The Abadie, Diamond and Hainmueller synthetic control, matched on an
arbitrary predictor matrix rather than on pre-period outcomes alone, with
the Abadie-L'Hour / Ben-Michael-Feller-Rothstein bias correction applied
on top. This is the estimator Orlova, Rury and Wiltshire (2023) use as
their preferred specification, via the Stata package allsynth
(Wiltshire 2022); fit_sdid() covers the synthetic-DiD and plain-DiD
comparisons they report alongside it.
Arguments
- Y
Outcome matrix, units x periods, controls first and the treated unit last – the layout
synthdid::panel.matrices()returns.- N0
Number of control units (the first
N0rows ofY).- T0
Number of pre-treatment periods (the first
T0columns).- X
Optional units x predictors matrix of time-invariant predictors, rows in the same order as
Y. Columns that are constant across units are dropped with a warning: they cannot discriminate between donors.- outcome_periods
Integer indices of pre-treatment columns of
Yto include as predictors, on top ofX.NULL(the default) uses every pre-treatment period. Pass a subset to reproduce a specification that matches on selected pre-treatment years.- bias_correct
Apply the bias correction.
TRUEby default.- bc_ridge
Ridge penalty for the bias-correction regressions, on standardised predictors. Small and positive; only binds when the positive-weight donors are few.
- bc_donors
Which donors the bias-correction regressions are fitted on.
"positive"uses only donors carrying positive weight, asallsynthdoes;"all"uses the whole pool, as augmented SCM does;"auto"(the default) uses the positive-weight donors when there are comfortably more of them than predictors and falls back to the whole pool otherwise, reporting when it does. The fallback matters: a sparse synthetic control can easily name fewer donors than there are predictors, and the regression is then rank-deficient exactly when the correction is being asked to do the most work.- v_method
How the diagonal of
Vis chosen."equal"(default) weights standardised predictors equally."mspe"searches for theVminimising pre-period MSPE, fromv_startsdeterministic starting points."synth"does the"mspe"search AND asks theSynthpackage for its own solution, then keeps whichever fits the pre-period better – see the note below on why that is not redundant."synth"needs the suggestedSynthpackage and falls back to"mspe"with a warning if it is missing.- v_starts
Restarts for the
"mspe"search, the first from equal weights and the rest scattered. One start is not enough – see the note in the code. Ignored whenv_method = "equal".- v_maxit
Nelder-Mead iterations per restart.
- penalty
Abadie-L'Hour (2021) penalty
lambda, non-negative. Zero (the default) is the classic synthetic control. A positive value adds a charge for each donor's own distance from the treated unit, which makes the weights unique and favours donors that resemble the treated unit individually rather than only on average. Its main use here is breaking ties when the pre-period is fitted exactly: any small positive value then picks, among the exact fits, the one built from the nearest donors. Larger values trade pre-period fit for locality, and aspenaltygrows the synthetic control becomes the single nearest neighbour. The scale is that of the standardised predictors.
Value
A list with att (mean corrected gap over the post-treatment
periods), gap (the corrected gap in every period, named by the
columns of Y), gap_raw (before correction), pre_rmspe,
post_rmspe, rmspe_ratio (post_rmspe / pre_rmspe), the same three
computed on the uncorrected gaps with a _raw suffix –
plus rmspe_ratio_hybrid (corrected post over uncorrected pre) – see
the code for which to hand rmspe_pvalue() – omega (donor weights, named by the control
rownames), v (the predictor weights used), n_predictors,
bias_correct, N0 and T0, plus two flags. exact_match is TRUE
when the predictors are matched exactly: the weights are then generally
not unique and the estimate depends on which optimal weights are
returned, so it warns (class sdidlab_exact_fit) unless a penalty is
set. exact_fit is TRUE when the whole pre-period is reproduced
exactly, which also leaves the RMSPE ratios undefined (NA).
Details
The weight solve is checked against an exact quadratic-programming
solution in the test suite: at convergence it reaches the same objective,
which is at least as low as Synth's interior-point solve.
It is deliberately separate from fit_sdid() rather than another
estimator = branch there. fit_sdid() delegates to synthdid, whose
sc_estimate() matches on pre-period outcomes only and has no bias
correction; bolting predictor matching onto it would change the meaning
of the existing "sc" option and put every stored run's numbers in play.
What the two steps do
Matching. Donor weights w sit on the simplex and minimise
\((X_1 - X_0 w)' V (X_1 - X_0 w)\), where the columns of X are the
predictors and V is diagonal. Predictors are standardised by their
cross-unit standard deviation first, so V = I means "weight every
predictor equally after standardisation" – the special case of ADH that
does not require the nested optimisation. v_method = "mspe" instead
searches over diagonal V for the one minimising pre-period MSPE of the
outcome, which is what Synth does by default.
Neither search dominates, which is why "synth" runs both. Synth
optimises V with optimx from a single equal-weight start and solves
for w with an interior-point QP; this package uses several
deterministic starts and an exact active-set solve (see
sc_weights_simplex()). The V surface
is rough, so the two searches can land in different local minima, and
either can find the better one. The objective is the same and is directly
comparable, so taking the better of the two is free accuracy and the
only defensible default when the surface is this rough.
When several convex combinations of donors fit equally well the
minimiser is not unique, and different solvers can return very different
comparison groups with the same loss. penalty resolves such ties;
without it, report omega rather than assuming it.
Bias correction. Even a good synthetic control leaves a residual discrepancy in the predictors, and that discrepancy biases the estimate. The correction regresses the outcome on the predictors among the donor units, one regression per period, and subtracts the model's prediction of the discrepancy. A ridge penalty keeps it defined when the positive- weight donors are fewer than the predictors.
Note the diagnostic consequence, which is a feature and not a bug: in any period whose outcome is itself a predictor, the corrected gap is mechanically zero, because the period's regression fits that column exactly. Seeing a row of exact zeros in the gap path is confirmation that the correction ran, not evidence of a perfect match. It is what produces the flat segments at matched years in bias-corrected SCM gap figures.
References
Abadie, A., Diamond, A. and Hainmueller, J. (2010). Synthetic Control Methods for Comparative Case Studies. Journal of the American Statistical Association 105(490), 493-505.
Abadie, A. and L'Hour, J. (2021). A Penalized Synthetic Control Estimator for Disaggregated Data. Journal of the American Statistical Association 116(536), 1817-1834.
Ben-Michael, E., Feller, A. and Rothstein, J. (2021). The Augmented Synthetic Control Method. Journal of the American Statistical Association 116(536), 1789-1803.
See also
rmspe_pvalue() for the permutation inference this feeds,
fit_sdid() for the synthetic-DiD and DiD comparisons.
Examples
set.seed(1)
Y <- matrix(rnorm(11 * 8), nrow = 11,
dimnames = list(c(paste0("d", 1:10), "treated"), 2010:2017))
Y[11, 7:8] <- Y[11, 7:8] + 3
fit_scm(Y, N0 = 10, T0 = 6)$att
#> Bias correction fitted on all 10 donors, not the 3 with positive weight.
#> ℹ 6 predictors need more donors than that to be identified.
#> [1] 2.394834