Lee-Wooldridge Difference-in-Differences via Unit-Specific Transformations • lwdidR

lwdidR is an R implementation of the Lee and Wooldridge (2026a, 2026b) panel difference-in-differences estimator. The estimator first transforms outcomes using only each unit’s pre-treatment observations, then estimates the treatment effect in a cross section.

Installation

# Install from GitHub
devtools::install_github("xiangao/lwdidR")

Overview

lwdidR implements the rolling difference-in-differences method of Lee and Wooldridge, drawing on two companion papers:

Lee, S. J., & Wooldridge, J. M. (2026a). Simple Transformation Approach to Difference-in-Differences Estimation for Panel Data. Journal of Business & Economic Statistics, 1–27. https://doi.org/10.1080/07350015.2026.2683047

Lee, S. J., & Wooldridge, J. M. (2026b). Simple Approaches to Inference with Difference-in-Differences Estimators with Small Cross-Sectional Sample Sizes. Working paper, SSRN 5325686. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5325686

The outcome transformation underlying the estimator is introduced in (2026a); the small-sample inference options (wild cluster bootstrap, permutation) and the staggered Castle Doctrine application replicated here come from (2026b). lwdidR is an R port of the original Stata lwdid command by Soo Jeong Lee and Jeffrey M. Wooldridge (https://github.com/Soo-econ/lwdid), whose v1.0 is based on (2026b).

The key idea is simple: residualize each unit’s outcome using its own pre-treatment observations, either by demeaning or detrending, and then run a pooled cross-sectional OLS.

The package supports: - Common-timing and staggered adoption designs - Transformation methods: demean, detrend, demeanq, detrendq - Standard errors: homoskedastic OLS, HC1, HC3, cluster-robust, wild cluster bootstrap, permutation - Cohort-specific ATTs and (g,r)-level event-study estimates - Large-N estimators (method = "ra"/"ipw"/"ipwra") with WATT(r) event study, wild cluster bootstrap, and sup-t bands

Large-N path (Lee & Wooldridge 2026a)

Setting method to "ra", "ipw", or "ipwra" activates the large-N path: it estimates ATT(g,t) by cohort and period (regression adjustment, inverse-probability weighting, or doubly-robust IPWRA), aggregates to the event-study WATT(r) path plus Pre/Post averages, and reports wild cluster bootstrap standard errors with simultaneous (sup-t) confidence bands.

# doubly-robust IPWRA, detrending transformation, with covariates
res <- lwdid(walmart, "log_wholesale_emp", "cid", "year", gvar = "first_year",
             rolling = "detrend", method = "ipwra",
             controls = c("x1", "x2", "x3"), reps = 999, seed = 1)
print(res)
plot(res)          # event-study plot with sup-t bands (requires ggplot2)

Large-N options: method, pre (pre-periods used in the transform), never (never-treated controls only), attgt (return ATT(g,t) cells), ydot (return transformed outcomes), reps/seed (bootstrap), level, nose.

Scope. The small-N path (default; Lee & Wooldridge 2026b) and the large-N path (method=; Lee & Wooldridge 2026a) are both implemented. The large-N estimators (RA/IPW/IPWRA), WATT(r) aggregation, and inference are validated against the Stata lwdid v2.4.2 output on lw_smoking and lw_walmart (see data-raw/stata-reference/).

Common-timing designs: `post` vs `dvar`

For common-timing designs (gvar = NULL), supply a binary post indicator. post can be used two ways, and lwdid() distinguishes them:

If post is the treatment-on indicator D_it (1 only for treated units in post periods), the ever-post units are the treated group — no extra argument needed (backwards-compatible behavior).
If post is a calendar post indicator (1 for all units once t ≥ first treated period), the treatment group lives in a separate column. Pass it via the new dvar argument, e.g. lwdid(..., post = "post", dvar = "treated"). If dvar is omitted, lwdid() auto-detects an unambiguous unit-invariant treat/treated/D/d column and errors if it cannot resolve one.

Quick Start

library(lwdidR)

# Load bundled Castle Doctrine dataset (50 US states, 2000-2010)
castle <- read.csv(system.file("extdata", "castle.csv", package = "lwdidR"))
castle$gvar <- castle$effyear
castle$gvar[is.na(castle$gvar) | castle$gvar == 0] <- NA

# Staggered design: demeaning with HC3 SEs
res <- lwdid(castle, "lhomicide", "sid", "year",
             gvar = "gvar", rolling = "demean", vce = "hc3")
print(res)

Lee-Wooldridge DiD (lwdidR)
Design:      staggered
Transf.:     demean
VCE:         HC3
--------------------------------------------------
Overall ATT:   0.0917
SE:            0.0612
t-stat:        1.4995
p-value:       0.1376

Replication: Lee & Wooldridge (2026b), Section 7.2

Castle Doctrine laws and log homicide rates:

Method	ATT	SE	t-stat
Demeaning (OLS SE)	0.0917	0.0571	1.607
Demeaning (HC3 SE)	0.0917	0.0612	1.500
Detrending (HC3 SE)	0.0666	0.0550	1.210

All results match the Castle Doctrine estimates reported in the text of Section 7.2 of Lee & Wooldridge (2026b) within tolerance 0.001 (data set used by Cunningham 2021). See vignette("castle_law") for full replication.

Main functions

Function	Description
`lwdid()`	Main estimator (common-timing and staggered)
`print.lwdid()`	Compact results display
`summary.lwdid()`	Results + period/cohort details

Documentation & vignettes

Full documentation: https://xiangao.github.io/lwdidR/

Page	Description
Castle Law replication	Full Lee-Wooldridge Section 7.2 replication
Simulation comparison	Simulation comparison across transformations
`lwdid()`	Main estimator
Reference index	All documented functions on one page

Algorithm Notes

Overall ATT (staggered): Not a delta-method average of cohort ATTs. Instead, a pooled cross-sectional regression: - Treated unit i: outcome = ydot_postavg for its own cohort g - Never-treated unit i: outcome = weighted average of ydot_postavg across all cohorts - Single OLS on N_treated + N_never_treated units

This matches the Python reference implementation and yields SE ≈ 0.061 (not 0.051 from naive averaging).

Citation

If you use lwdidR, please cite the Lee & Wooldridge papers and acknowledge the original Stata package.