cfcompare: R package implementing TROP and comparing it with other ATT estimators

Note: cfcompare is an independent R package. This is not officially endorsed or maintained by the authors of the TROP article (Athey, Imbens, Qu & Viviano, 2026). Official TROP software by the authors includes the Python package trop (ostasovskyi/TROP-Estimator) and the Stata command (justinwaddy/TROP). Please cite the original TROP paper when using the TROP estimator. See Citation for more details.

cfcompare implements the triply robust panel estimator (TROP) and puts it in a common R workflow with DID/TWFE, synthetic control, matrix completion, synthetic DID, and DIFP. This package enables not just to estimate TROP, but also to compare various ATT estimators on the same schema.

Use cfcompare when you want to:

run several ATT estimators on one binary-treatment panel with panel_compare();
get a shared tidy ATT table and shared plotting methods;
evaluate estimators by held-out RMSE with panel_rmse();
inspect TROP tuning through lambda grids with trop_sensitivity and plot_trop_surfaces (for CV-loss and ATT surfaces).

Note: all plots and tables shown in this README are based on a model sample dataset, and they are provided solely as examples of the visualizations that can be created with this package.

Installation

# install.packages("pak")
pak::pak("takuma1102/cfcompare")

You can also install from R-universe. This package will be submitted to CRAN as well down the line.

install.packages(
  "cfcompare",
  repos = c("https://takuma1102.r-universe.dev", "https://cloud.r-project.org")
)

Most of the core native engines run without optional estimator packages. Install synthdid, gsynth, augsynth, or did for the wrapped methods you plan to use.

Quick start

library(cfcompare)

# Compare multiple estimators.
cmp <- panel_compare(
  df,
  outcome = "y", treatment = "w", unit = "id", time = "t",
  methods = c("DID", "SDID", "SC", "MC", "DIFP", "TROP"),
  se = "bootstrap",  # You can also choose jackknife or "none" for running this faster.
  # covariates = c("x1", "x2"),
  anchor  = "pooled",  # faster method
  # or "per-cell" (more accurate; in accordance with the original paper)
  # or "auto" (choosing "per-cell" or "pooled" depending on the number of treated cells)
)

cmp$att                    # tidy ATT table, one row per method
autoplot(cmp)              # forest plot of ATT estimates and intervals (see below for an example plot)

Supported estimators

Method	Engine	`panel_compare()` default?	Notes
`TROP`	native R	yes	low-rank + two-way FE outcome model with unit/time weights
`DID`	native R	yes	two-way fixed effects
`SC`	`synthdid`	yes	skipped if `synthdid` is unavailable or the design is unsupported
`MC`	native R	yes	nuclear-norm matrix completion
`SDID`	`synthdid`	yes	skipped if `synthdid` is unavailable or the design is unsupported
`DIFP`	native R	yes	Doudchenko-Imbens-Ferman-Pinto estimator: demeaned SC with intercept
`gsynth`	`gsynth`	no	optional interactive-fixed-effects / MC-style engine
`augsynth`	`augsynth`	no	optional augmented synthetic control engine
`CS`	`did`	no	Callaway--Sant'Anna; requires an absorbing staggered treatment

By default, panel_compare() runs TROP, DID, SC, MC, SDID, and DIFP. Use methods = to specify an explicit set, or exclude = to drop one method from the default set. Optional engines whose package is missing, or whose design requirements are not met (e.g., in terms of covariates), are skipped with a message while the remaining methods still run.

Overview: what cfcompare can do

cfcompare has basic functions for estimating the TROP estimator (the official Python and Stata projects are referenced). It also adds an R-native comparison and diagnostic layer around TROP:

Goal	Entry point	Output
Compare multiple ATT estimators	`panel_compare()`	`cf_comparison` with a common `cf_att_tbl`, fits, panel data, and counterfactuals
Compare held-out predictive performance	`panel_rmse()`	ranked RMSE table; this is separate from TROP tuning cross-validation
Track estimation RMSE vs design size	`rmse_curve()` / `rmse_curves()`	`sqrt(E[(tau_hat - tau)^2])` vs N, T, #treated, or #post, with a known true ATT
Inspect TROP penalty sensitivity	`trop_sensitivity()` + `autoplot()`	lambda grid with CV loss and ATT at each grid point
Plot CV-loss / ATT surfaces	`plot_trop_surfaces()`	separate full-width CV-loss and ATT surface plots; returns surface matrices invisibly
Audit penalty components	`trop_ablation()`	table moving from full TROP toward MC and DID by constraining penalties
Compare inference choices	`compare_se_modes()`	one ATT estimate with bootstrap, jackknife, and/or placebo uncertainty rows
Reuse fitted counterfactuals	`counterfactual_matrix()`	common `N x T` estimated untreated-outcome matrix interface

Held-out RMSE comparison

panel_rmse() compares estimators by how well they predict held-out outcomes. The default metric = "placebo" repeatedly assigns a placebo block to control units and scores the resulting zero-effect placebo ATT.

r <- panel_rmse(
  df, outcome = "y", treatment = "w", unit = "id", time = "t",
  methods = c("DID", "SDID", "SC", "MC", "DIFP", "TROP"),
  horizon = 4, n_pseudo = 6, n_runs = 100, seed = 1
)

r            # ranked table: method, rmse, rmse_se, engine, note
autoplot(r)  # see below for an example plot

metric = "prediction" scores per-cell counterfactual prediction error.

For quick diagnostics on large panels, reduce the outer repetition counts, such as n_runs, and use lighter TROP controls via trop_control(n_cv_cells = , cv_cycles = , max_iter = , workers = ).

Note: This is a predictive error on held-out cells, not estimation error against a known effect. It is a different quantity from the estimation RMSE sqrt(E[(tau_hat - tau)^2]) reported by rmse_curves() and rmse_curve() (for a single-dimension plot) over Monte Carlo replications with a known true ATT. These two can rank methods differently.

TROP diagnostics

Run a single TROP fit when you need the selected penalties, per-treated-cell effects, or the fitted untreated counterfactual matrix.

fit <- trop(df, "y", "w", "id", "t")
fit$lambda                  # CV-selected (time, unit, nn) penalties
fit$tau_cells               # per-treated-cell effects
counterfactual_matrix(fit)  # fitted N x T untreated-outcome matrix
autoplot(fit)               # synthetic-control-style trajectory; see below

Inspect the TROP penalty surface by sweeping any two of the three penalties (time, unit, nn) and holding the third fixed.

g <- trop_sensitivity(
  df, "y", "w", "id", "t",
  axes = c("unit", "time"), # You can also choose "nn" for lambda_nn
  anchor  = "pooled"  # 
)

autoplot(g)                    # compact ggplot2 heatmap (see below)

surfaces <- plot_trop_surfaces(g, which = "both", ask = FALSE) (see below)
surfaces$cv_loss               # matrix behind the CV-loss surface
surfaces$att                   # matrix behind the ATT surface

Other targeted diagnostics are available when needed.

Use trop_ablation when you want to check how RMSEs change in accordance with the change in lambda penalties. (See also Table 5 in the original paper.)

trop_ablation(df, "y", "w", "id", "t")  # see below

Use compare_se_modes when you want to compare bootstrap and jackknife standard errors.

compare_se_modes(df, "y", "w", "id", "t", se = c("bootstrap", "jackknife"))  # see below

For event-study plots, use trop_event_study.

trop_event_study(fit, se = "bootstrap")

For placebo test relabeling control units as pseudo-treated, use trop_placebo_test() (penalties are held fixed).

pt <- trop_placebo_test(
  fit, B = 500L, alternative = "two.sided"
)

autoplot(pt)    # histogram of the null; observed ATT (red solid line) and 0 (dashed); see below.

Designs and inference

The treatment column is the active 0/1 treatment indicator for each unit-time cell. The TROP diagnosis handles block, staggered, and non-absorbing treatment patterns under the same schema. By contrast, SDID and SC are block-design methods and are skipped otherwise. CS is opt-in and requires absorbing staggered treatment.

Native estimators support se = "bootstrap", "jackknife", and "none" where applicable. These are practical resampling choices for the R workflow; use the TROP paper for formal inference conditions.

Numerical agreement with, and technical differences from the official packages

The TROP engine in this package is written in base R. For correctness checks, it has been compared with the official Python trop package on exactly comparable weighted-TWFE sample datasets, using trop_control(svd = "full"), and the results agree to numerical tolerance.

The below is how the original TROP paper, the official Python/Stata packages, and cfcompare handle main implementation choices.

	TROP Paper	Python	Stata	R (cfcompare)
Solver	Silent	CVXPY using SCS or OSQP	FISTA (accelerated proximal gradient)	Soft-impute-style proximal gradient (base R only)
Weight anchoring	Per-cell	Pooled only	Pooled (default) and per-cell (option)	Auto (choosing per-cell or pooled by treated-cell count) by default; per-cell and pooled as options
CV criterion	LOOCV	Placebo resampling (default) and k-fold (option)	Placebo resampling (default) and k-fold (option) for pooled; LOOCV for per-cell	LOOCV (default) and placebo resampling (option)
Placebo pattern for staggered treatment	N/A (LOOCV)	Common block	Actual per-unit patterns	Actual per-unit patterns; warning under LOOCV + staggered
Outcome pre-processing	Raw Y	Raw Y	Standardized (default) and raw Y (option)	Raw Y (default) and standardized (option)
Covariates	Yes	No	No	Yes
Inference	Bootstrap	None	Bootstrap	Bootstrap or jackknife (when available)

See the notes below for technical differences for this package's estimation from official packages.

Note 1 (solver): finite nuclear-norm penalties are solved with a proximal-gradient soft-impute routine, as opposed to FISTA or SCS, so that the implementation depends only on base R. Thus, exact digits need not perfectly match convex-solver implementations. To reconcile this difference, you can fix the penalties via lambda = list(time=, unit=, nn=) (bypassing CV), set svd = "full", and match its anchor.

Note 2 (cv): penalties are chosen by leave-one-control-cell-out prediction error by default; a placebo criterion is also available via trop_control(cv_method = "placebo").

Note 3 (per-cell or pooled): this package supports both anchoring modes of per-cell (solving a separate local weighting problem for each treated cell by using trop_matrix()) and pooled methods (constructing one set of weights for the entire treated group; faster than per-cell).

Note 4 (covariates): this package allows to use covariates through panel_compare.

Note 5 (lambda_nn's floor): for numerical stability, the solver adds a small scale-relative stabilising floor to every finite lambda_nn by default. (Reported lambda_nn as part of results do not include the samll floor.) lambda_nn = Inf (i.e., the DID/TWFE-like case) is never affected. trop_control(nn_floor = 0) disables this adjustment so penalties are applied exactly as supplied.

Note 6 (standardization): estimation uses the raw, non-standardized outcome by default, so that lambda values are on the outcome's natural scale. trop(standardize = TRUE) optionally standardizes the outcome.

More details

For more details, use the package help topics:

?panel_compare
?panel_rmse
?trop
?trop_sensitivity
?trop_ablation
?compare_se_modes

The source vignette is now being developed and the current version is stored in vignettes/cfcompare.Rmd; this will later be surfaced through pkgdown after being finalized.

Status

Experimental. The output schema is intended to be stable, but wrapped estimator packages can change their internal objects. Pin package versions in production code.

Citation

cfcompare is an independent package to implement TROP. If you use the TROP estimator, please cite the original JAM paper:

Athey, S., Imbens, G. W., Qu, Z., & Viviano, D. (2026). Triply Robust Panel Estimators. Journal of Applied Econometrics, 1--16. doi:10.1002/jae.70061.

Also cite the underlying packages for any wrapped estimators you use (synthdid, gsynth, augsynth, and did).