DSC — Distributional Synthetic Controls on Dube (2019) ====================================================== .. currentmodule:: mlsynth Path-A reproduction of the Distributional Synthetic Controls application (Gunsilius 2023) on the Dube (2019) minimum-wage panel. The authors' reference is the ``DiSCo`` R package (`Davidvandijcke/DiSCos `_), whose vignette analyses exactly this data. DSC fits simplex-constrained weights on the **quantile functions** of micro-level distributions: each ``(unit, time)`` cell is a sample, and the treated unit's counterfactual quantile function is a weighted average of the donors' (Agueh-Carlier barycenter / optimal transport). Data ---- ``basedata/dube_minwage.csv`` -- the ``DiSCo`` package's ``dube`` dataset (Dube 2019; ``adj0contpov`` by state-year), exported from ``dube.rda`` and **subsampled to 250 observations per state-year cell** (fixed seed) so the micro-panel is ~1 MB rather than 15 MB. 34 states (33 donors) x 7 years (1998-2004); Alaska (``fips = 2``) treated from 2003, the vignette's ``id_col.target = 2``, ``t0 = 2003``. Result ------ ================================ ========== Quantity DSC ================================ ========== ATT (mean post QTE) −0.15 Pre-period 2-Wasserstein fit 0.13 Placebo permutation p (2003) 0.91 Placebo permutation p (2004) 0.32 Donors 33 ================================ ========== The headline cross-check against the vignette is the placebo-permutation result: both post-year p-values **exceed 0.05** -- the vignette's stated "no spurious effect" -- and the small pre-period Wasserstein confirms close distributional tracking before treatment. .. note:: **No live DiSCo cross-validation here.** The ``DiSCo`` R package does not install on this environment's R version, and the vignette's weight/QTE numbers live in figures rather than text, so a value-for-value run isn't reproducible in CI. This case is therefore Path A on the authors' exact dataset and setup, with mlsynth's deterministic output pinned and anchored to the one quantitative claim the vignette states (``p > 0.05``). The subsampling (250 obs/cell) keeps the panel small; it shifts point values slightly from the full-data run but preserves the distributional structure and the inference conclusion. Reproduce --------- .. code-block:: bash python benchmarks/run_benchmarks.py dsc_dube