Changelog • fasttrop

fasttrop (development version)

Time weights now use direct exponential values, with zero after underflow. This corrects the identifiability test for distant cells in long panels. The correction applies to local estimation, cross-validation, and bootstrap draws.

trop() has the new argument covariates, and trop_matrix() has the new argument X. They add the term (X_{it}’) of Eq. 13 in Section 6.2 of the paper.
The new argument covariate_method selects how to calculate (). With "joint" (the default), each weighted fit calculates () together with the other parts of the model. With "residualize", one TWFE regression on the untreated cells calculates () first.
The global method can also use covariates. This is an extension of fasttrop. Section 6.2 of the paper adds covariates to the local method. The Python package from the authors does not have covariates.
The result has the new element covariates with the coefficients, the method, and the names of the covariates that the fit used.
The fixed effects absorb a covariate that does not change in time or does not change between units. The function gives a warning message for such a covariate.

New tests compare the results with covariates with two references. The script data-raw/oracle_covariates.py adds the covariate term to the code of the authors and also solves the problem with a second method.
The new script inst/validation/covariates.R runs a simulation with a known effect and a covariate that is different for the treated units.
The new scripts in inst/validation/stata compare fasttrop with the Stata package trop. They only call that package.
The new script inst/bench/covariate_timings.R measures the time cost of covariates.
The new article “Use covariates” explains the two procedures and the results.

This is the first version of the package.

The package calculates the Triply Robust Panel (TROP) estimator of Athey, Imbens, Qu, and Viviano (2026).
The local method (method = "local") is the estimator of the paper for each treated cell. The global method (method = "global") is the block estimator of the Python package from the authors.
Leave-one-out cross-validation or placebo cross-validation selects the tuning parameters.
The unit bootstrap of Algorithm 3 gives standard errors. The default confidence interval is a normal interval. A percentile interval is also available.

The model fits run in C with OpenMP and with the BLAS and LAPACK libraries of R. The number of threads does not change the results.
On Windows, the C code tells Windows not to use power throttling on its threads while it runs. Power throttling can make a long calculation in an R session slower. To keep the usual Windows policy, use trop_control(throttle = TRUE).
The function trop_threads() gives at most 2 threads when R CMD check sets the environment variable _R_CHECK_LIMIT_CORES_.

The package has a website with articles in plain words: https://farach.github.io/fasttrop/.
The vignette “Get started with fasttrop” shows a first analysis step by step.
citation("fasttrop") gives the references for the paper and the package.

The documents use the name “TWFE imputation” for the special case lambda = c(time = 0, unit = 0, nn = Inf). This method fits a two-way fixed effects model to the untreated cells only. In a staggered design, it is different from a conventional TWFE regression.
Two new tests compare TWFE imputation with a conventional TWFE regression. The test suite has 478 expectations.
The new script inst/validation/staggered.R runs simulations with a staggered design and a known effect. It compares a conventional TWFE regression, TWFE imputation, and the local method of TROP. The folder inst/validation has the saved results.
The article “How we checked fasttrop” has the results of the new simulations. The article “What TROP can and cannot do” has the limits.