igerber · wenddymacro · Aug 31, 2026 · Aug 31, 2026 · Aug 31, 2026 · Aug 31, 2026
diff --git a/README.md b/README.md
@@ -132,6 +132,7 @@ Full guide: `diff_diff.get_llm_guide("practitioner")`.
 - [Manipulation Testing](https://diff-diff.readthedocs.io/en/stable/api/regression_discontinuity.html) - Cattaneo, Jansson & Ma (2020) density-discontinuity test (`RDDensityTest`): rddensity 3.0 parity, robust bias-corrected inference, unrestricted/restricted models, mass-point adjustment
 - [Parallel Trends Testing](https://diff-diff.readthedocs.io/en/stable/api/diagnostics.html) - simple and Wasserstein-robust parallel trends tests, equivalence testing (TOST)
 - [Placebo Tests](https://diff-diff.readthedocs.io/en/stable/api/diagnostics.html) - placebo timing, group, permutation, leave-one-out
+- [TWFE Weight Diagnostics](https://diff-diff.readthedocs.io/en/stable/api/twfe_weights.html) - Baker et al. (2025) implicit weights a TWFE regression places on each ATT(g,t), against the ATT^O / ATT^simple targets, with the pre-trend contribution. Ports Callaway's `twfeweights` (MIT)
 - [Honest DiD](https://diff-diff.readthedocs.io/en/stable/api/honest_did.html) - Rambachan & Roth (2023) sensitivity analysis: robust CI under PT violations, breakdown values
 - [Pre-Trends Power Analysis](https://diff-diff.readthedocs.io/en/stable/api/pretrends.html) - Roth (2022) minimum detectable violation and power curves
 - [Power Analysis](https://diff-diff.readthedocs.io/en/stable/api/power.html) - analytical and simulation-based MDE, sample size, power curves for study design

diff --git a/benchmarks/R/generate_twfeweights_golden.R b/benchmarks/R/generate_twfeweights_golden.R
diff --git a/benchmarks/R/requirements.R b/benchmarks/R/requirements.R
@@ -20,6 +20,8 @@ required_packages <- c(
   "Synth",         # Abadie-Diamond-Hainmueller (2010) synthetic control (SyntheticControl R-parity; ships data(basque))
   "qte",           # Callaway qte package (Athey-Imbens CiC + QDiD R-parity; ships data(lalonde))
   "ptetools",      # Callaway ptetools (badcontrols dependency; bad-control DMLDiD lane black-box parity)
+  "BMisc",         # Callaway utility package (twfeweights dependency: weighted_ecdf, orig2t)
+  "DRDID",         # Sant'Anna & Zhao (2020) doubly-robust DiD (twfeweights AIPW dependency)
 
   # Utilities
   "jsonlite",      # JSON output for Python interop
@@ -28,7 +30,9 @@ required_packages <- c(
 
 # synthdid must be installed from GitHub
 github_packages <- list(
-  synthdid = "synth-inference/synthdid"
+  synthdid = "synth-inference/synthdid",
+  # TWFE weight diagnostics parity goldens (not on CRAN)
+  twfeweights = "bcallaway11/twfeweights"
 )
 
 install_if_missing <- function(pkg) {

diff --git a/benchmarks/data/twfeweights_golden.json b/benchmarks/data/twfeweights_golden.json
diff --git a/benchmarks/data/twfeweights_sim_panel.csv b/benchmarks/data/twfeweights_sim_panel.csv
diff --git a/benchmarks/data/twfeweights_unbalanced_panel.csv b/benchmarks/data/twfeweights_unbalanced_panel.csv
diff --git a/changelog.d/20260831-twfe-weight-diagnostics.md b/changelog.d/20260831-twfe-weight-diagnostics.md
@@ -0,0 +1,24 @@
+### Added
+- **TWFE weight diagnostics** (port of Brantly Callaway's `twfeweights` R
+  package, MIT): what a two-way fixed effects regression *implicitly* weights
+  on staggered-adoption data.
+  - `attgt_weights(results, aggregation="twfe"|"overall"|"simple")` reports the
+    weight a TWFE regression, ATT^O, or ATT^simple places on each ATT(g,t),
+    plus post-period negative-weight counts. Returns `ATTGTWeightsResult`.
+    The CS estimands honour the fit's `anticipation` window (an explicit
+    `anticipation=` on the frame path); `"twfe"` keeps `t >= g`.
+  - `decompose_twfe_weights(data, ..., method="fwl")` re-derives the estimate
+    from its ATT(g,t) building blocks and returns `TWFEDecompositionResult`
+    with `pre_period_contribution` - the sample contribution of the
+    pre-treatment cells, which can reflect parallel-trends violations or
+    sampling variation - and, with `balance_covariates=`, implicit-weight
+    covariate balance. `plot_twfe_weights()` renders either view (matplotlib
+    or plotly).
+  - Validation: rejects NaN / `-inf` cohort labels, covariate-adjusted fits
+    under `aggregation="twfe"`, unbalanced panels, non-finite outcomes /
+    covariates, duplicated or non-finite ATT(g,t) cells, an incomplete
+    group-time grid, and invalid sampling weights. Two structural gaps are
+    handled as R does instead of raising: a cohort with no estimable post cell
+    is dropped (`did`'s first-period drop, or running out of comparison
+    units), and under `control_group="not_yet_treated"` the CS estimands
+    average over each cohort's available post periods (`aggte`).
diff --git a/diff_diff/__init__.py b/diff_diff/__init__.py
@@ -298,6 +298,14 @@
     TROPResults,
     trop,
 )
+from diff_diff.twfe_weights import (
+    attgt_weights,
+    decompose_twfe_weights,
+)
+from diff_diff.twfe_weights_results import (
+    ATTGTWeightsResult,
+    TWFEDecompositionResult,
+)
 from diff_diff.two_stage import (
     TwoStageBootstrapResults,
     TwoStageDiD,
@@ -323,6 +331,7 @@
     plot_sensitivity,
     plot_staircase,
     plot_synth_weights,
+    plot_twfe_weights,
 )
 from diff_diff.wooldridge import WooldridgeDiD
 from diff_diff.wooldridge_results import WooldridgeDiDResults
@@ -459,6 +468,13 @@ def __getattr__(name: str) -> _Any:
     "TWFEWeightsResult",
     "chaisemartin_dhaultfoeuille",
     "twowayfeweights",
+    # TWFE weight diagnostics (Callaway `twfeweights` port) - distinct from
+    # the dCDH `twowayfeweights` surface above: these weight ATT(g,t)
+    # parameters, not (unit, time) cells.
+    "ATTGTWeightsResult",
+    "TWFEDecompositionResult",
+    "attgt_weights",
+    "decompose_twfe_weights",
     # WooldridgeDiD (ETWFE)
     "WooldridgeDiD",
     "WooldridgeDiDResults",
@@ -479,6 +495,7 @@ def __getattr__(name: str) -> _Any:
     "DurationDiDPretestResults",
     # Visualization
     "plot_bacon",
+    "plot_twfe_weights",
     "plot_event_study",
     "plot_group_effects",
     "plot_sensitivity",

diff --git a/diff_diff/dml_did.py b/diff_diff/dml_did.py
@@ -3362,6 +3362,7 @@ def fit(
             # stays admitted).
             is_survey_fit=survey_metadata is not None,
             bootstrap_results=bootstrap_results,
+            covariates=covariates,
         )
         self.results_ = results
         self.is_fitted_ = True

diff --git a/diff_diff/guides/llms-full.txt b/diff_diff/guides/llms-full.txt
@@ -1534,6 +1534,60 @@ results.print_summary()
 plot_bacon(results)
 ```
 
+### TWFE Weight Diagnostics
+
+What a TWFE regression implicitly weights on staggered data. Distinct from
+`twowayfeweights` (dCDH), which weights (unit, time) cells: these weight
+ATT(g,t) parameters. Ported from Brantly Callaway's `twfeweights` R package
+(MIT); methodology Baker, Callaway, Cunningham, Goodman-Bacon & Sant'Anna
+(2025).
+
+```python
+attgt_weights(
+    results,                        # CallawaySantAnnaResults, or a (g,t) frame
+    aggregation="twfe",             # "twfe" | "overall" (ATT^O) | "simple"
+    data=None, unit=None, time=None, first_treat=None,   # frame path only
+    weights=None,                   # unit-level sampling weights
+) -> ATTGTWeightsResult
+
+decompose_twfe_weights(
+    data,                           # balanced long panel (it re-estimates)
+    outcome=, unit=, time=, first_treat=,
+    method="fwl",
+    covariates=None,
+    base_period="first_period",     # or "gmin1"
+    balance_covariates=None,        # enables result.covariate_balance()
+    weights=None,
+) -> TWFEDecompositionResult
+
+plot_twfe_weights(result, kind="auto")   # "weights" | "balance"
+```
+
+`aggregation="twfe"` requires a fit with `base_period="universal"`,
+`control_group="never_treated"` AND no covariates (R twfe_weights' three
+restrictions); it raises otherwise. ATT^O and ATT^simple weights are
+non-negative and sum to one, so comparing `implied_att` across the three
+aggregations shows what the TWFE specification costs.
+
+Both entry points fail closed on input R never faced: NaN / `-inf` cohort
+labels (never-treated is exactly `0` or `+inf`), duplicated or non-finite
+ATT(g,t) cells, an incomplete group-time grid (`"twfe"` needs every cohort x
+period cell, the CS estimands every post cell), and sampling weights that are
+not finite, non-negative and positive-mass. Two structural gaps mirror R
+rather than raising, each with a `UserWarning`: a cohort with no estimable
+post cell is dropped from the table and the cohort shares (`did`'s
+first-period drop), and under `control_group="not_yet_treated"` the cells CS
+marks `zero_treated_control` are treated as structurally absent, so
+`"overall"`/`"simple"` average over each cohort's AVAILABLE post periods
+(`aggte`). `n_negative_post` / `negative_post_weight_share` report the
+pathology (negative weight on POST cells); `n_negative` counts pre cells too,
+and is near-half in every staggered design because the TWFE weights sum to
+zero over the full grid.
+
+`decompose_twfe_weights` takes the raw panel rather than a fitted result
+because it re-estimates. It is tied to `attgt_weights` by an identity:
+`attgt_weights(cs, aggregation="twfe").implied_att == decompose_twfe_weights(panel, ...).estimate`.
+
 ### StaggeredTripleDifference
 
 DEPRECATED in 3.9, removed in 4.0 (ledger row M-013). Use
@@ -2017,6 +2071,34 @@ Returned by `BaconDecomposition.fit()` (and the deprecated `bacon_decompose()` w
 
 **Methods:** `summary()`, `print_summary()`, `to_dataframe()`
 
+### ATTGTWeightsResult
+
+Diagnostic result from `attgt_weights`. No inference quintet - the
+decomposition is an algebraic identity.
+
+- `weights`: DataFrame with `group`, `time`, `post`, `weight`, `att`
+- `implied_att`: `sum(weight * att)` - the TWFE coefficient when
+  `aggregation="twfe"`
+- `n_negative`, `negative_weight_share`: the staggered-TWFE pathology
+- `aggregation`, `source`, `control_group`, `base_period`, `n_cells`
+- `summary()`, `to_dataframe()`, `to_dict()`
+
+### TWFEDecompositionResult
+
+Diagnostic result from `decompose_twfe_weights`.
+
+- `cells`: DataFrame with `group`, `time`, `post`, `att`, `weight`, `ess`,
+  `remainder`
+- `estimate` == `decomposition` + `remainder`
+- `pre_period_contribution`: sample contribution of the PRE-treatment cells
+  (can reflect differential pre-trends OR sampling variation; diagnostic
+  evidence, not proof the post-treatment assumption fails)
+- `post_only`, `effective_sample_size`, `covariates`, `base_period`
+- `covariate_balance(level="summary"|"cell", standardize=True,
+  post_only=True)`: implicit-weight covariate balance; raises when
+  `balance_covariates=` was not requested
+- `summary()`, `to_dataframe()`, `to_dict()`
+
 ### Comparison2x2
 
 Individual 2x2 DiD comparison (used in BaconDecompositionResults).
@@ -2532,6 +2614,28 @@ plot_bacon(
 )
 ```
 
+### plot_twfe_weights
+
+```python
+from diff_diff import plot_twfe_weights
+
+plot_twfe_weights(
+    results,                        # ATTGTWeightsResult | TWFEDecompositionResult
+    kind="auto",                    # "weights" | "balance" ("auto" picks balance
+                                    #   when the result carries a balance table)
+    standardize=True, absolute_value=True,   # balance view
+    annotate=False, ax=None, show=True,
+    backend="matplotlib",           # or "plotly"
+)
+```
+
+`kind="weights"` scatters weight against ATT(g,t), coloured by pre/post - points
+left of the vertical zero line carry negative weight; the title counts POST-only
+negatives. `kind="balance"` scatters unweighted against implicitly-weighted
+covariate differences; points near the horizontal axis are covariates the
+implicit weights balance. With `absolute_value=False` the reference diagonal
+spans `[-limit, limit]` so signed differences read correctly.
+
 ### plot_power_curve
 
 ```python

diff --git a/diff_diff/guides/llms.txt b/diff_diff/guides/llms.txt
@@ -92,6 +92,7 @@ The site is organized into 5 sections, each with a landing page:
 - [Manipulation Testing](https://diff-diff.readthedocs.io/en/stable/api/regression_discontinuity.html): Cattaneo, Jansson & Ma (2020) density-discontinuity manipulation test (`RDDensityTest`), parity with R rddensity 3.0 - boundary-adaptive local polynomial density estimation at the cutoff, robust bias-corrected inference, unrestricted/restricted models, jackknife/plugin variances, data-driven bandwidths, mass-point adjustment
 - [Parallel Trends Testing](https://diff-diff.readthedocs.io/en/stable/api/diagnostics.html): Simple and Wasserstein-robust parallel trends tests, equivalence testing (TOST)
 - [Placebo Tests](https://diff-diff.readthedocs.io/en/stable/api/diagnostics.html): Placebo timing, group, permutation, and leave-one-out diagnostics
+- [TWFE Weight Diagnostics](https://diff-diff.readthedocs.io/en/stable/api/twfe_weights.html): Baker et al. (2025) implicit weights on ATT(g,t) - `attgt_weights(results, aggregation='twfe'|'overall'|'simple')` takes a fitted `CallawaySantAnnaResults` (raw ATT(g,t) frame + panel as fallback) and returns the weight each estimand places on each group-time effect, with the negative-weight share; `decompose_twfe_weights(data, outcome=, unit=, time=, first_treat=, method='fwl', covariates=)` re-derives the TWFE estimate from its ATT(g,t) building blocks with `pre_period_contribution` (the pre-treatment cells' sample contribution, which can be pre-trends or sampling noise), and `result.covariate_balance()` reports implicit-weight covariate balance. Plot with `plot_twfe_weights`. R `twfeweights` 0.9.0 output parity
 - [Honest DiD](https://diff-diff.readthedocs.io/en/stable/api/honest_did.html): Rambachan & Roth (2023) sensitivity analysis — robust CI under parallel trends violations, breakdown values
 - [Pre-Trends Power Analysis](https://diff-diff.readthedocs.io/en/stable/api/pretrends.html): Roth (2022) Section II.A-B no-individually-significant (NIS) box-probability pretest power + minimum detectable violation; `pretest_form='nis'` (default) implements the paper's primary form, `pretest_form='wald'` retained as paper-supported alternative (Propositions 1+3+4 all apply); linear-violation MDV in Roth's γ units when relative-time labels are threaded through `fit()`; full Σ_22 routing on non-bootstrap CallawaySantAnna and SunAbraham adapters and on admitted CS-/StackedDiD-sourced `aggregate('event_study')` containers (StackedDiD persists its ES VCV in every inference mode)
 - [Power Analysis](https://diff-diff.readthedocs.io/en/stable/api/power.html): Analytical and simulation-based power analysis — MDE, sample size, power curves for study design

diff --git a/diff_diff/staggered.py b/diff_diff/staggered.py
@@ -7,7 +7,7 @@
 
 import bisect
 import warnings
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
 import numpy as np
 import pandas as pd
@@ -3056,6 +3056,7 @@ def fit(
             group_time_effects,
             is_survey_fit=survey_metadata is not None,
             bootstrap_results=bootstrap_results,
+            covariates=covariates,
         )
 
         self.is_fitted_ = True
@@ -5151,6 +5152,7 @@ def _build_aggregation_kit(
     *,
     is_survey_fit: bool = False,
     bootstrap_results: Optional["CSBootstrapResults"] = None,
+    covariates: Optional[Sequence[str]] = None,
 ) -> Optional["AggregationKit"]:
     """Distil the fit-time state post-fit re-aggregation needs.
 
@@ -5185,6 +5187,16 @@ def _build_aggregation_kit(
     # (or DDD) survey fit does not warn as "CallawaySantAnna" on post-fit
     # aggregate(). Legacy kits without the key default at the read site.
     bookkeeping["bootstrap_label"] = getattr(estimator, "_BOOTSTRAP_LABEL", "CallawaySantAnna")
+    # Covariate usage, recorded so downstream diagnostics can refuse designs
+    # their formulas do not cover (``attgt_weights(aggregation="twfe")``
+    # mirrors R twfe_weights' ``xformla == ~1`` restriction). Column NAMES
+    # only - never values - so the data-minimization contract holds.
+    bookkeeping["covariates"] = tuple(covariates or ())
+    # Panel balance, recorded so ``attgt_weights`` can reject an unbalanced
+    # fitted result: its cohort shares and E_t[D] assume the same units in
+    # every period. Defaults True when the key is absent (a legacy kit, or a
+    # producer that never computed it).
+    bookkeeping["is_balanced"] = bool((precomputed or {}).get("is_balanced", True))
 
     # Data minimization: the results object is picklable and users share
     # result artifacts, so the kit must not turn it into a carrier for raw