DiD result objects: what is where

Every DiD estimator returns a CausalResult. The headline is on the object (.estimate, .se, .ci, .pvalue, .n_obs); the tables and settings behind it are in .detail and the .model_info dict. This page lists those keys for the main estimators, so scripts do not have to guess.

Reading the tables: three accessors and tidy()

You rarely need model_info for the tables themselves:

accessor	returns	available on
result.attgt	ATT(g,t): group, time, relative_time, att, se, ci_lower, ci_upper, pvalue	callaway_santanna (and sp.did when it routes there), etwfe, wooldridge_did
result.event_study	one row per relative time: relative_time, att, se, ci_lower, ci_upper, pvalue	callaway_santanna, aggte(type='dynamic'), sun_abraham, did_imputation(horizon=...), gardner_did(event_study=True), stacked_did, event_study, etwfe_emfx(type='event')
result.aggregations	dict of tables: 'overall', plus 'dynamic', 'group', 'calendar' when the result holds them	every staggered fit above
result.tidy()	long table: term, estimate, std_error, statistic, p_value, conf_low, conf_high, type	every result

The accessors only read what the fit stored. They never refit. When a table is not there they raise ResultAccessorUnavailable, which is an AttributeError (so hasattr(result, "attgt") is False) and carries a recovery_hint naming the estimators that provide it. Two cases to know:

• etwfe / wooldridge_did store cohort-by-relative-time cells, which are ATT(g,t), not an event study. result.attgt reads them; for the event study call sp.etwfe_emfx(result, type='event').
• sp.aggte returns the aggregation, not the cells. Read .attgt on the fit you passed to aggte. aggregations does not compute the missing types; call sp.aggte(result, type=...) for those.

Intervals and p-values are each estimator's own. Where a stored table has none (gardner_did's dict, the etwfe cells) the accessors add the normal interval at the fit's alpha, the convention those estimators use.

tidy() has the same first seven columns on CausalResult and on EconometricResults (sp.regress, sp.feols, ...). CausalResult adds type ('main', 'event_study', 'group_time').

model_info keys

The tables below list the keys a fit with default options carries. Options add keys (for example calendar_time for date-typed time columns, survey_design with design=, boot_seed with a bootstrap). Keys that start with _ are internal and may change without notice. tests/test_did_result_accessors.py checks that every key listed in the four tables below is present on a default fit.

sp.callaway_santanna

.detail is the ATT(g,t) table (group, time, att, se, ci_lower, ci_upper, pvalue, relative_time). .estimate is the simple aggregation.

key	contents
estimator	2x2 engine per cell ('DR', 'IPW', 'REG')
control_group	'nevertreated' or 'notyettreated'
notyet_cutoff	not-yet-treated cutoff rule
base_period	'universal' or 'varying'
anticipation	anticipation periods
parallel_trends	dict describing the parallel-trends assumption used
pscore_trim	propensity-score trimming level
n_pscore_trimmed	units trimmed
weights	weights column or None
weighted	whether weights were used
covariates	covariate list or None
n_units	units in the panel
n_treated_units	eventually treated units
n_periods	periods
n_cohorts	treated cohorts
cohorts	list of cohorts
cohort_sizes	units per cohort
comparison_only_cohorts	cohorts used only as comparisons
dropped_gt_cells	(g,t) cells that could not be estimated
event_study	dynamic aggregation, a DataFrame (relative_time, att, se, ci_lower, ci_upper, pvalue)
pretrend_test	dict: statistic, df, n_cells, rank, rank_deficient, pvalue
clustervars	cluster variable(s) or None
bstrap	whether the multiplier bootstrap ran
biters	bootstrap draws (0 when analytic)
cband	uniform bands requested
boot_weight_type	multiplier weights
crit_val_uniform	uniform-band critical value (or None)
se_method	'analytic' or 'bootstrap'
inference_note	one-line description of the inference used
balance	panel balance handling

sp.aggte

.detail is the aggregated table. Keys: aggregation ('simple', 'dynamic', 'group', 'calendar'), balance_e, min_e, max_e, bstrap, n_boot, random_state, cband, crit_val_uniform, share_variance, base_period, vcov, vcov_source, cell_labels, influence_functions, overall_influence_function, vcv_pre, pre_event_times, source_method, inference_note, n_units; with type='dynamic' also event_study.

sp.sun_abraham

.detail is the event-study table.

key	contents
estimator	'Sun-Abraham IW'
headline	what .estimate is
other_aggregation	the other headline, with its value
summary_aggregation	'event_time' or 'fixest_att'
att_event_time	equal-weight mean of post-period effects
se_event_time	its SE
att_fixest_att	cell-weighted ATT (fixest summary(agg = "att"))
se_fixest_att	its SE
se_fixest_att_share	its SE with cohort-share variance
event_study	event-study DataFrame
event_times	relative times estimated
vcov_event_time	covariance of the event-time coefficients
pretrend_test	dict: statistic, df, pvalue, relative_times
pretest	pre-trend test type
pretest_periods	periods in the test
event_window	reported window
window_rule	'report', 'bin' or 'reference'
control_group	comparison group
control_cohort	reference cohort
n_cohorts	treated cohorts
cohorts	list of cohorts
anticipation	anticipation periods
ref_p	reference relative time(s)
dropped_unidentified_cohorts	cohorts dropped as unidentified
n_units_dropped_unidentified	their units
n_rows_dropped_unidentified	their rows
share_variance	cohort-share variance included
se_type	variance description
n_clusters	clusters
n_coeffs	interaction coefficients
dof_K	parameters counted for the small-sample factor
dof_fe_not_nested	fixed-effect parameters not nested in clusters
dof_convention	small-sample factor rule
fixed_effects	absorbed fixed effects
panel	panel input

sp.did_imputation

.detail is the event-study table when horizon= or pretrends= is given, else None.

key	contents
estimator	'BJS Imputation'
n_treated_obs	treated observations
n_control_obs	untreated observations (the Y(0) fit)
n_units	units
n_time_periods	periods
n_never_treated	never-treated units
cluster_var	cluster column
vce	'analytic' or 'bootstrap'
weights	weights column or None
y0_fe	fixed effects of the Y(0) model
y0_covariates	covariates of the Y(0) model
horizon_requested	whether horizon= was given

With horizon= or pretrends= it adds event_study, event_study_vcov, pretrend_method and event_study_convention; with pretrends= also pretrend_test and pretrends.

sp.drdid

.detail is a two-column summary (statistic, value).

key	contents
method	'improved' or 'traditional'
est_method	'dr', 'ipw', 'reg' or 'twfe'
engine	the DRDID estimator reproduced
drdid_reference	DRDID::<engine>
panel	True with id=, else repeated cross-sections
n_treated	treated units / observations
n_control	comparison units / observations
n_pre	pre-period rows
n_post	post-period rows
ps_mean_treated	mean propensity score, treated
ps_mean_control	mean propensity score, comparison
se_method	'influence_function'
covariates	covariate list

The panel path (id=) adds id, n_units, ps_flag and n_boot (unused); the repeated cross-section path adds variance_convention.

sp.gardner_did, sp.stacked_did, sp.etwfe

• gardner_did: event_study is a dict (horizon, coef, se, ci, n_obs, vcov) keyed by 'D_k+h' labels, present with event_study=True; vcov and cell_labels hold the full covariance. result.event_study turns it into the standard table.
• stacked_did: event_study (DataFrame), event_study_vcov, event_study_betas, event_study_se, cohorts, skipped_cohorts, window, cluster_var, cluster_stack, n_clusters, aggregation and aggregation_weights (the weight of each post event time in the headline ATT under aggregation='treated_obs', else None).
• etwfe: event_study holds cohort-by-relative-time cells (cohort, rel_time, period, estimate, se, ...), read through result.attgt; event_vcov is their covariance, cohort_vcov that of the per-cohort ATTs in .detail.