5  Callaway-Sant’Anna — didgpu_cs()

6 Callaway-Sant’Anna staggered DiD — didgpu_cs()

didgpu_cs() implements the staggered-adoption DiD estimator of @callaway2021difference, with all three inner estimators (OR, IPW, DR), all four aggregations (event / group / calendar / overall), and both bootstrap variants (cluster and multiplier).

6.1 Function signature

didgpu_cs(
  df,
  outcome, group, time, treatment,
  est_method     = c("DR", "OR", "IPW"),
  control_group  = c("notyet", "never"),
  aggregation    = c("event", "group", "calendar", "overall"),
  covariates     = NULL,
  weight         = NULL,
  bootstrap_reps = 0L,
  bootstrap_kind = c("cluster", "multiplier"),
  cluster        = NULL,
  ci_level       = 95,
  backend        = "auto",
  seed           = NULL,
  verbose        = TRUE
)

6.2 What it estimates

For each group-time \((g, t)\) with \(g\) = first-treated period and \(t\) = calendar period, the conditional ATT \(ATT(g, t) = \mathbb{E}\bigl[Y_t(1) - Y_t(0) \mid G_g = 1\bigr]\) is identified by a 2×2 DiD against either the not-yet-treated or the never-treated control group.

The package returns the full att_gt matrix, then aggregates it into the published event-study / group-time / calendar / overall quantities.

6.3 GPU acceleration

All three inner regressions (OR, IPW, DR) run on the GPU (src/cuda_cs_inner.cu) with per-row influence functions:

  • OR uses an in-thread Cholesky per (g, t) cell;
  • IPW / DR add a per-cell IRLS logistic propensity model that replicates stats::glm.fit to ~1e-8 (verified in tests/testthat/test-cuda-cs-wiring.R);
  • DR layers on the outcome-regression augmentation.

Per-cell influence functions also let the cluster and multiplier bootstrap SEs run entirely on the GPU.

Tip

Full chapter with worked examples, covariates, and the four aggregation shapes is in progress. The function documentation (?didgpu_cs) is complete and the README’s “Method families” section sketches typical usage.

6.4 See also

  • @callaway2021difference — the method paper.
  • did — the reference R implementation didgpu_cs targets.

6.5 References