Permute treatment in space across every donor
Source:R/scm_placebo_permutation.R
scm_placebo_permutation.RdRefits fit_scm() once per unit, each time pretending that unit is the
treated one and that the rest of the pool are its donors. The result is
the reference distribution rmspe_pvalue() ranks against, and the set of
grey placebo paths behind an Abadie-style gap figure.
Arguments
- Y
Outcome matrix, units x periods, controls first and the treated unit last.
- N0
Number of control units.
- T0
Number of pre-treatment periods.
- X
Optional units x predictors matrix, rows aligned with
Y.- ...
Passed to
fit_scm()–outcome_periods,bias_correct,v_method,bc_donorsand friends. The same arguments are used for the treated fit and every placebo, which is the point.
Value
A list with results, a tibble of one row per unit (unit,
treated, att, pre_rmspe, post_rmspe, rmspe_ratio and their
_raw counterparts computed before bias correction), gaps, a
units x periods matrix of the corrected gap paths, and gaps_raw, the
same before correction.
Details
The treated unit is always excluded from the donor pool of a placebo fit. If it were left in, a placebo could reproduce itself out of the genuinely treated unit's post-treatment trajectory, which would import the effect being tested into the null distribution.
Everything is deterministic: the simplex solve is Frank-Wolfe from a
fixed start and the bias correction is a regression, so no seed is
involved and repeated calls agree exactly. That is the difference between
this and the placebo variance in fit_sdid(), which is simulated.
See also
rmspe_pvalue() for the ranking that turns this into a
p-value, and fit_scm() for the fit it repeats.
Examples
set.seed(1)
Y <- matrix(rnorm(6 * 8), nrow = 6,
dimnames = list(c(paste0("d", 1:5), "treated"), 2010:2017))
Y[6, 7:8] <- Y[6, 7:8] + 3
scm_placebo_permutation(Y, N0 = 5, T0 = 6)$results
#> Bias correction fitted on all 4 donors, not the 2 with positive weight.
#> ℹ 6 predictors need more donors than that to be identified.
#> Bias correction fitted on all 4 donors, not the 3 with positive weight.
#> ℹ 6 predictors need more donors than that to be identified.
#> Bias correction fitted on all 4 donors, not the 3 with positive weight.
#> ℹ 6 predictors need more donors than that to be identified.
#> Bias correction fitted on all 4 donors, not the 3 with positive weight.
#> ℹ 6 predictors need more donors than that to be identified.
#> Bias correction fitted on all 4 donors, not the 3 with positive weight.
#> ℹ 6 predictors need more donors than that to be identified.
#> Bias correction fitted on all 5 donors, not the 3 with positive weight.
#> ℹ 6 predictors need more donors than that to be identified.
#> # A tibble: 6 × 12
#> unit treated att pre_rmspe post_rmspe rmspe_ratio pre_rmspe_raw
#> <chr> <lgl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 d1 FALSE -0.162 0.342 0.950 2.78 0.595
#> 2 d2 FALSE 0.190 0.437 0.774 1.77 0.783
#> 3 d3 FALSE 0.185 0.371 1.16 3.12 0.728
#> 4 d4 FALSE -0.365 0.942 1.02 1.08 1.05
#> 5 d5 FALSE -0.297 0.566 1.52 2.69 0.916
#> 6 treated TRUE 3.23 1.06 3.23 3.04 1.15
#> # ℹ 5 more variables: post_rmspe_raw <dbl>, rmspe_ratio_raw <dbl>,
#> # rmspe_ratio_hybrid <dbl>, exact_fit <lgl>, exact_match <lgl>