The Chained Difference-in-Differences

Estimates group-time average treatment effects in panels with missing outcomes by the chained difference-in-differences method of Bellégo, Benatia and Dortet-Bernadet (2025) . The package extends 'did' and retains the att_gt() and aggte() interfaces. Chained estimation combines observed changes within units by the generalized method of moments. It supports inverse probability weighting, outcome regression and doubly robust estimation. Missing outcomes can be handled by observation weighting, imputation of outcome changes, or an augmented combination of both. Users can choose comparison groups and parallel trends assumptions, fit linear cohort-by-date fixed effects, test trends before treatment, and inspect fitted regression models. A separate analysis estimates the effects of starting and stopping a binary treatment under no carryover. Ordinary mode uses the estimators of Callaway and Sant'Anna (2021) .


cdid cdid logo

cdid brings widely used difference-in-differences estimators into a common framework for complete and incomplete panels. It reproduces documented specifications of established methods and extends them with chained comparisons, observation weighting, change imputation and doubly robust corrections for selective missingness, by leveraging the chained difference-in-differences (chained DiD).

CRAN package page · Release notes · User guide · Author website

This guide describes cdid 1.3.0.1. Articles label the version used for recorded output. Check your installed version with packageVersion("cdid") and consult the release notes.

Installation

Install the current CRAN release in a normal R session:

install.packages("cdid")
library(cdid)

Install 1.3.0.1 from GitHub when the CRAN release is older or when you want to follow the main branch:

# Run once if remotes is not installed:
install.packages("remotes")
remotes::install_github("dvbn/cdid")
library(cdid)

The CRAN package includes a short vignette covering an incomplete panel, estimation, diagnostics, aggregation and plotting. To build it when installing from GitHub, add build_vignettes = TRUE and have knitr, rmarkdown and Pandoc available. Open it with vignette("cdid", package = "cdid").

Requirements and build tools

cdid requires R 4.1.0 or later and the CRAN packages did 2.5.1 or later and DRDID 1.4.0 or later. It refuses to load with earlier versions of did or DRDID. If an error reports an outdated dependency, run install.packages(c("did", "DRDID")) and restart R. You do not need GitHub versions of did or DRDID. Modified forks can have different interfaces even when their version numbers exceed the minimum.

Installing from source compiles the package's C code, so the usual tools for building R packages from source must be available:

  • Windows: Rtools matching your R version.
  • macOS: Xcode Command Line Tools, see R for macOS tools.
  • Linux: a C/C++ compiler and make, including tools needed by dependencies.

Use a standard CRAN installation of R. If a CRAN mirror or platform binary still offers an older release, install the current source from the main CRAN server with install.packages("cdid", repos = "https://cran.r-project.org", type = "source"). The optional pinned validation setup installs exact dependency versions for checks of a revision.

An abrupt "R session aborted" can come from incompatible local R installations and compiled packages. Check which R version RStudio uses and reinstall the affected dependencies with that R installation. For help, include sessionInfo() and the installation output in an issue.

Get started

The example uses the mpdta county panel shipped with did. Treatment is adopted once and kept, so first.treat holds each county's adoption year, with 0 for counties never treated. That column is the cohort variable gname.

fit <- cdid::att_gt(
  yname = "lemp", tname = "year", idname = "countyreal",
  gname = "first.treat", xformla = ~lpop, data = did::mpdta,
  bstrap = FALSE, cband = FALSE
)
summary(fit)

event_study <- cdid::aggte(fit, type = "dynamic", bstrap = FALSE, cband = FALSE)
cdid::ggdid(event_study)

summary(fit) shows the ATT(g,t) table under a heading that names the chained estimator, then a short record: the available cells and the reasons for unavailable ones, the unit table, the handling of missing outcomes, the parallel trends assumption, the GMM weighting and the inference performed. aggte() summarizes the stored cells without refitting, here by time since adoption. cdid::tidy(fit) and cdid::glance(fit) give the same information as data frames. The defaults are those of did: doubly robust treatment adjustment, never-treated controls and no anticipation. bstrap = FALSE requests analytical standard errors for fast interactive use.

Get started walks through a simulated panel with recorded output, the four aggregation types, plots, export and a comparison with did. The practical notebook lets you edit and rerun every example. The user guide orders the remaining pages as a reading path.

What the package does

A treatment cohort is the set of units first treated in the same period. For cohort g and period t, ATT(g,t) is the average effect of the treatment in period t on the units of cohort g. Callaway and Sant'Anna (2021, Journal of Econometrics) define these effects for staggered adoption, estimate each one by comparing the cohort with units never treated or not yet treated, and aggregate them into summaries such as event studies. The did package implements their estimators, which adjust for covariates by outcome regression (OR), inverse probability weighting (IPW) or a doubly robust (DR) combination of the two.

cdid extends this framework to incomplete panels with the chained DiD of Bellégo, Benatia and Dortet-Bernadet (2025, Journal of Econometrics). did handles an unbalanced panel either by keeping only the units observed at every date or, with allow_unbalanced_panel = TRUE, through repeated cross section (RCS) estimators that compare outcome levels. Chained DiD uses the changes of each unit between the dates at which it is observed and combines all available short comparisons into the effects by the generalized method of moments (GMM). The targets remain the ATT(g,t) of the cohorts of the original study population, including their units with missing outcomes, and the assumptions concern the missing changes rather than the missing levels.

Why changes within units

Write the outcome of unit i at date t as $Y_{it} = \alpha_i + u_{it}$, where $\alpha_i$ is a fixed individual effect and $u_{it}$ holds everything that varies over time: trends, the treatment effect and noise. The change between two observed dates, $Y_{it} - Y_{is} = u_{it} - u_{is}$, contains no $\alpha_i$. Suppose treated units with persistently low outcomes are more likely to leave the panel while controls leave at random. The treated units still observed later have a higher mean $\alpha_i$, so comparing their outcome levels with earlier ones adds this change in composition to the treatment effect. If attrition depends only on $\alpha_i$ and $\alpha_i$ is unrelated to $u_{it} - u_{is}$, the treated units that remain still have the mean change of their cohort.

This is the main advantage of chained DiD for incomplete panels: observation may depend on the unobservable individual effect, as long as that effect is unrelated to the changes, and the observed levels need not stay representative of the population. Removing the fixed effect does not by itself make missing changes ignorable. If the persistent component is related to trends or to heterogeneous treatment effects, observation that depends on it also selects the changes.

The assumptions

Within a cohort, the units observed at both dates of a comparison must be representative of that cohort with respect to the mean outcome change over that pair, directly or conditional on covariates with an appropriate adjustment. This equates population expectations of changes. It does not require every unit to have the same change or realized sample means to agree. Parallel trends, no anticipation and overlap must also hold. Handling missing outcomes, Parallel trends and Treatment adjustment state these assumptions.

Neither approach is better in every design. When the sample at each date represents the population, comparisons of levels can be valid too. Identify effects when the observed units change composition works through a design in which changes identify the effects and levels do not. Differencing removes the persistent unit component, while the transitory noise of each short change accumulates along a chain, so which approach is more precise depends on the design, as Bellégo, Benatia and Dortet-Bernadet (2025, Section 2.1.3 and Proposition 1) analyze. How to compare precision explains how to compare them on matched targets.

Two modes

cdid imports did and keeps its function names and argument order: att_gt(), aggte(), ggdid(), tidy(), glance() and conditional_did_pretest(). Most upstream defaults are retained, including treatment DR in att_gt().

  • Chained mode is the default for panel data. It estimates the usable short comparisons between pairs of dates at which units are observed and combines them by GMM. An effect that these comparisons cannot recover, given cohort and control support and the assumptions, is reported as unavailable with a reason, never as zero.
  • Ordinary mode, requested with mode = "ordinary" and used automatically for repeated cross sections, passes the call unchanged to did. allow_unbalanced_panel = NULL resolves according to the mode. Moving from did explains the differences and the added arguments.

Data and model decisions

The decisions below come in the order you meet them. Each links to the guide page that states the assumption behind the choice.

1. Prepare the panel and the unit table

Supply a long data frame with one row per unit and date: a unit identifier, a finite numeric date, a numeric outcome with missing values kept as NA, and the treatment column. Never replace a missing outcome with zero or a carried-forward value. The units in data form the study population. To include study units with no observed outcome, pass a unit table with one row per unit through frame. Units observed once or never supply no change but stay in the cohort shares, the observation models and the standard errors.

Covariates that change over time are read at each cohort's reference date from the dated rows in data. A unit can contribute its change from years 3 to 4 even if its year-2 outcome is missing. If the comparison uses a covariate from year 2 and you know its value, supply it on the year-2 row with y = NA. If the covariate is unknown, att_gt() warns and excludes the unit from that cohort's comparisons under the default missing_reference = "exclude". The unit can still contribute to other cohorts for which its required covariates are known. If treated units are excluded, the effect describes the remaining treated units. Set missing_reference = "error" to stop and list the missing inputs. Sampling weights that change within a unit need an explicit fix_weights rule.

Prepare your panel and unit table · Add a known reference covariate row · Investigate missingness and data structure

2. Choose the comparison group and the parallel trends assumption

control_group selects never-treated or not-yet-treated controls, as in did. pt_assumption = "post", the default, assumes parallel trends from each cohort's reference period onward. "all" adds the same assumption from a chosen common start date pt_start up to each reference period, so that earlier dates can also serve as baselines. "ctrl" keeps PT-post and adds common untreated trends among the clean control cohorts only. Valid added restrictions can make more effects available or improve precision, but neither gain is guaranteed, and false restrictions bias the estimates. base_period changes only how cells before treatment are reported, not the assumption.

When a cohort and its eligible controls are observed over different windows and share no date pair, set cross_pair = TRUE to link their separate within-unit changes through calendar dates where the trend assumption applies. chain_report(fit) shows which comparisons identify each effect and the weights that combine them. The default gmm_weighting = "optimal" estimates the moment covariance from the data. "within" weights comparisons by their within-unit covariance under uncorrelated level shocks.

Parallel trends · Add PT-all restrictions · Cross-pair chaining, PT-ctrl and within-unit weighting · Short pre-treatment windows · GMM weighting and covariance regularization

3. Adjust for treatment

est_method makes each cohort comparable with its controls by IPW ("ipw"), OR ("reg") or DR ("dr", the default), as in did. In chained mode, OR and DR regress the controls' outcome changes on covariates. With the default outcome_pooling = "separate", each distinct comparison has its own coefficients. "pooled" shares slopes per calendar step across comparisons with local intercepts, the specification of versions before 1.2.0.0. Custom did-style IPW functions are accepted on the cohort route.

Treatment adjustment · Choose IPW, OR or DR in an incomplete panel

4. Adjust for missing outcomes, separately

selection states the assumption about missing outcomes. The default "none" fits no observation model and requires the observed units of each pair to be representative of their cohort's mean change. "conditional_mtar" allows observation to depend on covariates and adjusts for it. selection_est_method then chooses observation weighting ("ipw"), which weights each observed change by the inverse of its fitted probability of being observed at both dates, change imputation ("imputation"), which predicts the actual changes with regressions fitted to the observed changes, or the augmented correction ("dr"), which combines the two. All three work with treatment IPW, OR and DR, and selection_xformla can give the observation model its own covariates.

Treatment DR with the augmented correction is doubly robust in both layers: for the treatment, either the propensity model or the outcome regression may be misspecified, and for the missing outcomes, either the observation model or the imputation model, provided the other model of each layer and the remaining assumptions are correct. Augmenting the missing-outcome layer does not give treatment IPW or OR the treatment-DR guarantee.

Handling missing outcomes · Which models must be correct · Choose an observation model · Impute actual changes · Sampling information for the observation model · Robustness

5. Choose the treatment design

The examples above use an absorbing treatment: units adopt once and stay treated, and effects may evolve with time since adoption. Choose the switching model for entry and exit effects, or the linear fixed-effects specification for an adoption panel.

Design What to supply What you estimate Aggregate
Adoption cohorts (default) gname, the first treatment date, with 0 for units never treated ATT(g,t) for each cohort and date, so effects can evolve with time since adoption aggte(), including event studies
Binary treatment that can stop and restart dname, the treatment status at every study date, including dates with missing outcomes The immediate effect at each entry or exit, under a static model with no carryover of past treatment fit$aggregate, a switcher-weighted average
Linear cohort-by-date fixed effects gname and control = list(preset = "cohort_time_fe") Cohort-by-date effects in a linear specification with unit fixed effects and a common untreated calendar path aggte()

The dname model assumes that past treatment has no carryover beyond current status. Identification also depends on parallel trends and support for each entry or exit comparison. Entrants are compared with units that stay untreated and exiters with units that stay treated, and both estimates are reported on the same treated-minus-untreated scale. Treatment and missing-outcome adjustments are chosen as above. Supplying both gname and dname is an error, and absorbing data supplied through dname still requests the static model. The cohort-by-date preset is a linear specification whose regression fit matches R etwfe and Stata jwdid on the same inputs, with chained handling of gaps and the same treatment and missing-outcome adjustments.

When treatment can stop and restart · Cohort-by-date fixed effects · Options reference

Diagnostics and inference

  • Availability. tidy(fit) returns one row per requested cell with available, support.status and support.reason. Unavailable cells are NA with a reason. An aggregate that needs an unavailable cell is NA unless na.rm = TRUE requests the restricted target. See Diagnose errors and unavailable results.
  • Identification paths. chain_report(fit) lists the comparisons behind each effect and the weights used to combine them. See Cross-pair chaining.
  • Fitted models. Set control = list(keep_nuisance = TRUE) in att_gt(), then call nuisance_models(fit) to see the fitted equations, coefficient tables and standard errors of the treatment, outcome, observation and imputation models.
  • Pretrend tests. The fit records a joint Wald test of the cells before treatment when its covariance allows it. conditional_did_pretest() runs the conditional test. Failure to reject does not establish parallel trends. See Pretrend tests.
  • Uncertainty. bstrap, cband, biters and alp follow did: multiplier bootstrap standard errors and simultaneous bands by default, analytical standard errors with bstrap = FALSE. clustervars accepts one cluster variable beyond the unit. Standard errors use complete unit-level influence functions that include the fitted models. compute_inffunc = FALSE returns point estimates only. See Inference, Results and Uncertainty.

User guide

The guide home orders these pages as a reading path with stages.

Topic Pages
Start with your panel Get started, prepare your data, inspect missingness, hands-on notebook
Work with gaps Incomplete panels, changing sample composition, short pre-treatment windows
Choose comparisons Treatment adjustment, compare IPW, OR and DR, parallel trends, PT-all example, cross-pair chaining
Adjust for missing outcomes Overview, observation models, change imputation, sampling information, robustness
Choose the treatment design When treatment can stop and restart, cohort-by-date fixed effects
Read results Options, results, GMM weighting, inference, compare precision, troubleshooting
Compare software Familiar DiD models for incomplete panels, reproduce R and Stata models, moving from did

References

Chained IPW (est_method = "ipw") and observation weighting for missing outcomes follow Bellégo, Benatia and Dortet-Bernadet (2025). The package extends these methods with chained OR and DR, change imputation, the augmented correction, PT-all and PT-ctrl, cross-pair chaining, within-unit GMM weighting and the linear cohort-by-date preset. The static dname model adapts the static binary specialization of de Chaisemartin, D'Haultfoeuille, Pasquier, Sow and Vazquez-Bare to complete and incomplete panels. The methods pages of the guide state the assumptions each extension needs.

Roadmap

Further efficiency improvements are planned. Python and Stata versions are planned as well, with separate repositories and documentation.

Questions and bug reports

Use GitHub Issues to ask about using the package or choosing a method and to report reproducible bugs. Include your package version, sessionInfo() and a short reproducible example.

Contributing

See Contributing to propose a change or help with a reported issue.

Support the project

If cdid is useful to your work, you can buy me a coffee.

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.

install.packages("cdid")

1.3.0.1 by David Benatia, 10 hours ago


https://github.com/dvbn/cdid


Report a bug at https://github.com/dvbn/cdid/issues


Browse source code at https://github.com/cran/cdid


Authors: David Benatia [cre, aut] , Christophe Bellégo [aut] , Joel Cuerrier [aut] , Vincent Dortet-Bernadet [aut]


Documentation:   PDF Manual  


GPL-3 license


Imports did, DRDID, parallel, rlang

Suggests testthat, codetools, digest, ggplot2, knitr, rmarkdown, generics, jsonlite, waldo, withr


See at CRAN