didgpu

GPU-accelerated, checkpointed difference-in-differences for econometric scale

Welcome to didgpu

didgpu is a single R package that gives you five published heterogeneity-robust difference-in-differences estimator families, each with:

• ⚡ a CUDA backend that drops checkpointed bootstrap times from hours to minutes on a single GPU;
• 💾 per-cell checkpointing to disk, so a 12-hour cluster bootstrap that crashes at hour 11 resumes from cell 10,994 instead of starting over;
• 🎯 bit-for-bit numerical equivalence with the reference R packages (DIDmultiplegtDYN, did, fect, …) verified by a continuous randomized differential-test suite of tens of thousands of panels.

You can use any of the estimators from a single, consistent interface, mix GPU and CPU backends per call, and resume any long-running fit after a crash, an OOM, or a reboot.

0.1 The five estimator families

Family	Function	Reference
Dynamic, heterogeneity-robust DiD	didgpu()	de Chaisemartin & D’Haultfoeuille (2024)
Callaway-Sant’Anna staggered DiD	didgpu_cs()	Callaway & Sant’Anna (2021)
Fixed-effects counterfactual (FEct/IFEct/MC)	didgpu_fect()	Liu, Wang & Xu (2024); Athey et al. (2021)
Instantaneous DID_M (non-absorbing binary)	didgpu_did_static()	de Chaisemartin & D’Haultfoeuille (2020)
Continuous treatment, no stayers	didgpu_did_continuous()	de Chaisemartin & D’Haultfoeuille (2024)

Each one has a chapter in this book.

0.2 Why “GPU”?

The CUDA backend isn’t a gimmick — it is the difference between a bootstrap-CI workflow that’s iterated overnight and one you can re-run in the time it takes to refill a coffee mug. On a single RTX-class card the typical speedup is 10×–100× for the cluster bootstrap and 5×–30× for the per-cell DID kernel, depending on panel size.

You do not need a GPU to use didgpu. All five estimators have pure-R and Rcpp fallbacks; the GPU backend is optional and detected at install time. CRAN’s build farm does not have a GPU and the package still installs and runs correctly there.

0.3 Why “checkpointed”?

Long bootstrap runs will die at some point — OOM, power outage, Windows update reboot, network blip on a slurm node. didgpu writes one small RDS file per bootstrap replicate. If the run dies, you re-launch exactly the same command and it picks up from where it stopped. No code changes, no tryCatch plumbing, no babysitting.

0.4 How to read this book

• Part I (Getting Started) — install the package, run your first estimator. Start here if you’ve never used didgpu.
• Part II (Estimators) — one chapter per estimator family. Each chapter shows the math behind it, the function signature, common options, and an end-to-end worked example.
• Part III (Advanced) — GPU setup, sensitivity analysis, reproducibility/checkpointing, bootstrap variants. Reach for these once you’re comfortable with at least one estimator.
• Part IV (Reference) — methodology notes (the math), FAQ, and the changelog.

0.5 Cite

If didgpu played a role in published work, please cite both this package and the relevant method papers:

citation("didgpu")

Tip

The package print method always reminds you which method paper goes with the estimator you just ran. There is a one-line citation() for each estimator family.

0.6 Status

didgpu is actively maintained. Source and issue tracker: https://github.com/JoshuaAmmons/didgpu.

The package’s numerical correctness against the reference R packages is re-tested on every push via a large randomized differential testing suite. The current status badge:

✅ Last full sweep: ~17,000 informative comparisons, 0 failures
✅ Main 1285-case battery (all backends, all flags): clean
✅ Bit-for-bit equivalence with DIDmultiplegtDYN 2.3.x across the
   weighted x flag matrix