The efficient robust estimator of Borusyak, Jaravel and Spiess (2024), Theorem 2. Unit and period effects are fitted on untreated observations only, untreated potential outcomes are imputed for the treated cells, and the treated residuals are averaged with the requested weights. Unlike a two-way fixed-effects regression, no already-treated observation contributes to the counterfactual.
Arguments
- y
Outcome, long format.
- D
Binary 0/1 absorbing treatment, long format.
- unit, time
Panel identifiers.
- X
Optional time-varying covariates (vector or matrix, one column per covariate), long format.
- weights
Target weights:
NULLor"equal"(default) averages treated cells equally;"cohort"weights by cohort size; or supply your own per-observation weights. All three routes are available, since the paper's Theorem 2 imputes identically whatever the estimand.
Value
A list with estimate, se, ci,
tau_it, imputed_y0, weights,
linearity_residual, se_note, event,
cohort_att, pretrend_by_rel,
pretrend_max_abs, pretrend_note,
unit_effects, period_effects,
covariate_coef, n_treated_cells,
n_untreated_cells, n_units, n_periods,
no_forbidden_comparisons, method.
References
Borusyak, K., Jaravel, X. and Spiess, J. (2024). Revisiting event study designs: robust and efficient estimation. Review of Economic Studies, 91(6), 3253-3285.
Examples
set.seed(1)
nu <- 6; T <- 5
unit <- rep(1:nu, each = T)
time <- rep(1:T, nu)
ft <- rep(c(Inf, Inf, Inf, 4, 4, 3), each = T)
D <- as.integer(time >= ft)
y <- rnorm(nu * T) + D * 1.5
morie_boryis(y, D, unit, time)
#> $estimate
#> [1] 1.012478
#>
#> $se
#> [1] 0.6437863
#>
#> $ci
#> [1] -0.2493203 2.2742757
#>
#> $tau_it
#> [,1] [,2] [,3] [,4] [,5]
#> [1,] -0.7066393 -0.02792423 -0.7649001 1.5813997 -0.08193605
#> [2,] -0.9065197 0.26999573 0.8031875 0.5560344 -0.72269798
#> [3,] 1.5227426 0.26942261 -0.4593651 -2.1374341 0.80463402
#> [4,] -0.3456817 -0.44832042 0.7940022 2.0867774 1.46189490
#> [5,] 0.3205738 0.05235070 -0.3729245 -1.0214509 1.19016388
#> [6,] 0.1155244 -0.11552440 0.3518148 1.2598075 1.75833640
#>
#> $imputed_y0
#> [,1] [,2] [,3] [,4] [,5]
#> [1,] 0.08018553 0.21156755 -0.07072856 0.01388114 0.4114438
#> [2,] 0.08605130 0.21743332 -0.06486279 0.01974691 0.4173096
#> [3,] -0.01096140 0.12042063 -0.16187548 -0.07726578 0.3202969
#> [4,] 0.30074813 0.43213016 0.14983405 0.23444375 0.6320064
#> [5,] 0.59840357 0.72978560 0.44748949 0.53209919 0.9296619
#> [6,] -0.17165314 -0.04027111 -0.32256722 -0.23795752 0.1596052
#>
#> $weights
#> [,1] [,2] [,3] [,4] [,5]
#> [1,] 0.11904762 0.11904762 0.04761905 -0.1428571 -0.1428571
#> [2,] 0.11904762 0.11904762 0.04761905 -0.1428571 -0.1428571
#> [3,] 0.11904762 0.11904762 0.04761905 -0.1428571 -0.1428571
#> [4,] -0.07142857 -0.07142857 -0.14285714 0.1428571 0.1428571
#> [5,] -0.07142857 -0.07142857 -0.14285714 0.1428571 0.1428571
#> [6,] -0.21428571 -0.21428571 0.14285714 0.1428571 0.1428571
#>
#> $linearity_residual
#> [1] 1.088019e-14
#>
#> $se_note
#> [1] "clustered on unit using the estimator's exact linear weights; treated cells contribute their deviation from the average effect, which is the conservative choice"
#>
#> $event
#> $event$`0`
#> [1] 0.4723805
#>
#> $event$`1`
#> [1] 1.303955
#>
#> $event$`2`
#> [1] 1.758336
#>
#>
#> $cohort_att
#> $cohort_att$`2`
#> [1] 1.12332
#>
#> $cohort_att$`3`
#> [1] 0.9293463
#>
#>
#> $pretrend_by_rel
#> $pretrend_by_rel$`-3`
#> [1] -0.01255397
#>
#> $pretrend_by_rel$`-2`
#> [1] -0.09348177
#>
#> $pretrend_by_rel$`-1`
#> [1] 0.1018511
#>
#>
#> $pretrend_max_abs
#> [1] 0.1018511
#>
#> $pretrend_note
#> [1] "the imputation uses EVERY untreated period, so parallel trends is assumed throughout the pre-period, not only just before adoption; these residuals are zero in expectation under that assumption"
#>
#> $unit_effects
#> [1] 0.11621089 0.12207667 0.02506397 0.33677350 0.63442894 -0.13562777
#>
#> $period_effects
#> [1] -0.03602537 0.09535666 -0.18693945 -0.10232975 0.29523293
#>
#> $covariate_coef
#> NULL
#>
#> $n_treated_cells
#> [1] 7
#>
#> $n_untreated_cells
#> [1] 23
#>
#> $n_units
#> [1] 6
#>
#> $n_periods
#> [1] 5
#>
#> $no_forbidden_comparisons
#> [1] "no treated observation enters the fit, so an already-treated unit cannot act as a control"
#>
#> $method
#> [1] "Borusyak-Jaravel-Spiess (2024) imputation estimator"
#>