API Reference

All estimators follow a common two-step API:

solver = <SolverClass>(O, Z, ...)   # attach data
result = solver.fit(...)            # run the algorithm

Result Objects

Every fit() call returns a subclass of Result. The table below lists all attributes you can read off each result object.

Common attributes (available on every result)

Attribute	Description
result.tau	ATT estimate — scalar, or array for multi-treatment solvers.
result.baseline	Fitted counterfactual panel (N × T ndarray).
result.O	Original observed outcome panel (N × T).
result.Z	Original treatment mask (N × T).
result.std_tau	Standard deviation of tau (if inference was run).
result.covariance_tau	Covariance matrix of tau (if available).
result.inference_method	String describing the inference approach (or None).
result.residuals	O - baseline for all cells (computed property).
result.effect_matrix	Residuals restricted to treated cells (O - baseline) * (Z > 0).
result.z_pattern	Detected treatment pattern: 'block', 'staggered', or 'non-monotone'.
result.untreated_rmse	RMSE on all Z=0 cells (computed property).
result.control_rmse	RMSE on pure control units across all time periods (computed property).
result.pre_exposure_rmse	RMSE on all units before first treatment (computed property).
result.untreated_r2	R² on Z=0 cells — scale-free model-fit diagnostic.
result.summary()	Print a formatted diagnostics table.
result.plot_actual_vs_counterfactual(unit)	Interactive Plotly figure for one unit.

Estimator-specific attributes

Estimator	Attribute	Description
DID	row_fixed_effects	Unit FE vector (N,)
	column_fixed_effects	Time FE vector (T,)
	beta	Covariate coefficients (K,) — None when no covariates.
SDID	unit_weights	Donor unit weights w_sdid (N,). Nonzero entries = synthetic-control pool.
	time_weights	Time weights l_sdid (T,). Concentrated near treatment onset.
MC-NNM	M	Low-rank component of counterfactual (N×T)
	row_fixed_effects	Unit FE vector (N,)
	column_fixed_effects	Time FE vector (T,)
	beta	Covariate coefficients (K,) — None when no covariates.
DC-PR	M / baseline	De-biased counterfactual panel (N×T)
	std / std_tau	Sandwich SE for tau.
OLS SC	beta	List of simplex weight vectors, one per treated unit. result.beta[i] is a (n_control,) array.
	individual_te	Per-unit ATT list [unit_idx, tau_hat] or [unit_idx, tau_hat, p_val].
	control_units	Row indices of donor units (list[int])
	treatment_units	Row indices of treated units (list[int])
	V	Predictor importance weights (list) when covariates are used.
Covariance PCA	U	Left factor matrix (N × r ndarray)
RSC	M / baseline	Low-rank counterfactual panel (N×T)

Quick access examples:

# DID
result = DIDPanelSolver(O, Z).fit()
result.row_fixed_effects    # unit FE, shape (N,)
result.column_fixed_effects # time FE, shape (T,)

# SDID
result = SDIDPanelSolver(O, Z).fit()
result.unit_weights         # donor weights, shape (N,)
result.time_weights         # time weights, shape (T,)

# MC-NNM
result = MCNNMPanelSolver(O, Z).fit()
result.M                    # low-rank component only
result.row_fixed_effects    # unit FE, shape (N,)
result.column_fixed_effects # time FE, shape (T,)
result.beta                 # covariate coefs (or None)

# DC-PR
result = DCPanelSolver(O, Z).fit()
result.std                  # sandwich SE

# OLS Synthetic Control
result = OLSSCPanelSolver(O, Z).fit()
result.beta[0]              # donor weights for first treated unit
result.individual_te        # [[unit_idx, tau_hat], ...]
result.control_units        # donor row indices
result.treatment_units      # treated row indices

# Covariance PCA
result = CovariancePCAPanelSolver(O, Z).fit()
result.U                    # left factor matrix (N × r)

# RSC
result = RSCPanelSolver(O, Z).fit()
result.M                    # low-rank counterfactual

Difference-in-Differences (DID)

Synthetic Difference-in-Differences (SDID)

De-biased Convex Panel Regression (DC-PR)

Matrix Completion with Nuclear Norm Minimisation (MC-NNM)

Covariance PCA

OLS Synthetic Control (SC)

Robust Synthetic Control (RSC)

Data Utilities

Synthetic Data Generation

Use generate() to create a fully controlled panel from a low-rank factor model. The function returns (O, Z, tau_true) so ground-truth evaluation is always possible.

Semi-Synthetic Experiments

Use real panel data as a baseline, inject a known synthetic treatment effect, and benchmark estimators under controlled conditions.