4 Dynamic DiD — `didgpu()`

5 Dynamic difference-in-differences — `didgpu()`

didgpu() implements the heterogeneity-robust event-study DiD estimator of @dechaisemartin2024dynamic. It is the flagship estimator in this package and bit-for-bit numerically equivalent to the reference R package DIDmultiplegtDYN across every commonly-used option.

5.1 What it estimates

For each event-time \(k \in \{1, 2, \dots, \text{effects}\}\), the per-event-time DID is the average causal effect of being exposed to a strictly higher treatment dose for \(k\) periods relative to a clean, not-yet-switched comparison group within the same baseline-treatment cohort. The estimator pools across “switcher-in” (treatment goes up) and “switcher-out” (treatment goes down) directions via Neyman-weighted averaging.

The published Effects matrix has one row per event-time, with:

Estimate, SE, LB.CI, UB.CI — the point estimate and its inference;
N (unweighted observations), Switchers (unweighted switcher cells), N.w (weighted observation mass), Switchers.w (weighted switcher mass) — the four sample-size columns the reference reports.

A parallel Placebos matrix reports pre-treatment falsification effects at horizons \(-1, -2, \dots, -\text{placebo}\).

5.2 Function signature

didgpu(
  df,
  outcome, group, time, treatment,
  effects               = 1L,
  placebo               = 0L,
  switchers             = "",                # "" / "in" / "out"
  same_switchers        = FALSE,
  same_switchers_pl     = FALSE,
  normalized            = FALSE,
  only_never_switchers  = FALSE,
  dont_drop_larger_lower= FALSE,
  trends_lin            = FALSE,
  trends_nonparam       = NULL,
  controls              = NULL,
  weight                = NULL,
  continuous            = NULL,
  predict_het           = NULL,
  bootstrap_reps        = 0L,
  cluster               = NULL,
  ci_level              = 95,
  backend               = "auto",            # "auto"/"r"/"cpu"/"cuda"/"reference"
  checkpoint_dir        = NULL,
  seed                  = NULL,
  n_workers             = NULL,
  verbose               = TRUE
)

5.3 Common option recipes

5.3.1 Heterogeneity by group covariate

Tip

Full chapter content with worked examples is in progress. The options above each correspond to a section of the reference paper @dechaisemartin2024dynamic; this chapter will expand to a worked example per option. In the meantime, the function documentation (?didgpu) is complete and the package README links every option to the relevant section of the paper.

5.4 See also

@dechaisemartin2024dynamic — the method paper.
DIDmultiplegtDYN — the reference R implementation didgpu targets bit-for-bit.