GitHub

Reference

TASC Estimation

TASC.StateSpaceParams — Type
StateSpaceParams(A, H, Q, R, m0, P0)

Parameters for the linear Gaussian state-space model

x_t = A * x_{t-1} + q_t,  q_t ~ N(0, Q)
y_t = H * x_t + r_t,      r_t ~ N(0, R)

where rows of Y are units and columns are time periods.

TASC.TASCResult — Type
TASCResult

Fitted Time-Aware Synthetic Control model. The treated unit is assumed to be the first row of the input matrix.

TASC.fit_tasc — Function
fit_tasc(Y; d, T0, n_em=100, n_post=0, tol=1e-4, q_diag=true, r_diag=true)

Fit Time-Aware Synthetic Control on the pre-intervention columns 1:T0. Y must be an N x T matrix. By default, row 1 is treated and rows 2:N are donors. Post-intervention treated entries are ignored during counterfactual prediction. Pass treated_rows to mask multiple treated units after T0.

TASC.predict_counterfactual — Function
predict_counterfactual(model, Y)

Use a fitted TASCResult to estimate the treated unit's untreated counterfactual path. The filter uses all units before T0; after T0, it uses only donor rows, then an RTS smoother borrows information across time.

Returns a named tuple with target, donors, variance, effect, state_mean, and state_covariance.

TASC.predict_post_intervention — Function
predict_post_intervention(model, Y)

Forecast the treated path after T0 by propagating the smoothed state at T0 forward with A, without using post-intervention donor observations.

TASC.TASCPlot — Type
TASCPlot

Lightweight plotting wrapper for RecipesBase/Plots.jl. Construct with tasc_plot(model, Y) and render with plot(...) after loading Plots.

Preprocessing

Core helpers:

  • panel_matrix(data, T0; target = 1, donors = nothing)
  • transform(M; method = :standard)
  • inverse_transform(M)
  • hsvt(X; rank = 2, p = 1.0)
  • denoise(M; num_sv, p = 1.0, filter_method = :HSVT, do_transform = false)
  • get_energy(s)
  • get_approx_rank(s; threshold = 0.95)

Synthetic Control Baselines

Core helpers:

  • fit_synthetic_control(pre_donor, pre_target; method = :ols, lambda = nothing, fit_intercept = true)
  • predict(model, donor)
  • score(model, donor, target)
  • predict_and_mse(model, donor, target_true)

Simulation

TASC.simulate_tasc — Function
simulate_tasc(; N=20, T=80, d=3, seed=1)

Generate synthetic panel data from the TASC state-space model. Returns (Y, params, signal), where Y and signal are N x T.

Additional generators:

  • gen_A, gen_H, gen_cov, and gen_dirichlet_params
  • generate_model_data and generate_multiple_layers
  • generate_rank_1_matrix and generate_rank_k_matrix
  • generate_sine_wave, generate_linear_dataset, generate_new_sine_dataset, generate_sine_dataset_A, and generate_sine_dataset_B
  • make_approx_low_rank