Comparison of Augmented Synthetic Control Method (ASCM)

This notebook compares the implementation of Augmented Synthetic Control in causalis and pysyncon.

This notebook presents the ascm comparison research workflow and key analysis steps.

Data Generation

We use the generate_scm_gamma_26 function to generate synthetic panel data.

Result

Ground-truth ATTE is 1.919269

	unit_id	calendar_time	y	y_cf	mu_cf	mu_treated
0	donor_1	2000-01	9.971611	9.971611	10.059081	10.059081
1	donor_1	2000-02	10.344379	10.344379	10.097467	10.097467
2	donor_1	2000-03	10.998498	10.998498	11.354098	11.354098
3	donor_1	2000-04	11.508717	11.508717	11.277716	11.277716
4	donor_1	2000-05	11.125281	11.125281	10.098791	10.098791

We use the AugmentedSyntheticControl model with lambda_aug=50.

Result

Causalis ATTE: 3.945634

	value
field
estimand	average_post_effect
model	AugmentedSyntheticControl
inference	average_att_ttest
value	3.9456 (ci_abs: 0.3664, 7.5249)
value_relative	16.0707 (ci_rel: 1.4922, 30.6491)
alpha	0.0500
p_value	0.0417
is_significant	True
post_outcome_d_mean	28.3185
pointwise_post_period_average	3.7668
effect_by_time	[{'period': 2003-02, 'estimate': 3.32949405792...
cumulative_effect	22.6007
n_significant_periods	0
n_donors	40
n_pre_periods	37
n_post_periods	6
time	2026-03-14

We prepare the Dataprep object and fit the AugSynth model.

Result

PySyncon ATTE: 3.766599

Result

Method ATTE Error 0 Ground Truth 1.919269 0.000000 1 Causalis 3.945634 2.026365 2 PySyncon 3.766599 1.847330

The estimates differ because:

Time Period Alignment: causalis automatically correctly identifies the pre-treatment period for the treated unit. In pysyncon, manually specifying pre_periods requires filtering by the treated unit to avoid including periods where donors are "pre-treatment" but the treated unit is already treated.
De-meaning (Intercept): pysyncon performs row-wise (cross-sectional) de-meaning by the donor average at each time point before applying Ridge. This effectively includes a time-specific intercept $\alpha_t$ . This also alters the direction of the Ridge penalty.
Constraint Specification: causalis explicitly enforces the sum-to-one constraint on augmented weights via a Lagrange multiplier in the original space, while pysyncon relies on cross-sectional de-meaning to satisfy the constraint (which is mathematically equivalent to solving with an intercept).