NEWS

sdid 0.1.0

First release of sdid, which builds on synthdid 0.0.9 by Arkhangelsky, Athey, Hirshberg, Imbens, and Wager. BSD 3-Clause License.

sdid() estimates the ATT, following the Stata command sdid. It takes a long format panel, detects staggered adoption, and fits SDiD, SC, or DiD via method ("sdid", "sc", "did").
Inference must be chosen with vce: "bootstrap", "jackknife", "placebo", or "noinference". reps defaults to 50. seed makes replications reproducible without changing the session's RNG state.
sdid() prints a header, replication progress, and the ATT with its z statistic, p-value, and confidence interval at level. graph = FALSE skips the plot; quietly = TRUE prints nothing.
alternative sets a two-sided (default) or one-sided ("greater", "less") test. One-sided tests give one-sided p-values and intervals, labeled P>z or P<z.
Results have print, summary, coef, vcov, confint, nobs, weights, and plot methods, documented in ?sdid and ?sdid-methods.
tidy() and glance() methods work with modelsummary and broom. tidy(cohorts = TRUE) adds a row per adoption period, and labels names the rows.
Two vignettes, getting started and plotting, are rewritten around sdid().

Cohort effects are averaged with weights proportional to treated units times post-treatment periods (Clarke et al. 2024, Algorithm 1).
Bootstrap, jackknife, and placebo resample the whole panel jointly (Algorithms 2–4) and report cohort and overall standard errors.
Each cohort's z statistic, p-value, and interval are printed and stored in cohorts.
Reproduces the cohort estimates, overall effects, and jackknife SE of Clarke et al. (2024) on their gender quota data.

covariates takes time-varying covariates. covariate.method is "optimized" (default; coefficients estimated jointly with the weights) or "projected" (two-way fixed effects regression on untreated observations; Kranz 2022).
projection.sample sets the projected regression's sample: never-treated units (default, as in Stata) or all untreated observations.
The optimized method standardizes covariates by default (standardize = TRUE), warns on reaching max.iter (class sdid_max_iter_warning), and errors if coefficients diverge.
beta is named by covariate. With the optimized method and staggered adoption, it is a covariate × cohort matrix.
Standard errors re-estimate the coefficients in every replication.

sdid_plot(), sdid_units_plot(), sdid_placebo(), and sdid_placebo_plot() replace their synthdid_* versions and take sdid() results. Staggered designs get one panel per cohort.
sdid_plot() styles: "minimal" (default) and "effects", which adds a simplified DiD diagram.
Axes get about five ticks covering the data. Whole-number time axes get whole-number ticks only. x.breaks and y.breaks set ticks manually.
sdid_plot() and sdid_units_plot() accept a named list of single-cohort results, one panel each. facet and facet.vertical are removed; lambda.comparable defaults to FALSE.
Stacked panels have open frames, their own time-weight axes, and cohort labels at the adoption period. A red dashed line marks treatment.
sdid_units_plot() is redesigned: type = "scatter" (default) plots each control's weight against its own estimate; type = "bars" shows weighted controls ordered by weight. New arguments: label.units, cohort, weighted.only. Uses ggrepel when installed. negligible.threshold, negligible.alpha, and units are removed.
sdid_placebo() and sdid_placebo_plot() gain placebo.time to choose when the in-time placebo starts. Single-adoption designs only.
Plotting no longer attaches ggplot2 and requires ggplot2 >= 3.4.0.

Unit and time weights are estimated in C++. Bootstrap, jackknife, and placebo replications run in parallel, with identical results for any thread count under the same seed. Set threads with sdid.num.threads or num.threads.
Long weight solves switch to a Gram matrix gradient, much faster on large panels. Controlled by sdid.solver and sdid.gram.max.mb.
Replications can be interrupted with Ctrl-C.

Tibbles and data.tables are accepted. They previously failed with "There is no variation in treatment status".
Fixed a line-search edge case that could produce NaN weights.
The in-time placebo for SC no longer fails with a duplicated omega.intercept argument.
Replications use the fitted estimator's sparsify settings, not the defaults.
A treatment column stored as doubles no longer warns.
The PENN documentation gives the correct number of rows.

synthdid_estimate(), sc_estimate(), did_estimate(), panel.matrices(), synthdid_se(), synthdid_effect_curve(), and timesteps() are no longer exported.
synthdid_rmse_plot(), synthdid_controls(), simulate_dgp(), estimate_dgp(), and lindsey_density_estimate() are removed.
Underlying estimates (in estimates and from sdid_placebo()) have class sdid_estimate and print on one line without computing an SE. Their summary() and format() methods are removed, and the methods dependency is dropped.
The original plot style, standard errors in plots, and several estimates in one shared panel are removed.
Options are renamed with the sdid. prefix. The verbosity option is removed; use quietly.