feat: v0.9.16 — breadth expansion (new-methods pass across 15 sub-packages)

Ships 30+ new modules across 15 sub-packages plus 7 dedicated test
files, bumping StatsPAI's coverage of the applied-causal-inference
landscape closer to v1.0. This is a breadth release in the spirit
of the earlier v1.0 upgrade pass (see memory project_v1_upgrade),
not an architectural change.

Spec: docs/superpowers/specs/2026-04-20-v0916-breadth-expansion.md

New estimators and analyses by area
  Causal forests:      multi_arm_forest, iv_forest,
                       survival/causal_forest (Cui-Kosorok 2023)
  Causal discovery:    FCI with latent confounders (Zhang 2008),
                       ICP / nonlinear_icp (Peters-Bühlmann-Meinshausen)
  DAG tooling:         do_calculus, identification, counterfactual,
                       swig, llm_dag
  DTR:                 q_learning, a_learning, snmm (Robins 2004)
  TMLE:                ltmle (van der Laan-Gruber 2012)
  Proximal:            negative_controls (Lipsitch / Miao-Shi-Tchetgen),
                       pci_regression (Cui-Tchetgen 2020)
  Interference:        network_exposure (Aronow-Samii 2017),
                       peer_effects
  Dose-response:       vcnet + scigan (Nie-Brunskill-Wager 2021)
  Matching:            genmatch (Diamond-Sekhon 2013)
  Sensitivity:         rosenbaum_bounds (Rosenbaum 2002)
  Spatial:             spatial_did, spatial_iv (Kelejian-Prucha 1998)
  Policy learning:     ope direct / IPS / SNIPS / doubly-robust
  Time series:         its (interrupted time series)
  Censoring / survival IPCW:  ipcw (Robins-Finkelstein 2000)
  Target-trial emulation: protocol, emulate, ccw, diagnostics
  g-formula:           ice (iterated conditional expectation)
  Transport:           identify, generalize, weighting
  Bounds:              balke_pearl (Balke-Pearl 1997)
  Mediation:           four_way_decomposition (VanderWeele 2014)
  Neural causal:       cevae (Louizos et al. 2017),
                       gnn_causal
  Many-weak-IV:        iv/many_weak
  OPE package:         new top-level ope/
  Bayes DID:           bayes_did gains cohort= parameter and
                       dedicated BayesianDIDResult type

New tests (7 files)
  test_target_trial.py, test_dag_scm.py, test_icp.py,
  test_ope_cevae.py, test_transport.py, test_gformula_ice.py,
  test_bayes_mte_bivariate_normal.py

Test hardening gap (declared up front): 20+ of the 30+ modules
ship without a dedicated test file. The import chain is healthy
(collection: 2373 / 2381 pass; 559 passed in the focused regression
over affected sub-packages), but per-module correctness tests are
deferred to follow-up commits per spec §4.4 (P0: ltmle, fci;
P1: dtr trio, negative_controls, network_exposure).

Version
  0.9.15 → 0.9.16 in pyproject.toml and src/statspai/__init__.py.

Regression evidence
  pytest --collect-only -q:  2373 collected, 8 deselected in 11s
  pytest on affected packages: 559 passed, 3 skipped, 0 failed in 339s

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brycewang-stanford

docs/superpowers/specs/2026-04-20-v0916-breadth-expansion.md

@@ -0,0 +1,190 @@
+# StatsPAI v0.9.16 — v1.0 breadth expansion (20 new modules)
+
+**Author:** Bryce Wang · **Date:** 2026-04-20 · **Status:** design → implementation
+
+## 1. Motivation
+
+Between the v0.9.15 tidy-multiterm spec (landed as db050de) and this
+release, 20 new modules were authored to cover previously-unsupported
+methods — most of them requested by users working on applied causal
+projects where StatsPAI was already close but missing a specific
+estimator. This spec catalogs the batch, flags gaps, and defines the
+ship gate.
+
+This is a **breadth pass**, analogous to the v1.0 upgrade recorded in
+memory `project_v1_upgrade.md`, but scoped tighter: each module is a
+standalone addition with no cross-cutting architectural change.
+
+## 2. Inventory
+
+20 new modules, ~4,800 LOC, organised by sub-package.
+
+> **Note:** this inventory was captured at the start of review. The
+> actual release grew to **27+ modules** and **3 dedicated test files**
+> as the author added more coverage during the sign-off window (DAG
+> do-calculus / SWIG / counterfactual, Balke-Pearl bounds, ICP causal
+> discovery, four-way decomposition, peer effects, transport package,
+> plus `tests/test_dag_scm.py` and `tests/test_icp.py`). The authoritative
+> file list is the release commit diff.
+
+### 2.1 Causal forests family (648 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `causal/iv_forest.py` | 200 | Forest-based LATE | Athey-Tibshirani-Wager 2019 |
+| `causal/multi_arm_forest.py` | 176 | Multi-arm CATE | Tibshirani-Wager 2020 |
+| `survival/causal_forest.py` | 272 | Survival CATE | Cui-Kosorok-Athey-Wager 2023 |
+
+### 2.2 Causal discovery (726 LOC)
+| Module | LOC | Output | Reference |
+|---|---|---|---|
+| `causal_discovery/fci.py` | 430 | PAG with latent-confounder edges | Spirtes-Meek-Richardson 1995, Zhang 2008 |
+| `dag/identification.py` | 296 | Identifiability certificate on a user DAG | Pearl back-door/front-door + ID algorithm |
+
+### 2.3 Dynamic treatment regimes (522 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `dtr/q_learning.py` | 185 | Q-function backward induction | Murphy 2003 |
+| `dtr/a_learning.py` | 173 | Advantage / blip functions | Robins 2004, Moodie et al. 2007 |
+| `dtr/snmm.py` | 164 | Structural nested mean model | Robins 1994 |
+
+### 2.4 TMLE extensions (381 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `tmle/ltmle.py` | 381 | Longitudinal ATE under static regime | van der Laan-Gruber 2012 |
+
+### 2.5 Proximal causal inference (505 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `proximal/negative_controls.py` | 312 | NCO/NCE calibration + double-NC ATE | Lipsitch 2010, Miao-Shi-Tchetgen 2018/2020 |
+| `proximal/pci_regression.py` | 193 | Proximal 2-stage regression | Cui-Tchetgen 2020 |
+
+### 2.6 Interference / spillover (323 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `interference/network_exposure.py` | 323 | HT estimator for exposure mappings | Aronow-Samii 2017 |
+
+### 2.7 Dose-response / continuous treatment (253 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `dose_response/vcnet.py` | 253 | Varying-coefficient dose-response (+ SCIGAN) | Nie-Brunskill-Wager 2021 |
+
+### 2.8 Matching (231 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `matching/genmatch.py` | 231 | Genetic-algorithm matching | Diamond-Sekhon 2013 |
+
+### 2.9 Diagnostics / sensitivity (295 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `diagnostics/rosenbaum.py` | 295 | Rosenbaum Γ sensitivity bounds | Rosenbaum 2002 |
+
+### 2.10 Spatial econometrics (376 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `spatial/did.py` | 208 | DID with spatial spillovers | Delgado-Florax 2015, Butts 2021 |
+| `spatial/iv.py` | 168 | Spatial IV | Kelejian-Prucha 1998 |
+
+### 2.11 Policy learning (222 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `policy_learning/ope.py` | 222 | Off-policy evaluation (IPS, DR, switch) | Dudík-Erhan-Langford-Li 2014 |
+
+### 2.12 Time series causal (203 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `timeseries/its.py` | 203 | Interrupted time series | Wagner et al. 2002 |
+
+### 2.13 Censoring / survival weights (266 LOC)
+| Module | LOC | Estimand | Reference |
+|---|---|---|---|
+| `censoring/ipcw.py` | 260 | Inverse-probability-of-censoring weights | Robins-Finkelstein 2000 |
+
+### 2.14 Target trial emulation (503 LOC, **only module WITH tests**)
+| Module | LOC | Purpose |
+|---|---|---|
+| `target_trial/protocol.py` | 151 | Protocol specification primitives |
+| `target_trial/emulate.py` | 125 | Emulation pipeline |
+| `target_trial/ccw.py` + `ccw_internal.py` | 188 | Clone-censor-weight backbone |
+| `target_trial/diagnostics.py` | 61 | Balance + positivity diagnostics |
+| **`tests/test_target_trial.py`** | **188 LOC** | **sole dedicated test** |
+
+## 3. Re-export surface
+
+Main `src/statspai/__init__.py`: +35 / −8 lines, registers 13 new
+top-level symbols:
+`multi_arm_forest`, `iv_forest`, `genmatch`, `proximal_regression`,
+`ltmle`, `vcnet`, `scigan`, `fci`, `network_exposure`, `q_learning`,
+`a_learning`, `snmm`, `causal_survival_forest`.
+
+Sub-package `__init__.py` diffs: 12 files, 1–4 lines each, all pure
+re-exports.
+
+`sp.*` smoke check: all 13 new symbols resolve on import.
+
+## 4. Ship gate
+
+### 4.1 Blockers identified
+
+| # | Blocker | Severity |
+|---|---|---|
+| B1 | 22 of 23 new modules have **no dedicated tests** — only `target_trial` does | HIGH |
+| B2 | Version string stays at `0.9.14` — not bumped to `0.9.16` | MED |
+| B3 | `CHANGELOG.md` not updated | LOW |
+
+### 4.2 Mitigation strategy — **WIP release**
+
+This release ships as **`v0.9.16.dev` / "breadth expansion WIP"** with
+a clear documentation disclaimer. Reasoning:
+
+- Writing production tests for 22 modules upfront would gate the
+  release by 2–4 weeks and leave finished code rotting in working
+  tree.
+- The import chain is healthy (`sp.*` all resolve), so users who
+  discover a module and read its docstring + paper refs get a
+  working-in-principle implementation.
+- The known-gap signalling pattern matches `memory/project_v1_upgrade`
+  and `project_synth_v090_review` precedent: ship, then harden with
+  follow-up test commits per module.
+
+### 4.3 Required pre-commit actions
+
+Only what is strictly needed to avoid shipping a broken release:
+
+1. Bump version `0.9.14 → 0.9.16` in `pyproject.toml` + `src/statspai/__init__.py`.
+2. Run full test suite (no new tests added, but must confirm no
+   regressions against the 2,319 existing tests).
+3. Commit as **one release commit**: `feat: v0.9.16 — breadth expansion (20 new modules, test hardening follow-up)`.
+
+### 4.4 Deferred (follow-up commits)
+
+Per-module test files, one PR/commit per module, roughly prioritised by user-visibility:
+
+| Priority | Module | Reason |
+|---|---|---|
+| P0 | `tmle/ltmle.py` | Most complex algorithm (recursive backward induction); highest failure surface |
+| P0 | `causal_discovery/fci.py` | PAG output format is sticky to API changes |
+| P1 | `dtr/{q,a,snmm}_learning.py` | 3-way API consistency |
+| P1 | `proximal/negative_controls.py` | Paper-sensitive math |
+| P1 | `interference/network_exposure.py` | MC exposure probs are numerically delicate |
+| P2 | All others | Smaller and / or simpler |
+
+## 5. Commit strategy
+
+**Single commit**, following the project's historical release pattern
+(`a546ead feat: v0.9.12 —`, `2fd88de feat: v0.9.13 —`, etc.):
+
+```
+feat: v0.9.16 — breadth expansion (20 new modules, test hardening follow-up)
+```
+
+The commit body explicitly declares the WIP test status so future
+`git blame` readers see the gap.
+
+## 6. Success criteria
+
+1. `git push origin main` succeeds.
+2. `pip install .` produces a 0.9.16 wheel.
+3. `import statspai as sp` + 13 new symbol lookups still resolve after
+   checkout.
+4. Full existing test suite stays green (no regressions).
+5. At least one follow-up commit per P0 module within 2 weeks.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "StatsPAI"
-version = "0.9.15"
+version = "0.9.16"
 description = "The Agent-Native Causal Inference & Econometrics Toolkit for Python"
 readme = "README.md"
 license = {text = "MIT"}

src/statspai/__init__.py

@@ -22,7 +22,7 @@
 >>> sp.outreg2(result, filename="results.xlsx")
 """
 
-__version__ = "0.9.15"
+__version__ = "0.9.16"
 __author__ = "Biaoyue Wang"
 __email__ = "brycew6m@stanford.edu"
 
@@ -89,7 +89,7 @@
     Absorber, demean, absorb_ols, hdfe_ols, FEOLSResult,
 )
 from .causal_impact import causal_impact, CausalImpactEstimator, impactplot
-from .mediation import mediate, MediationAnalysis, mediate_sensitivity, mediate_interventional
+from .mediation import mediate, MediationAnalysis, mediate_sensitivity, mediate_interventional, four_way_decomposition, FourWayResult
 from .bartik import bartik, BartikIV, ssaggregate, shift_share_se
 from .output.outreg2 import OutReg2, outreg2
 from .output.modelsummary import modelsummary, coefplot
@@ -164,8 +164,9 @@
 from .regression.logit_probit import logit, probit, cloglog
 from .regression.glm import glm, GLMRegression, GLMEstimator
 from .regression.count import poisson, nbreg, ppmlhdfe
-from .neural_causal import tarnet, cfrnet, dragonnet, TARNet, CFRNet, DragonNet
-from .causal_discovery import notears, NOTEARS, pc_algorithm, PCAlgorithm, lingam, LiNGAMResult, ges, GESResult, fci, FCIResult
+from .neural_causal import tarnet, cfrnet, dragonnet, TARNet, CFRNet, DragonNet, gnn_causal, GNNCausalResult
+from .neural_causal.cevae import cevae, CEVAE, CEVAEResult
+from .causal_discovery import notears, NOTEARS, pc_algorithm, PCAlgorithm, lingam, LiNGAMResult, ges, GESResult, fci, FCIResult, icp, nonlinear_icp, ICPResult
 from .tmle import tmle, TMLE, super_learner, SuperLearner, ltmle, LTMLEResult
 from .policy_learning import policy_tree, PolicyTree, policy_value, direct_method, ips, snips, doubly_robust, OPEResult
 from .conformal_causal import conformal_cate, ConformalCATE
@@ -174,7 +175,7 @@
 from .matrix_completion import mc_panel, MCPanel
 from .dose_response import dose_response, DoseResponse, vcnet, scigan, VCNetResult
 from .bounds import lee_bounds, manski_bounds, BoundsResult, horowitz_manski, iv_bounds, oster_delta, selection_bounds, breakdown_frontier, balke_pearl, BalkePearlResult
-from .interference import spillover, SpilloverEstimator, network_exposure, NetworkExposureResult
+from .interference import spillover, SpilloverEstimator, network_exposure, NetworkExposureResult, peer_effects, PeerEffectsResult
 from .dtr import g_estimation, GEstimation, q_learning, QLearningResult, a_learning, ALearningResult, snmm, SNMMResult
 from .multi_treatment import multi_treatment, MultiTreatment
 from .robustness import spec_curve, SpecCurveResult, robustness_report, RobustnessResult, subgroup_analysis, SubgroupResult
@@ -185,8 +186,24 @@
     rule1 as do_rule1, rule2 as do_rule2, rule3 as do_rule3,
     apply_rules as do_calculus_apply, RuleCheck,
     swig, SWIGGraph, SCM,
+    llm_dag, LLMDAGResult,
 )
 
+# === Transportability (Pearl-Bareinboim + Dahabreh-Stuart) ===
+from . import transport
+from .transport import (
+    TransportWeightResult,
+    identify_transport, TransportIdentificationResult,
+)
+
+# === Off-Policy Evaluation (contextual bandits) ===
+from . import ope
+from .ope import OPEResult
+
+# === Parametric g-formula (iterative conditional expectation) ===
+from . import gformula
+from .gformula import ice as gformula_ice_fn, ICEResult
+
 # === Target Trial Emulation (JAMA 2022 framework) ===
 from . import target_trial
 from .target_trial import (