TROP implementations compared: the paper, Python, Stata and R

The Triply RObust Panel (TROP) estimator of Athey, Imbens, Qu and Viviano (2026) is available in four places:

• The paper, which defines the estimator, its weights and its leave-one-out cross-validation.
• Python: the official trop package (zhaonanq/TROP-Estimator).
• Stata: the official trop command (justinwaddy/TROP).
• R: this package, cfcompare, an independent implementation that is not affiliated with the authors of the paper.

The implementations make different choices where the paper leaves room (how weights are anchored when many cells are treated, how penalties are cross-validated, how inference is done), and those choices change between releases. This page records them as dated snapshots: each table describes the four implementations as they stood at that date, and older snapshots are kept unchanged so that the evolution stays visible. Jump to the latest snapshot.

As of July 2026

Python trop v0.1.9, Stata trop v0.2.5, cfcompare 0.1.0.9006.

	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)

What changed between July and October 2026

Python trop

• v0.1.10 (July 2026): Monte Carlo simulation tools that generate panels from features of a real dataset, to compare estimators; the estimator objective rewritten with cvxpy.sum_squares.
• v0.1.11 (September 2026): unit-level bootstrap inference, trop.bootstrap (controls-only resampling by default, or stratified). The repository moved to zhaonanq/TROP-Estimator.

Stata trop

• v0.2.6 (August 2026): in the pooled group(time) mode, unit distances now mask the other treated units and the target periods; new pooled_treat_distance(unit() time()) option to override a group’s distance to itself.
• v0.2.7 (August 2026): adaptive cross-validation search, vce(jackknife), progress bars.
• v0.2.8 (August 2026): the adaptive search starts with a sweep of the full grids.
• v0.2.9 (September 2026): in group(time), the treated units are masked from the fit and the effect is estimated from the residual.

cfcompare

• 0.2.0 (October 2026): pooled anchoring per treatment spell (Stata’s group(time) estimands, with per-spell effects in fit$tau_groups); pooled time weights centred on the spell midpoint; pooled_treat_distance; controls-only bootstrap resampling; delete-one-unit jackknife over every unit. See the changelog for details.
• 0.2.1 (October 2026): this website; the comparison tables moved here from the README.

As of October 2026

Python trop v0.1.11, Stata trop v0.2.9, cfcompare 0.2.1.

	TROP Paper	Python (v0.1.11)	Stata (v0.2.9)	R (cfcompare 0.2.1)
Solver	Silent	CVXPY (SCS for finite lambda_nn, OSQP for lambda_nn = Inf), sum_squares formulation	FISTA (accelerated proximal gradient with adaptive restart), warm-start nuclear path	Soft-impute-style proximal gradient (base R only; optional truncated SVD via RSpectra), warm-start nuclear path
Estimands / weight anchoring	Per-cell	Pooled only, over a single trailing block (treated_periods)	group(time) (default): one pooled estimand per treatment spell (same start and end, pooled across units), cell-weighted average; group(cell): per-cell	Auto (per-cell when at most max_cells treated cells, else pooled) by default; anchor = "per_cell" or "pooled" (per treatment spell, as Stata’s group(time); per-spell effects in fit$tau_groups)
Pooled unit distance	N/A (per-cell eq. (3))	RMS distance to the average treated path over the pre-periods	RMS distance to the group’s average path over the periods in which no group member is treated; each comparison unit’s own treated cells masked (v0.2.6)	Same as Stata; identical to Python on a block design
Pooled time-weight centre	N/A (per-cell abs(t - s))	T - tp/2 (half a period after the block midpoint)	Spell midpoint (first + last)/2 (v0.2.4)	Spell midpoint by default; trop_control(time_center = "python") reproduces the Python convention, which trop_matrix() always uses
Own-group distance	N/A	Treated units compared with their own average (non-zero in general)	Same; pooled_treat_distance(unit(#) time(#)) overrides the members’ unit distance and the in-spell time distance (v0.2.6)	Same; trop_control(pooled_treat_distance = list(unit =, time =))
Units with no usable comparison period	Undefined (0/0)	Not applicable (block design: every unit shares the pre-periods)	Unit weight set to 0 when lambda_unit > 0 (included when lambda_unit = 0)	Largest finite distance (smallest positive weight)
Other treated cells in the fit	Excluded	Excluded (block design)	Excluded; tau is the mean residual over the target cells (v0.2.9)	Excluded; tau is the mean residual over the target cells
CV criterion	LOOCV	Placebo resampling (default), k-fold; cycle, joint or adaptive search	Placebo resampling (default) or k-fold for group(time); LOOCV for group(cell); cycle, joint or adaptive search	LOOCV (default) or placebo resampling (cv_method = "placebo"); coordinate-descent (cycle) search
Placebo pattern for staggered / general treatment	N/A (LOOCV)	Common trailing block	Actual per-unit patterns, grouped by treatment spell	Actual per-unit patterns, grouped by treatment spell; warning under LOOCV + staggered
Outcome pre-processing	Raw Y	Raw Y	Standardized (always); lambdas on the standardized scale	Raw Y (default) and standardized (standardize = TRUE)
Covariates	Yes	No	No	Yes
Bootstrap	Stratified block bootstrap	trop.bootstrap: unit-level bootstrap, controls-only resampling (default) or stratified; percentile CI	vce(bootstrap): stratified block bootstrap; percentile CI (no inference by default)	se = "bootstrap" (default): stratified (default) or controls-only (boot_resample = "controls"); percentile or normal CI
Jackknife	–	–	vce(jackknife): delete-one-unit over every unit	se = "jackknife": delete-one-unit over every unit (replicates that leave no treated unit are skipped)

Numerical agreement with the official Python package

The TROP engine in cfcompare 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 matrix-in entry point trop_matrix() keeps the Python package’s signature and conventions for such cross-checks.

Notes on cfcompare’s implementation choices

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; exact digits therefore need not match convex-solver implementations. To reconcile the difference, fix the penalties via lambda = list(time =, unit =, nn =) (bypassing cross-validation), set trop_control(svd = "full"), and match the other implementation’s anchoring.

Cross-validation. Penalties are chosen by leave-one-control-cell-out prediction error by default (the paper’s criterion); a placebo criterion that resamples the actual treatment patterns is available via trop_control(cv_method = "placebo"), and is the recommended choice under staggered adoption.

Per-cell or pooled anchoring. anchor = "per_cell" solves a separate locally weighted problem for each treated cell. anchor = "pooled" uses one set of weights per treatment spell group – treated cells with the same start and end period, pooled across units, the group(time) estimands of the Stata command – and is faster. Under block adoption there is one group; under staggered adoption one per cohort; a unit that switches in and out of treatment contributes one group per spell. A group’s unit distance is to the average path of its units over the periods in which none of them is treated, each comparison unit’s own treated cells are masked, and the time weights are centred on the spell midpoint. The per-group effects are returned in fit$tau_groups, and the ATT is their cell-weighted average. trop_control(pooled_treat_distance = list(unit = 0, time = 0)) gives the group’s own units and periods the full weight (the Stata command’s pooled_treat_distance()). trop_matrix() is always pooled over a single trailing block (the Python signature).

Covariates. Time-varying covariates can be supplied to trop() and panel_compare(); they enter the low-rank term additively, as in Section 6.2 of the paper.

Nuclear-norm floor. For numerical stability, the solver adds a small scale-relative floor to every finite lambda_nn by default (the reported lambda_nn does not include it). lambda_nn = Inf (the DID/TWFE-like case) is never affected, and trop_control(nn_floor = 0) applies the penalties exactly as supplied.

Standardization. Estimation uses the raw outcome by default, so lambda values are on the outcome’s natural scale; trop(standardize = TRUE) standardizes the outcome as the Stata command does, and then reports lambda on the standardized scale.

Inference. se = "bootstrap" resamples units with replacement, keeping each unit’s full time series: treated and control units separately by default (the stratified block bootstrap of the paper and of the Stata command), or only the control units with trop_control(boot_resample = "controls") (the default of the Python package’s trop.bootstrap). se = "jackknife" deletes one unit at a time over every unit, as the Stata command’s vce(jackknife); replicates that leave no treated unit are skipped.