DiD module — API reference

Core estimators

sp.callaway_santanna(data, y, g, t, i, ...)

Callaway & Sant'Anna (2021) staggered DID. See the guide for an end-to-end walkthrough.

Key arguments:

Name	Default	Meaning
estimator	'dr'	'dr' (Sant'Anna–Zhao doubly robust), 'ipw', or 'reg'
control_group	'nevertreated'	Or 'notyettreated'
base_period	'universal'	Or 'varying'
anticipation	0	Periods of anticipation (CS2021 §3.2)
panel	True	False for repeated cross-sections
x	None	Covariate list

sp.aggte(result, type='simple', ...)

Unified aggregation with Mammen (1993) multiplier bootstrap.

type ∈ {'simple', 'dynamic', 'group', 'calendar'}.

Relevant arguments:

Name	Default	Meaning
n_boot	1000	Multiplier-bootstrap replications
random_state	None	RNG seed
balance_e	None	Balance cohorts across e ∈ [0, balance_e]
min_e / max_e	-inf / inf	Event-time window
cband	True	Attach uniform confidence band columns

Result's detail frame carries cband_lower / cband_upper for all aggregations except simple.

sp.cs_report(data_or_result, ..., save_to=None)

One-call report card composing the full pipeline. See the guide. Returns a CSReport dataclass with .plot(), .to_markdown(), .to_latex(), .to_excel(), and .to_text() methods.

sp.sun_abraham(data, y, g, t, i, ...)

Sun & Abraham (2021) interaction-weighted event study with Liang–Zeger cluster-robust sandwich SEs.

sp.did_imputation(data, ..., horizon=None)

Borusyak, Jaravel & Spiess (2024) imputation estimator. Also available through the unified entry point:

sp.did(df, y="y", treat="first_treat", time="year", id="unit",
       method="bjs", horizon=list(range(-4, 5)))

horizon attaches the BJS event-study table to result.model_info["event_study"], so result.plot() and sp.honest_did(result, ...) work the same way as CS/SA event studies.

sp.did_multiplegt(data, ..., placebo=0, dynamic=0)

de Chaisemartin–D'Haultfoeuille DID for switch-on-off treatments, with dCDH 2024 joint placebo Wald and average cumulative effect.

Inference

sp.event_study_vcov(result, allow_diagonal=True)

The joint covariance of an event study, read off whichever estimator produced it (CS / aggte, event_study, sun_abraham, gardner_did, did_imputation, stacked_did, lp_did, did_multiplegt_dyn, etwfe). Returns times, beta, vcov, joint and source; joint=False flags the cases where only a diagonal or block-diagonal matrix exists.

sp.uniform_bands(result, alpha=0.05, which='all', window=None)

Sup-t simultaneous band over the covered event times — the object an event-study plot needs, since a pointwise interval covers one horizon at a time. Falls back to the conservative Sidak critical value when the covariance is not joint.

sp.did_few_treated(data, y, id, time, treat, method='conley_taber')

Conley–Taber (2011) and Ferman–Pinto (2019) inference for designs with one or a handful of treated clusters, where the cluster-robust variance over-rejects whatever the total cluster count. The placebo distribution is built from the control groups and inverted; method='ferman_pinto' additionally rescales it for the heteroskedasticity unequal group sizes generate (needs group_size=).

sp.did_calibrated_simulation(data, y, id, time, cohort, estimators=..., effect=0.0)

Scores candidate estimators on a placebo simulation calibrated to your own panel: the estimated dynamic effect is removed with the imputation fit, the adoption pattern is redrawn, a known effect is injected, and every candidate is refit n_sims times. Returns bias, RMSE, coverage, rejection rate and the ratio of mean reported SE to realised dispersion, each with a Monte Carlo standard error, plus .best('rmse'). The redrawn assignment is random, so parallel trends holds by construction and the table is about the estimators, not the design. stacked_did and lp_did estimate over an event window, which need not cover the same cells as the overall ATT — read a gap for them as a possible estimand difference.

sp.cs_jackknife(data, y, g, time, id, type='simple')

Delete-one-cluster (CV3) jackknife of a Callaway–Sant'Anna aggregate, against R didjack and Stata csdidjack.

Sensitivity

sp.honest_did(result, e, m_grid=None, method='smoothness')

sp.breakdown_m(result, e, method='smoothness')

Rambachan & Roth (2023). Accept any CausalResult carrying an event study; the fixed-length interval uses the joint covariance sp.event_study_vcov recovers, and result.attrs['interval'] records whether the returned table is the FLCI or the worst-case-bias fallback.

sp.bjs_pretrend_joint(result, data, ..., n_boot=300, seed=None)

Cluster-bootstrap joint Wald pre-trend test for BJS imputation results. Upgrades the default sum-of-z² test (which assumes pre-period independence) to the full covariance-aware test.

Visualisation

sp.ggdid(result, ax=None, ...)

aggte() visualiser with uniform-band overlay. Dispatches on result.model_info['aggregation'].