predict(np.column_stack([X, np.full(len(X), a)]))
opv = sp.ope.doubly_robust(X_test, A_test, R_test, pi_b=pi_b, pi_e=pi_e,
                           reward_model=reward_model)
print(f"Policy value (DR): {opv.value:.3f} ± {opv.se:.3f}")
# IPS/SNIPS need no reward model: sp.ope.snips(A_test, R_test, pi_b=pi_b, pi_e=pi_e)

B.5 Uncertainty + fairness + robustness

# 5.1 Conformal prediction intervals on CATE — distribution-free coverage.
# sp.conformal_causal exposes conformal_cate / conformal_ite / conformal_continuous /
# conformal_fair / conformal_interference and more — pick by estimand.
cp = sp.conformal_causal.conformal_cate(train, y="revenue_30d", treat="treatment",
                                         covariates=X_cols, alpha=0.10)   # 90% PI

# 5.2 Subgroup fairness audit — DP / EO gaps across protected attributes.
# fairness_audit audits a BINARY classifier: BOTH `predictions` and `labels` must be 0/1
# columns. (A meta-learner result has no .predict — score with your own classifier.)
holdout = holdout.assign(
    targeted  = (g_t.predict_proba(holdout[X_cols])[:, 1] > 0.5).astype(int),   # binary decision
    responded = (holdout["revenue_30d"] > holdout["revenue_30d"].median()).astype(int),  # binary label
)
fair = sp.fairness.fairness_audit(holdout, predictions="targeted",
                                  protected="gender", labels="responded",
                                  threshold=0.10)

# 5.3 Sensitivity dashboard — a TEXT/numeric dashboard (.summary() + numeric attrs),
# NOT a figure (no .plot()/.savefig()).
print(sp.sensitivity_dashboard(dml, train).summary())

# 5.4 (Reuse AER §7 robustness) Spec curve over nuisance/control choices.
# se_types ∈ {'nonrobust','hc1'/'robust','cluster'}; sc.plot() → (fig, ax).
sc = sp.spec_curve(train, y="revenue_30d", x="treatment",
                   controls=[X_cols[:1], X_cols[:3], X_cols],
                   se_types=["nonrobust", "robust"])
fig, ax = sc.plot()
fig.savefig("figures/figB6_spec_curve.png", dpi=300)

B.6 Reporting checklist (ML-causal-specific footer)

When producing the Table-2 footer, include — in addition to the AER stars/SE language:

• Nuisance learners used (e.g., "outcome: SuperLearner[xgb, rf, lasso, nn]; treatment: same")
• Cross-fitting: number of folds, sample-splitting scheme
• Overlap diagnostic: PS distribution range, % trimmed
• CATE summary: mean / 5–95% range / share with CATE > 0
• Policy value: off-policy DR value vs. random / vs. always-treat baselines
• Conformal coverage: empirical coverage of nominal 1−α PI on holdout
• Fairness audit: subgroup CATE gaps vs. acceptable thresholds

Doubly-robust DML / DR-Learner / TMLE are preferred over single-robust S- or T-learner alone. Report S- or T-learner only as a baseline in the horse race. Always check overlap before reporting any IPW-flavored estimator.

Method Catalog

Classical

Choose by FE structure:

• No FE / single low-cardinality FE → sp.regress (statsmodels OLS wrapper)
• High-dim FE absorption (y ~ x | fe1 + fe2) → sp.feols (pyfixest backend, AER workhorse)
• Two-way panel (entity × time) → sp.panel(...) (linearmodels backend, standard panel diagnostics)

sp.regress("y ~ x1 + x2", df, cluster="firm_id")                       # OLS — `|` is NOT FE here
sp.feols  ("y ~ x1 + x2 | firm_id + year", df, vcov={"CRV1":"firm_id"})# OLS + 2-way FE absorbed
sp.feols  ("y ~ x1 + x2 | firm_id",        df, vcov={"CRV1":"firm_id+year"})  # 2-way cluster
sp.fepois ("count ~ x1 + x2 | firm_id",    df, vcov={"CRV1":"firm_id"})# Poisson + FE (count outcomes)
sp.feglm  ("y ~ x1 + x2 | firm_id", df, family=

> Content truncated for page performance. Open the source repository for the full SKILL.md file.