Skip to contents

Triply RObust Panel (TROP) estimator

Source: R/trop.R
trop.Rd

Fits the TROP estimator of Athey, Imbens, Qu & Viviano (2026) on a long panel. TROP combines a low-rank-plus-two-way-fixed-effects outcome model with exponential-decay unit weights (upweighting controls similar to the treated) and time weights (upweighting periods near the treated periods). Penalties are chosen by leave-one-out cross-validation on the control cells. The estimator nests DID/TWFE, matrix completion and synthetic-control-type weighting as special cases.

Usage

trop(
  data,
  outcome,
  treatment,
  unit,
  time,
  covariates = NULL,
  lambda = NULL,
  standardize = FALSE,
  anchor = c("auto", "per_cell", "pooled"),
  se = c("bootstrap", "auto", "jackknife", "none"),
  grids = NULL,
  control = trop_control(),
  verbose = FALSE
)

Arguments

data

A long data.frame with one row per unit-time.

outcome, treatment, unit, time

Column names (strings). treatment must be a 0/1 indicator of active treatment in that cell (works for block, staggered, and non-absorbing designs).

covariates

Optional character vector of column names for time-varying covariates. When supplied, the low-rank term is augmented additively as \(L_{js} = X_{js}'\phi + R_{js}\) (Athey, Imbens, Qu & Viviano 2025, Section 6.2): the covariate-linear part is unpenalised and the nuclear norm applies to the residual \(R\). Penalty cross-validation is performed on the covariate-residualised model. Covariates must be fully observed (including in treated cells, where they enter the counterfactual). The fitted coefficients are returned in $phi.

lambda

Optional named list list(time=, unit=, nn=) to fix the penalties and skip cross-validation. When standardize = TRUE, supplied penalties are interpreted on the standardized-outcome scale. Whether nn is supplied here or chosen by cross-validation, the solver adds the small scale-relative stabilising floor nn_floor (see trop_control()) to every finite nn (the reported $lambda$nn stays nominal and the resolved floor is stored in $lambda$nn_floor); set control = trop_control(nn_floor = 0) to apply nn exactly as given.

standardize

Logical; if TRUE, the outcome is standardized internally as (Y - mean(Y)) / sd(Y) (over all observed cells) before fitting, and the ATT, standard error, confidence interval, per-cell effects and counterfactual are mapped back to the raw outcome scale. This matches the convention of the official Stata command (v0.2.1+); as there, the selected lambda values are reported on the standardized scale, which makes them comparable across outcomes and across implementations. Default FALSE (raw outcome, the previous behaviour). The returned panel$Y is on the fitting (standardized) scale, with the centre/scale stored in $scaling.

anchor

How weights are anchored to treated cells: "per_cell" re-solves eq. (2) with cell-specific weights for every treated cell (faithful to the paper); "pooled" solves once per treatment spell group – treated cells with the same start and end period, pooled across units, the group(time) estimands of the official Stata command – with weights anchored to that group (unit distances to the average trajectory of the group's units over the periods in which none of them is treated, time distances to the midpoint of the spell; see trop_control()'s time_center). Simultaneous permanent adoption is one group, staggered adoption gives one group per cohort, and a unit that switches in and out of treatment contributes one group per spell; the ATT is the treated-cell-weighted average of the group effects, which are returned in $tau_groups. "auto" (default) uses per_cell when there are at most max_cells treated cells and pooled otherwise. trop_matrix() has no anchor argument; it is always pooled over a single trailing block.

se

Standard-error method: "bootstrap" (default; unit-level block bootstrap, stratified by treatment status unless trop_control(boot_resample = "controls")), "jackknife" (delete-one-unit over every unit, treated and control, as the Stata command's vce(jackknife); replicates that leave no treated unit are skipped), "auto" (jackknife when there are at least 2 treated units, else bootstrap), or "none".

grids

Optional list of penalty grids; see trop_control().

control

A list of solver/CV settings from trop_control().

verbose

Logical; print CV progress.

Value

An object of class trop: a list with the ATT estimate, std.error, conf.low/conf.high, the selected penalties lambda, any covariate coefficients phi, per-cell effects (tau_cells), the per-spell-group effects tau_groups (pooled anchor only: one row per group with its start/end period, unit and cell counts and estimate; the ATT is their n_cells-weighted average), the estimated counterfactual matrix, the reshaped panel, and scaling (the outcome centre/scale used internally; the identity unless standardize = TRUE). All effect and outcome-level components are on the raw outcome scale; panel$Y and lambda are on the fitting scale.

References

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

Examples

df <- sim_panel(N = 14, T = 9, n_treated = 3, t0 = 6, seed = 1)
# fixed penalties keep this example fast and deterministic; omit `lambda`
# to choose (time, unit, nn) by cross-validation
fit <- trop(df, "y", "w", "id", "t",
            lambda = list(time = 0.1, unit = 0.5, nn = 2), se = "none")
fit
#> Triply RObust Panel (TROP) estimator
#>   ATT = 0.5904
#>   penalties: lambda_time=0.1, lambda_unit=0.5, lambda_nn=2
#>   lambda_nn stabilising floor: +0.00466 (added in the solver; trop_control(nn_floor = 0) disables)
#>   estimated rank(L)=2, anchor=per_cell, design=block
#>   treated cells=12 across 3 unit(s)