A Penalized Synthetic Control Estimator

for Disaggregated Data

Alberto Abadie Jérémy L’Hour

MIT CREST
August 12, 2021

Abstract

Synthetic control methods are commonly applied in empirical research to esti-

mate the effects of treatments or interventions on aggregate outcomes. A syn-
thetic control estimator compares the outcome of a treated unit to the outcome
of a weighted average of untreated units that best resembles the characteristics of
the treated unit before the intervention. When disaggregated data are available,
constructing separate synthetic controls for each treated unit may help avoid in-
terpolation biases. However, the problem of finding a synthetic control that best
reproduces the characteristics of a treated unit may not have a unique solution.
Multiplicity of solutions is a particularly daunting challenge when the data in-
cludes many treated and untreated units. To address this challenge, we propose
a synthetic control estimator that penalizes the pairwise discrepancies between
the characteristics of the treated units and the characteristics of the units that
contribute to their synthetic controls. The penalization parameter trades off
pairwise matching discrepancies with respect to the characteristics of each unit
in the synthetic control against matching discrepancies with respect to the char-
acteristics of the synthetic control unit as a whole. We study the properties of
this estimator and propose data-driven choices of the penalization parameter.

Alberto Abadie, Department of Economics, Massachusetts Institute of Technology and NBER,

abadie@mit.edu. Jérémy L’Hour, CREST, ENSAE Paris and Insee, jeremy.l.hour@ensae.fr. We benefited
from detailed comments by Guido Imbens, and from discussions with Victor-Emmanuel Brunel, Xavier
D’Haultfœuille, Laurent Davezies, Yannick Guyonvarch, Clément de Chaisemartin, David Margolis and
participants at the 14th IZA Labor Market Policy Evaluation, the 10th French Econometrics Conference,
the MIT/NBER Conference on Synthetic Control and Related Methods and CREST and Université de
Montréal seminars. Charlie Rafkin, Rahul Singh, and Jaume Vives provided expert research assistance.
NSF support through grant SES-1756692 is gratefully acknowledged. R and Matlab scripts are available at
https://github.com/jeremylhour/pensynth.
 1. Introduction

Synthetic control methods (Abadie and Gardeazabal, 2003; Abadie et al., 2010, 2015; Doud-
chenko and Imbens, 2016) are often applied to estimate the treatment effects of aggregate
interventions (see, e.g., Kleven et al., 2013; Bohn et al., 2014; Hackmann et al., 2015; Cun-
ningham and Shah, 2018). Suppose we observe data for a unit that is affected by the
treatment or intervention of interest, as well as data on a donor pool, that is, a set of un-
treated units that are available to approximate the outcome that would have been observed
for the treated unit in the absence of the intervention. The idea behind synthetic controls
is to match each unit exposed to the intervention or treatment of interest to the weighted
average of the units in the donor pool that most closely resembles the characteristics of the
treated unit before the intervention. Once a suitable synthetic control is selected, differences
in outcomes between the treated unit and the synthetic control are taken as estimates of the
effect of the treatment on the unit exposed to the intervention of interest.
The synthetic control method is akin to nearest neighbor matching estimators (Dehejia
and Wahba, 2002; Abadie and Imbens, 2006; Imbens and Rubin, 2015) but departs from
nearest neighbor matching methods in two important aspects. First, the synthetic control
method does not impose a fixed number of matches for every treated unit. Second, instead of
using a simple average of the matched units with equal weights, the synthetic control method
matches each treated unit to a weighted average of untreated units with weights calculated
to minimize the discrepancies between the treated unit and the synthetic control in the
values of the matching variables. Synthetic control estimators retain, however, appealing
properties of nearest neighbor matching estimators. In particular, like nearest neighbor
matching estimators, synthetic control estimators use weights that are non-negative and
sum to one. In addition, synthetic control weights are often sparse. That is, like nearest
neighbor matching estimators, they only assign positive weights to a relatively small number
of untreated units. Sparsity and non-negativity of the weights, along with the fact that
synthetic control weights sum to one and define a weighted average, are important features
that allow the use of expert knowledge to evaluate and interpret the estimated counterfactuals

1
(see Abadie et al., 2015). As shown in Abadie et al. (2015), similar to the synthetic control
estimator, a regression-based estimator of the counterfactual of interest–i.e., the outcome for
the treated in the absence of an intervention–implicitly uses a linear combination of outcomes
for the untreated with weights that sum to one. However, unlike synthetic control weights,
regression weights are not explicit in the outcome in the procedure, they are not sparse,
and they can be negative or greater than one, allowing unchecked extrapolation outside the
support of the data and complicating the interpretation of the estimate and the nature of
the implicit comparison. While many applications of the synthetic control framework have
focused on cases where only one aggregate unit is exposed to the intervention of interest,
the method has found recent applications in contexts with disaggregated data, where data
sets contain multiple treated units. In some cases, especially in cases with a small number
of treated units, the interest may lie on the treament effects for each of the treated. In
other cases, especially in settings with a large number of treated units, the interest may lie
on the average effect of the treatment among the treated (see, e.g., Acemoglu et al., 2016;
Gobillon and Magnac, 2016; Kreif et al., 2016). In such settings, one could simply construct
a synthetic control for an aggregate of all treated units. However, interpolation biases may
be much smaller if the estimator of the aggregate outcome that would have been observed for
the treated in the absence of the treatment is based on the aggregation of multiple synthetic
controls, one for each treated unit.
Using synthetic controls to estimate treatment effects with disaggregated data creates
some practical challenges. In particular, when the values of the matching variables for a
treated unit fall in the convex hull of the corresponding values for the donor pool, it may
be possible to find multiple convex combinations of untreated units that perfectly reproduce
the values of the matching variables for the treated observation. That is, the best synthetic
control may not be unique. One practical consequence of the curse of dimensionality is that,
even for a moderate number of matching variables, each particular treated unit is unlikely
to fall in the convex hull of the untreated units, especially if the number of untreated units
is not very large. As a result, lack of uniqueness is rarely an issue in settings with one

2
or a small number of treated units and, if it arises, it can typically be solved by ad-hoc
methods, like increasing the number of covariates or by restricting the donor pool to units
that are similar to the treated units. In settings with many treated and many untreated
units, non-uniqueness may be an important consideration and a problem which is harder to
solve.
More generally, in contrast to common aggregate data settings with a small donor pool
(see, e.g., Abadie and Gardeazabal, 2003; Abadie et al., 2010), in settings with a large
number of units in the donor pool, single untreated units may provide close matches to
the treated units in the data. Therefore, in such settings, the researcher faces a trade-off
between minimizing the covariate discrepancy between each treated unit and its synthetic
control as a whole (synthetic control case) and minimizing the covariate discrepancy between
each treated unit and each unit that contributes to its synthetic control (matching case).
This article provides a generalized synthetic control framework for estimation and infer-
ence. We introduce a penalization parameter that trades off pairwise matching discrepancies
with respect to the characteristics of each unit in the synthetic control against matching
discrepancies with respect to the characteristics of the synthetic control unit as a whole.
This type of penalization is aimed to reduce interpolation biases by prioritizing inclusion
in the synthetic control of units that are close to the treated in the space of the matching
variables. Moreover, we show that as long as the penalization parameter is positive, the
generalized synthetic control estimator is unique and sparse. If the value of the penalization
parameter is close to zero, our procedure selects the synthetic control that minimizes the
sum of pairwise matching discrepancies (among the synthetic controls that best reproduce
the characteristic of the treated units). If the value of the penalization parameter is large,
our estimator coincides with the pair-matching estimator. We study formal properties of the
penalized synthetic control estimator and propose data-driven choices of the penalization
parameter. We propose, in addition, a bias-corrected version of the penalized synthetic con-
trol estimator, which is analogous to the one applied to matching estimators in Rubin (1973)
and Abadie and Imbens (2011). We show that the bias-correction substantially improves the

3
properties of penalized synthetic control estimators.
Doudchenko and Imbens (2016), Athey et al. (2021), Amjad et al. (2018), Arkhangelsky
et al. (2018) and Chernozhukov et al. (2021) have also proposed penalization schemes for
synthetic controls and related methods. Doudchenko and Imbens (2016), Arkhangelsky
et al. (2018) and Chernozhukov et al. (2021) use an L1 penalty term (lasso), an L2 penalty
term (ridge), or a combination of both (elastic net) to regularize synthetic control weights.
This is different from our penalization scheme, which depends on the matching discrepancy
between the treated unit and the units in the synthetic control. Athey et al. (2021) assume
an underlying sparse factor structure for the outcome under no treatment and adapt matrix
completion techniques to estimate a counterfactual. Their estimator penalizes the complexity
of the factor structure. The estimator in Amjad et al. (2018) uses low-rank approximation
techniques to de-noise the outcomes for the units in the donor pool. Then, potential outcomes
without the treatment for the treated are estimated as linear combinations of de-noised
outcomes for the units in the donor pool, with ridge-regularized coefficients. Bias-corrected
synthetic control estimators have been independently studied in Ben-Michael et al. (2021)
and Arkhangelsky et al. (2018).
The rest of the article is organized as follows. Section 2 presents the penalized synthetic
control estimator and discusses several of its geometric properties. Section 3 discusses per-
mutation inference. Section 4 presents ways to choose the penalization term. Section 5
illustrates the properties of the estimator through simulations. Section 6 contains an appli-
cation. Section 7 contains a summary of the article and conclusions. The appendix gathers
the proofs.

2. Penalized Synthetic Control

2.1. Synthetic Control for Disaggregated Data

Assume we observe n units, some of which are exposed to the treatment or intervention of
interest. We code the treatment status of unit i using the binary variable Di , so Di = 1 if i
is treated and Di = 0 otherwise. To define treatment effects we adopt a potential outcomes

4
framework, as in Rubin (1974). Let Y1i and Y0i be random variables representing potential
outcomes under treatment and under no treatment, respectively, for unit i. The effect of the
treatment for unit i is Y1i − Y0i . Realized outcomes are defined as

Y1i if Di = 1,
Yi =
Y0i if Di = 0.

Let Xi be a (p × 1)-vector of pre-treatment predictors of Y0i . We assume that we observe

(Yi , Xi ) = (Y1i , Xi ) for n1 treated observations and (Yi , Xi ) = (Y0i , Xi ) for n0 untreated
observations. Combining data for treated and nontreated we obtain the pooled data set,
{(Yi , Di , Xi )}ni=1 , with n = n0 + n1 . To simplify notation, we reorder the observations in
the data so that the n1 treated observations come first. The quantities of interest are the
treatment effects on the treated units, τi = Y1i − Y0i for i = 1, . . . , n1 , and/or the average
treatment effect on the treated (ATET):
n1
1 X

τ= Y1i − Y0i . (1)
n1 i=1

Many estimators of τ , are of the form,

n n
1 X 1 X
Yi Di − Yi (1 − Di )Vi . (2)
n1 i=1 n0 i=1

Popular estimators of this type in micro-econometrics include most notably regression (An-
grist and Pischke, 2008; Abadie et al., 2015), matching estimators (Rosenbaum and Rubin,
1983; Dehejia and Wahba, 2002; Abadie and Imbens, 2006), and propensity score weighting
estimators (Hirano et al., 2003). For example, in the case of the pair-matching estimator, the
weight Vi given to control unit i is equal to the number of times control unit i is the nearest
neighbor of a treated unit, rescaled by n0 /n1 . The synthetic control method (Abadie and
Gardeazabal, 2003; Abadie et al., 2010, 2015; Doudchenko and Imbens, 2016) also belongs
to this class of estimators. It matches each treated unit to a “synthetic control”, that is, a
weighted average of untreated units with weights chosen to make the values of the predic-
tors of the outcome variable of each synthetic control closely match the values of the same
predictors for the corresponding treated units.

5
 For any (p × 1) real vector X and any (p × p) real symmetric positive-definite matrix
Γ, define the norm kXk = (X 0 ΓX)1/2 . Because Γ is diagonalizable with strictly positive
eigenvalues, we can always transform the vector X so that the matrix Γ becomes the (p × p)
identity matrix. As a result, without loss of generality, we will consider only Γ = I. In the
synthetic control framework, model selection–that is, the choice of the variables included in
X–is operationalized through the choice Γ, which rescales or weights each predictor in X
according to its predictive power on the outcome (see Abadie et al., 2010). In a setting with
many treated and untreated units, the standard synthetic control estimation procedure is as
follows:

1. For each treated unit, i = 1, . . . , n1 , compute the n0 -vector of weights Wi∗ = (Wi,n
∗
1 +1
,...,
∗
Wi,n ) that solves
2
n
X
minn Xi − Wi,j Xj (3)

Wi ∈R 0
j=n1 +1

s.t. Wi,n1 +1 ≥ 0, . . . , Wi,n ≥ 0,

Xn
Wi,j = 1,
j=n1 +1

∗
where Wi,j is the weight given to control unit j in the synthetic control unit corre-
sponding to treated unit i. A synthetic control estimate of the effect of the treatment
on treated unit i is
n
X
∗
τbi = Yi − Wi,j Yj .
j=n1 +1

2. Averaging the treatment effects on the treated produces a synthetic control estimate
of τ ,
n1
" n
#
1 X X
∗
τb = Yi − Wi,j Yj . (4)
n1 i=1 j=n +1 1

Notice that τb is the estimator in equation (2) reweighting each nontreated unit, j = n1 +
1, . . . , n, by Vj = (n0 /n1 ) ni=1 ∗ ∗
P 1
Wi,j , with Wi,j = 0 for i ≥ n1 + 1.

6
 To simplify the exposition, so far we have described a simple cross-sectional setting. The
extension to the more common panel data setting for synthetic controls, where the same
units are observed for a number of periods—before and after the intervention happens for
the treated—is immediate and we will use it in later sections. Panel data settings with mul-
tiple treated units also raise the possibility that different treated units adopt the treatment
at different points in time. Staggered adoption of a treatment (Athey and Imbens, 2021;
Ben-Michael et al., 2019) can easily be accommodated in the synthetic control framework,
although it creates some implementation challenges related to the choice of a meaningful av-
erage of the individual treatment effects as a target parameter, and the fact that the donor
pool changes in time. Moreover, even in cross-sectional settings or when all treated units
adopt the treatment at the same time, τ in equation (1) is by no means the only possible tar-
get parameter of interest. For instance, if the data consist of a number of cities or states, one
may wish to calculate a population-weighted average treatment effect. This is, again, easy
to implement in a synthetic control framework like the one in this article, where the effect
of the treatment is estimated separately for each treated unit. Abadie (2020) provides an
introduction to synthetic control estimation and discusses feasibility and data requirements.

2.2. Penalized Synthetic Control

The main contribution of this article is to propose a penalized version of the synthetic control
estimator in equation (3). For treated unit i and given a positive penalization constant λ,
∗
the penalized synthetic control weights, Wi,j (λ), solve
n
2 n
X X
min Xi − Wi,j Xj + λ Wi,j kXi − Xj k2 (5)

Wi ∈Rn0
j=n1 +1 j=n1 +1

s.t. Wi,n1 +1 ≥ 0, . . . , Wi,n ≥ 0,

Xn
Wi,j = 1.
j=n1 +1

The penalized synthetic control estimates are

n
X
∗
τbi (λ) = Yi − Wi,j (λ)Yj .
j=n1 +1

7
for the unit-level treatment effects, τi , and
n1
" n
#
1 X X
∗
τb(λ) = Yi − Wi,j (λ)Yj (6)
n1 i=1 j=n +1
1

for the average effect on the treated, τ .

The tuning parameter λ sets the trade-off between componentwise fit and aggregate fit.
The choice of the value of λ is important and will be discussed in Section 4. The penalized
synthetic control estimator encompasses both the synthetic control estimator and the nearest
neighbor matching as special polar cases. At one end of the spectrum, as λ → 0, the penal-
ized estimator becomes the synthetic control that minimizes the sum of pairwise matching
discrepancies among the set of synthetic controls that best reproduce the characteristics of
the treated units. Our motivation to choose among synthetic controls that fit the treated
unit equally well by minimizing the sum of pairwise matching discrepancies is to reduce
worst-case interpolation biases. At the other end of the spectrum, as λ → ∞, the penalized
estimator becomes the one-match nearest neighbor matching with replacement estimator in
Abadie and Imbens (2006).
Let X0 be the (p×n0 ) matrix with column j equal to Xn1 +j , and let ∆i be the (n0 ×1) vec-
tor with j-th element equal to kXi −Xn1 +j k2 . Moreover, let ∆N
i
N
= minj=1,...,n0 kXi −Xn1 +j k2
be the smallest discrepancy between unit i and the units in the donor pool. Finally, let Wi∗ (λ)
be a solution to (5), and ∆∗i (λ) = kXi −X0 Wi∗ (λ)k2 be the square of the discrepancy between
unit i and the (penalized) synthetic control. The next lemma establishes bounds on ∆∗i (λ)
and ∆0i Wi∗ (λ).

Lemma 1 (Discrepancy Bounds) For any λ ≥ 0

0 ≤ ∆∗i (λ) ≤ ∆N
i
N
,

and for λ > 0

1 + λ NN
∆N
i
N
≤ ∆0i Wi∗ (λ) ≤ ∆i .
λ
All proofs are in the appendix.

8
 The first result in Lemma 1 states that the synthetic unit is contained in a closed ball
p
of center Xi and radius equal to the distance to the nearest neighbor, ∆N i
N
. The second
result implies that the tuning parameter λ controls the compound discrepancy between the
treated unit and the units that contribute to the synthetic control, ∆0i Wi∗ (λ).
The specific penalty term in equation (5) is one of many possible alternatives. For
2 2
instance, in the spirit of elastic nets, one could add an L2 penalty term, γ(Wi,n 1 +1
+· · ·+Wi,n ),
to the objective function in equation (5). The L1 penalty term in equation (5) has the
advantage of producing easy-to-interpret sparse solutions, which is also a feature of the
standard synthetic control estimator. In the absence of the penalty term (that is, when
λ = 0), the problem in (5) can be solved by projecting Xi on the convex hull of X0 . Existence
of sparse solutions follows from Carathéodory’s theorem. However, if λ = 0 the solution to
the problem in (5) may not be unique if Xi belongs to the convex hull of the columns of
X0 . Adopting λ > 0 penalizes solutions with potentially large interpolation biases created by
large matching discrepancies and produces uniqueness and sparsity as stated in the following
result.

Theorem 1 (Uniqueness and Sparsity) Suppose that any submatrix composed by rows
of [X00 1n0 ∆i ] has full rank, where 1n0 is the (n0 × 1) vector of ones. Then, if λ > 0 the
optimization problem in equation (5) admits a unique solution Wi∗ (λ) with at most p + 1
non-zero components.

Notice that the condition that any submatrix composed by rows of the n0 × (p + 2) matrix
[X00 1n0 ∆i ] has full rank implies that there are no two control units with the same values
of the predictors. It also implies that there is no set of control units of cardinality p + 2 or
larger such that the values of the predictors belong to a sphere with center at Xi .

Example: Consider a simple numerical example with only one covariate. Suppose, there is
one treated unit with X1 = 2 and three control units with X2 = 1, X3 = 4 and X4 = 5. This
simple setting is depicted in Figure 1.
Notice that X1 belongs to [1, 5], the convex hull of the columns of X0 , and ∆1 = (1, 4, 9)0 .
Consider first the case with λ = 0. Then, W ∗ (0) = (2/3, 1/3, 0)0 and W ∗∗ (0) = (3/4, 0, 1/4)0

9
 Figure 1: A simple example

X2 X1 X3 X4
d u d d
1 2 3 4 5

are the only two sparse solutions (with number of non-zero weights not greater than p+1 = 2)
to (5). The first sparse solution, W ∗ (0), interpolates X1 = 2 using X2 = 1 and X3 = 4. The
second sparse solution, W ∗∗ (0) is of lower quality relative to W ∗ (0) in terms of compound
discrepancy, as it uses an interpolation scheme that replaces X3 with X4 , an observation
farther away from X1 . As a result, W ∗ (0) is preferred over W ∗∗ (0) in terms of worst case
interpolation bias. However, the better compound fit of W ∗ (0) is not reflected in a better
value in the objective function in (3). Moreover, because any convex combination of W ∗ (0)
and W ∗∗ (0) is also a solution, the problem in (3) has an infinite number of solutions, W0∗ =
{aW ∗ (0) + (1 − a)W ∗∗ (0) : a ∈ [0, 1]}. Let V̄ (a) = aW ∗ (0) + (1 − a)W ∗∗ (0). The compound
discrepancy of V̄ (a) is
∆0i V̄ (a) = 3 − a.

W ∗ (0), which is obtained making a = 1, produces the lowest compound discrepancy among
all the solutions to equation (3).
When λ > 0, however, the program (5) has a unique solution, which is sparse:

(2 + λ/2, 1 − λ/2, 0)0 /3 if 0 < λ ≤ 2,


∗
W (λ) =
(1, 0, 0)0 if λ > 2.

Notice that W ∗ (λ) never puts any weight on X4 . As λ → ∞, W ∗ (λ) selects the nearest
neighbor match, and as λ → 0, W ∗ (λ) converges to W ∗ (0), the (non-penalized) synthetic
control in W0∗ with the smallest compound discrepancy. 

2.3. Geometric Properties of Penalized Synthetic Controls

In this section, we use Delaunay tessellations to characterize the geometric properties of

penalized synthetic control estimators. The results in this section provide a geometric inter-
pretation of penalized synthetic controls, and imply that there exists an estimator for the

10
case λ → 0 that does not depend on approximating the limit estimate with the one obtained
for an arbitrary small value of λ.
A Delaunay triangulation tessellates the convex hull of a set of points, {x1 , . . . , xn }, in R2
into triangles with vertices in {x1 , . . . , xn }. Each triangle of the Delaunay triangulation of
{x1 , . . . , xn } is such that its circumscribing circle does not contain any point in {x1 , . . . , xn }
in its interior. Delaunay triangulations generalize to higher dimensions, in which case they
are often referred to as Delaunay tessellations. A Delaunay tesselation in R3 is a collection
of tetrahedrons with vertices in {x1 , . . . , xn } such that their circumscribing spheres do not
contain points in {x1 , . . . , xn } in their interiors. More generally, a Delaunay tesselation in
Rp is a collection of p-simplices with vertices in {x1 , . . . , xn } such that their circumscribing
hyperspheres do not contain points of {x1 , . . . , xn } in their interiors (see, e.g., Boissonnat and
Yvinec, 1998; Okabe et al., 2000). We will refer to the simplices of a Delaunay tessellation as
Delaunay simplices. The set {x1 , . . . , xn }, along with the collection of segments connecting
the vertices of each p-simplex of a Delaunay tessellation, constitutes the Delaunay graph
induced by the tessellation. For the remainder of this section, we will assume that every
Delaunay tessellation is done on the convex hull of a set of points in general quadratic
position. We say that n points in Rp are in general quadratic position when (i) for k =
2, . . . , p, no k + 1 points lie in a (k − 1)-dimensional hyperplane of Rp (non-collinearity), and
(ii) no p + 2 points lie on the boundary of an hypersphere in Rp (non-cosphericity) (see, e.g.,
Okabe et al., 2000). If all the points in the set {x1 , . . . , xn } are in general quadratic position,
then the Delaunay tessellation of the convex hull of {x1 , . . . , xn } exists and is unique. The
assumption of general quadratic position is fairly innocuous. Realizations of random vectors
drawn from a distribution that is continuous with respect to the Lebesgue measure are in
general quadratic position with probability one.
The next theorem provides a characterization of the units contributing to a particular
synthetic control, X0 Wi∗ (λ) with λ > 0, as vertices of the Delaunay simplex containing
X0 Wi∗ (λ) in the Delaunay tessellation of Xn1 +1 , . . . , Xn .

Theorem 2 (Delaunay Property I) Let Wi∗ (λ) be a solution to the penalized synthetic

11
control problem in (5) with λ > 0. Consider the Delaunay tessellation induced by the columns
of X0 . Then, for any control unit j = n1 + 1, . . . , n, such that Xj is not a vertex of the
∗
Delaunay simplex containing X0 Wi∗ (λ), it holds that Wi,j (λ) = 0.

This result provides a notion of proximity between a synthetic control and each of the units
that contribute to it. Theorem 2 also provides a simple way to compute the solution for the
“pure synthetic control” case (λ → 0) that does not entail the choice of an arbitrarily small
value of λ to use in (5). Recall that when λ = 0, the problem of minimizing kXi − X0 W k
subject to the weight constraints has typically infinite solutions if X1 belongs to the convex
hull of the columns of X0 , in which case Xi = X0 W for all solutions. In the presence
of multiple solutions, the pure synthetic control case selects the solution that produces the
lowest compound discrepancy, W 0 ∆i , among all W such that Xi = X0 W . Directly solving (5)
for λ → 0 requires, in practice, a choice of a small value for λ. It also creates computational
difficulties, as the minimization problem is close to one with multiple solutions. Theorem 2
provides a solution to these problems, because it implies that the solution of (5) for λ → 0
can assign positive weights only to the vertices of the simplex in the Delaunay tessellation
of Xn1 +1 , . . . , Xn that contains the projection of Xi on the convex hull of the columns of X0 .
As a result, it is enough to solve (5) allowing positive weights only on the observations that
represent the vertices of the Delaunay face that contains the projection of Xi on the convex
hull of the columns of X0 . In high-dimensional settings, however, the large computation
costs of Delaunay triangulations may make this approach unfeasible.
Consider the Delaunay graph induced by the Delaunay tessellation of the convex hull of
a set of points representing the predictor values for a treated unit, i, and for all the units in
the donor pool. The next theorem shows that, in such a graph, the treated unit is connected
to all the untreated units that contribute to the synthetic control of unit i.

Theorem 3 (Delaunay Property II) Let Wi∗ (λ) be a solution to the penalized synthetic
control problem in (5) with λ > 0. Consider the Delaunay tessellation induced by the columns
of X0 and the treated Xi , and denote Ii as the indices of the points in {Xn1 +1 , . . . , Xn } that
are connected to Xi in the corresponding Delaunay graph. For any j 6∈ Ii , it holds that

12
 ∗
Wi,j (λ) = 0.

Theorem 3 provides a notion of proximity between a treated unit and the units contributing
to its penalized synthetic control. It therefore restricts the donor pool to these units con-
nected to the treated and as such provides a way to simplify the computation of the synthetic
control for λ > 0.
Figure 2 illustrates Lemma 1 and Theorems 2 and 3 in two dimensions. The top-left
panel (a) displays the treated unit (black cross) and the Delaunay triangulation of untreated
units. The top-right panel (b) draws the trajectory of the synthetic unit as λ changes (as λ
increases, the solution drifts toward the nearest neighbor and away from the treated – solid
black line) and the circle centered on the treated of radius equal to the distance between the
treated and its nearest neighbor. Notice that the synthetic unit is never located outside of
this circle, as per Lemma 1. The bottom left panel (c) shows the four untreated units that
have a non-zero weight across some solutions of the penalized synthetic control as λ changes
(black dots). They are the vertices of the two triangles where the synthetic unit is located,
as per Theorem 2. The bottom-right panel (d) shows that these units are also connected to
the treated in the augmented Delaunay triangulation (that includes the treated unit), as per
Theorem 3.
Notice that being connected to the treated unit in the augmented Delaunay triangulation
is a necessary but not sufficient condition for a unit to contribute to the synthetic control.
This can easily be seen in Figure 2 (d). For example, the nearest neighbor to the treated
unit is connected to the treated unit in the augmented Delaunay graph. This is, in fact,
a general property: the (undirected) nearest neighbor graph is always a subgraph of the
Delaunay graph. However, there are positive values of λ for which the penalized synthetic
control estimator puts zero weight on the nearest neighbor of the treated unit. This is
implied by the fact that the penalized synthetic control does not lie on a Delaunay simplex
with the nearest neighbor as one of the vertices and Theorem 2. Although the objective
function in equation (6) is a combination of the the objective functions minimized by the
unpenalized synthetic control and the nearest neighbor matching estimator, the penalized

13
synthetic controls estimator is not in general a combination of the unpenalized synthetic
control estimator and the nearest neighbor matching estimator. Kellogg et al. (2020) propose
averaging synthetic control and nearest-neighbor matching estimators.

2.4. Bias-Corrected Synthetic Control

We will also consider bias-corrected versions of synthetic control estimators. We adopt a

bias correction analogous to that implemented in Rubin (1973) and Abadie and Imbens
(2011) for matching estimators. Let µ
b0 (x) be a regression predictor of the outcome, Yi , of
an untreated unit with covariate values Xi = x. A bias-corrected version of the synthetic
control estimator in equation (6) is
n1
" n
#
1 X
X
∗


τbBC (λ) = Yi − µ
b0 (Xi ) − Wi,j (λ) Yj − µ
b0 (Xj ) . (7)
n1 i=1 j=n +1
1

Like in Abadie and Imbens (2011), the bias correction in equation (7) adjusts for mismatches
between the characteristics of the treated units and the characteristics of each of the units
that contribute to the synthetic controls. Depending on the setting and the nature and
quantity of data, µ
b0 can be a parametric or a non-parametric regression. A bias correction
of this type has been independently studied in Ben-Michael et al. (2021), who propose using
ridge regression to estimate µ
b0 (x).

3. Permutation Inference

In this section, we adapt the inferential framework in Abadie et al. (2010) to the penalized
synthetic control estimators of Section 2. Like in Abadie et al. (2010), our inferential exer-
cises compare the value of a test statistic to its permutation distribution induced by random
reassignment of the treatment variable in the data set. Aside from simulation errors, this
inferential exercise is exact by construction, regardless of the number of units in the data.
We next describe two possible implementations that employ different test statistics and per-
mutation schemes. Alternative test statistics and permutation schemes are possible and, in
practice, the choice among them should take into account the nature of the parameter(s)

14
 Figure 2: Geometric properties of penalized synthetic control estimator

(a) Treated unit (black cross) on the Delaunay (b) The synthetic unit as λ changes (solid black
triangulation of untreated units (dashed lines). line) and the circle centered on the treated
of radius equal to the distance between the
treated and its nearest neighbor.

● ●

● ●

● ●
● ●

(c) The four untreated (black dots) that have (d) Treated unit and synthetic units on the
a non-zero weight in some solutions of the pe- Delaunay triangulation augmented with the
nalized synthetic control as λ changes. treated unit.

15
of interest (e.g., individual vs. aggregate effects), the characteristics of the intervention that
is the object of the analysis, and the structure of the data set. Randomized reassignment
of the treatment in the data is taken here as a benchmark against which we evaluate the
rareness of the value of a test statistic, and it may not reflect the actual and typically un-
known treatment assignment process (see Abadie et al., 2010, 2015). Firpo and Possebom
(2018) propose a procedure to assess the sensitivity of permutation inference to deviations
from the reassignment benchmark. The permutation procedure outlined in this section is
conditional on the data and its validity does not depend on the nature of the mechanism
used to generate the data set. Alternative inferential procedures for synthetic controls have
been proposed by Chernozhukov et al. (2021) and Cattaneo et al. (2019), among others, and
they are summarized in Abadie (2020). While this section focuses exclusively on p-values,
permutation distributions are easy to visualize and report, and they contain important ad-
ditional information, like the signs and magnitudes of the test statistics (Abadie, 2020). In
addition, as in Firpo and Possebom (2018), confidence intervals around synthetic control
estimates can be obtained by inverting the results on statistical tests based on the p-values
in this section.

3.1. Inference on Aggregate Effects

Here we outline a simple permutation procedure that employs a test statistic, Tb, that mea-
sures aggregate effects for the treated. Examples of aggregate statistics of this type are the
synthetic controls estimators in equations (6) and (7). Similar to Abadie et al. (2010), in a
panel data setting Tb can be based on the ratio between the aggregate mean square prediction
error in a post-intervention period and a pre-intervention period. Let Yit be the observed
outcome for unit i at time t, and let τbit be as in equations (6) and (7) but with Yi and Yj
replaced by Yit and Yjt , respectively. Then, the ratio between the aggregate mean square
prediction error in a post-intervention period T1 ⊆ {T0 + 1, . . . , T } and a pre-intervention
period T0 ⊆ {1, . . . , T0 } is
n1
!2 , n1
!2
X X X X
τbit (λ) τbit (λ) . (8)
t∈T1 i=1 t∈T0 i=1

16
Let D obs = (D1 , ..., Dn ) be the observed treatment assignments in the data. We will write
Tb(D obs ) to indicate the value of the test statistic in the data, and Tb(D) to indicate the value
of the test statistic when the treatment values are reassigned in the data as indicated in D.
The procedure is as follows:

1. Compute the test statistic in the original data Tb(D obs ).

2. At each iteration, b = 1, . . . , B, permute at random the components of D obs to obtain

Tb(D (b) ).

3. Calculate p-values as the frequencies across iterations of the values of Tb(D (b) ) more
extreme than Tb(D obs ). Typically, for two-sided tests:
B
!
1 X n
(b) obs
o
pb = 1+ 1 |Tb(D )| ≥ |Tb(D )| .
B+1 b=1

For one sided tests:

B
!
1 X n o
pb = 1+ 1 Tb(D (b) ) ≥ Tb(D obs ) ,
B+1 b=1
or !
B
1 X n o
pb = 1+ 1 Tb(D (b) ) ≤ Tb(D obs ) .
B+1 b=1

3.2. Inference Based on the Sum of Rank Statistics of Unit-Level Treatment

Effects Estimates

Similar to Dube and Zipperer (2015), we propose a test based on the rank statistics of the
unit-level treatment effects. Unlike the test in Dube and Zipperer (2015), we calculate the
permutation distribution directly from the data. The test we employ is based on the sum
of the ranks of n1 × (B + 1) unit-level test statistics for the treated in the data and in B
random permutation of observed treatments. Individual treatment effects, Tbi , may be based
on differences in outcomes between treated and synthetic controls,
n
X
∗
Yi − Wi,j (λ)Yj ,
j=n1 +1

17
bias corrected versions of the unit-level treatment effects,
n
!
X
∗
(Yi − µ
b0 (Xi )) − Wi,j (λ)Yj − µ
b0 (Xj ) .
j=n1 +1

The test statistic may be based on unit-level versions of the mean squared prediction error
ratio in equation (8). The procedure is implemented as follows:

1. Compute unit-level test statistic for the treated, Tbi , for i = 1, . . . , n1 , under the actual
treatment assignment, D obs .

2. At each iteration b = 1, . . . , B, permute at random the components of D obs to obtain

(b) (b)
Tbi (D (b) ) for the treated. Denote these estimates Tb1 , . . . , Tbn1 (in arbitrary order).

(1) (1) (B) (B)

3. Calculate the ranks R1 , . . . , Rn1 , R1 , . . . , Rn1 , . . ., R1 , . . . , Rn1 associated to the
(1) (1) (B)
n1 ×(B+1) individual treatment effect estimates Tb1 , . . . , Tbn1 , Tb1 , . . . , Tbn1 , . . ., Tb1 , . . . ,
(B)
Tbn1 (or of their absolute values or negative values) and the sums of ranks for each
P 1 (b)
permutation, SR = ni=1 Ri , SR(b) = ni=1
P 1
Ri , b = 1, . . . , B.

4. Calculate p-values as:

B
!
1 X  (b)
pb = 1+ 1 SR ≥ SR .
B+1 b=1

4. Penalty Choice

We present two data-driven selectors for the penalty term, λ. In the context of treatment
effects estimation, cross-validation is complicated by the absence of data on a ground truth
(that is, on the values of Y0 for the treated units in the post-intervention periods, see Athey
and Imbens, 2016). Since synthetic controls are often applied to panel data, we consider a
balanced panel data setting with T periods and T0 < T pre-intervention periods. As before,
we define Yit as the outcome for unit i at time t. Adaptation of (5) and (6) to the panel
data setting is straightforward by allowing Xi to potentially include multiple pre-intervention
values of the outcome variable and of other predictors of post-intervention outcomes.

18
 The first selector proposed in this section is based on cross-validation on the outcomes on
the untreated units in the post-intervention period. The second selector minimizes prediction
error in a hold-out pre-intervention period.

4.1. Leave-One-Out Cross-Validation of Post-Intervention Outcomes for the

Untreated

This section discusses a leave-one-out cross-validation procedure to select λ by minimizing

mean squared prediction error for the untreated units in the post-intervention period. The
procedure is as follows:

1. For each control unit i = n1 + 1, ..., n, and each post-intervention period, t = T0 +

1, . . . , T , calculate
n
X
∗
τbit (λ) = Yit − Wi,j (λ)Yjt ,
j=n1 +1
j6=i

∗
where Wi,j (λ) is a synthetic control for unit i that is produced by the donor pool
{n1 + 1, . . . , n}\{i}.

2. Choose λ to minimize some measure of loss, such as the sum of the squared prediction
errors for the individual outcomes,
n
X T
X  2
τbit (λ) ,
i=n1 +1 t=T0 +1

or for the average outcomes

T n
!2
X X
τbit (λ) .
t=T0 +1 i=n1 +1

4.2. Pre-Intervention Holdout Validation on the Outcomes of the Treated

An alternative selector of λ is based on validation over the outcomes for the treated on a
hold out pre-intervention period. To simplify the exposition, and because it may be the most
natural choice, we will only describe the case when the validation period comes immediately

19
before the intervention, although other choices are possible. Let h and k be the lengths
of the training and validation periods, respectively. The validation period comprises the k
dates immediately before the intervention, and the training period comprises the h dates
immediately before the validation period. The procedure is as follows:

1. For each treated individual, i, and validation period, t ∈ {T0 − k + 1, . . . , T0 }, compute

n
X
∗
τbit (λ) = Yit − Wi,j (λ)Yjt ,
j=n1 +1

∗
where Wi,j solve (5) with X1 , . . . , Xn measured in the training period.

2. Choose λ to minimize a measure of error, such as the sum of the squared prediction
for the individual outcomes,
n1
X T0
X  2
τbit (λ) ,
i=1 t=T0 −k+1

or the squared prediction error of the aggregate outcomes,

T0 n1
!2
X X
τbit (λ) ,
t=T0 −k+1 i=1

in the validation period.

5. Simulations

This section reports the results of a Monte Carlo experiment that investigates the properties
of the penalized synthetic control estimator relative to its unpenalized version (λ = 0) and
to the nearest neighbor matching estimator in a panel data framework.
The data generating process is as follows. Let Xmi be the m-th component of Xi . The sim-
ulation design includes two periods: a pre-intervention period (t = 1), and a post-intervention
period (t = 2). Irrespective of the treatment status, the outcome at date t ∈ {1, 2} is gener-
ated by Yit = ( pm=1 Xmir
P
) /β +εit with r a positive constant governing the degree of linearity
of the outcome function. Hence, the treatment effects, τi , are equal to zero. The error terms
εit are generated as independent values from the standard normal distribution. For the n1

20
treated units, Xi , is a vector of dimension p with independent entries uniformly distributed
on [.1, .9]. For the n0 control units, Xi is a vector of the same dimension with independent en-
√
tries distributed as U , where U is uniform on [0, 1]. We set β = var ( pm=1 Xmi
p
r
P
|Di = 1),
so that var(Yi,t |Di = 1) = 2 and the signal-to-noise ratio for the treated is equal to one.
We compare the performances of synthetic control and matching estimators with M
matches per unit (see, e.g. Abadie and Imbens, 2006). We will consider these two estimators
with a fixed choice and a data-driven choice of λ and M . Under the fixed procedure, we
impose λ → 0 for the synthetic control and M = 1 in the matching estimator, encompassing
both polar cases of the penalized synthetic control estimator highlighted in this paper. The
case λ → 0 is referred to as the “pure synthetic control”. Among all the solutions to the
unpenalized synthetic control optimization problem in equation (3), it selects the one with the
smallest componentwise matching discrepancy, nj=n1 +1 Wi,j kXi − Xj k2 . The computation
P

of the pure synthetic control estimator is based on the result in Theorem 2 and discussion
thereafter. The pure synthetic control estimator is not to be confused with the non-penalized
synthetic control (λ = 0), for which we also report results, and which does not take into
account the compound discrepancy. The data-driven choice of λ and M uses the first period
outcome to minimize the mean square error (MSE) over that period. In other words, we
follow the second procedure in Section 4. At each simulation step, λ and M are chosen so
as to minimize mean square error,
n1 n
!2
1 X X
∗
Yi1 − Wi,j (λ)Yj1 ,
n1 i=1 j=n +1
1

and
 2
n1
1 X
Yi1 − 1
X
Yj1  ,
n1 i=1 M
j∈JM (i)

with respect to λ and M , respectively, where JM (i) is the set of indices of the M control units
that are the nearest to treated unit i as measured by the Euclidean norm. The parameter
λ is selected over a grid of positive values. This implies that the penalized synthetic control
estimator, which is sparse, does not nest the unpenalized one, which is not necessarily sparse

21
if the treated unit falls inside the convex hull defined by the values of the predictors in the
donor pool. The number of matches is selected over the set of positive integers not greater
than 20. We also report a bias-corrected version of the estimators as in Section 2.4, based
on a linear specification for the regression function, µ
b0 .
For each configuration and each estimator, we report four statistics computed on the
estimates of the treatment effects on the treated units in the second period. The first
statistic is the individual root mean square error (RMSE indiv.), computed as the square
root of average individual-level MSE across simulations,
B n1 
!1/2
1 X 1 X (b)
 2
τbi2 .
B b=1 n1 i=1

The second is the aggregate-level RMSE (RMSE aggreg.) across simulations,


B n1
!2 1/2
1 1
X X (b)
τb  .
B b=1 n1 i=1 i2

The third is the absolute value of the bias across simulations (|Bias|),
 B n1

1 X 1 X 
(b) 
τbi2  .

B n1 i=1


b=1

The last is the average density defined as the average number of untreated units used as
controls for each treated unit, i.e., number of non-zero entries of Wi∗ (λ) or number of matches
in the optimized matching procedure.
The results are reported in Tables 1, 2 and 3 for n0 ∈ {20, 40, 100} respectively, each
time with n1 = 10. Table 4 reports results for n1 = 100, n0 = 400. Each table is divided
into sixteen blocks corresponding to a particular value of (p, r). Each block is divided
into two parts: the upper half reports the results without bias-correction and the lower
half reports results with a bias-correction based on a linear specification of the regression
function. Results are color-coded column-by-column within each half-block on a continuous
color scale. For the upper half-block, the scale varies from dark blue (minimum column
value) to light yellow (maximum column value). For the lower half-block, the scale varies
from bright red (minimum column value) to light yellow (maximum column value).

22
 Some clear patterns emerge from the results in Tables 1-3. First, for most parameter
values the penalized and the pure synthetic control estimators outperform the matching
procedures across all three measures of performance. This advantage appears to be increasing
with p, the dimension of the covariates. Second, in terms of aggregate RMSE and bias, the
unpenalized synthetic control estimator shows mixed results, especially when p is small and
r is large, but catches up with the pure synthetic control estimator as p increases, which
is expected. Indeed, the pure and unpenalized synthetic control estimators coincide for
treated units outside of the convex hull of the untreated. As the dimensionality of the
matching variables increases, the probability that a treated unit falls outside the convex hull
of the untreated increases. In terms of individual RMSE, the unpenalized synthetic control
estimator behaves very well at the cost of large reductions in sparsity and, therefore, at the
cost of interpretability of the individual estimates. Third, the advantage of the penalized and
pure synthetic control estimators with respect to the bias slightly decreases as the degree
of the outcome function r increases. When r is relatively large, the matching procedure
displays a low bias as expected, albeit at the expense of a very large individual RMSE.
These three observations are magnified in Table 4 where the penalized synthetic control
performs consistently well in each of the sixteen blocks. The biases of the estimators go
down substantially when we adopt the bias-correction procedure of Section 2.4. Here, it is
more difficult to rank estimators based on the simulation, as the amount of bias corrected
by the procedure is different for each estimator in a way that may be directly linked to
the simulation design. That said, the overall patterns of relative performance of the bias-
corrected estimators is similar to that of the the non-corrected estimators, albeit with more
muted differences in performance. Overall, the penalized synthetic control estimator strikes
a favorable bias-variance trade-off in Tables 1-4 by combining the strength of matching and
(un-penalized) synthetic control.

23
 Table 1: Monte-Carlo Simulations, n1 = 10, n0 = 20
r=1 r = 1.2 r = 1.4 r=2
RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density
indiv. aggreg. indiv. aggreg. indiv. aggreg. indiv. aggreg.
p = 2, average number of treated outside convex hull: 4.35
Pen. Synth. 1.3681 0.6014 0.2096 2.2814 1.3618 0.5987 0.2096 2.2823 1.3567 0.5964 0.2108 2.2853 1.3493 0.5919 0.2135 2.2796
Unpen. Synth. 1.2953 0.6297 0.2084 11.7803 1.2921 0.6353 0.2367 11.7803 1.2914 0.6430 0.2646 11.7803 1.3023 0.6767 0.3466 11.7803
Pure Synth. 1.3437 0.6008 0.2009 2.5124 1.3395 0.6000 0.2062 2.5124 1.3364 0.5998 0.2112 2.5124 1.3319 0.6017 0.2260 2.5124
Matching 1.5280 0.6368 0.2357 1.5233 0.6330 0.2293 1.5198 0.6295 0.2229 1.5157 0.6219 0.2060
Opt. Matching 1.3603 0.6749 0.4174 4.5260 1.3541 0.6709 0.4095 4.4750 1.3493 0.6673 0.4033 4.4100 1.3381 0.6591 0.3935 4.4120
Pen. Synth. (BC) 1.3261 0.5714 0.0045 1.3263 0.5713 0.0061 1.3281 0.5715 0.0154 1.3441 0.5773 0.0423
Unpen. Synth. (BC) 1.2535 0.5968 0.0140 1.2554 0.5970 0.0320 1.2608 0.5995 0.0497 1.2935 0.6183 0.1030
Pure Synth. (BC) 1.3034 0.5708 0.0065 1.3041 0.5700 0.0015 1.3068 0.5704 0.0036 1.3234 0.5766 0.0175
Matching (BC) 1.4492 0.5948 0.0031 1.4504 0.5948 0.0148 1.4539 0.5962 0.0324 1.4744 0.6071 0.0814
Opt. Matching (BC) 1.2271 0.5285 0.0063 1.2293 0.5292 0.0303 1.2359 0.5319 0.0531 1.2676 0.5505 0.1184
p = 4, average number of treated outside convex hull: 8.84
Pen. Synth. 1.4731 0.8102 0.5713 2.9922 1.4749 0.8215 0.5868 2.9868 1.4768 0.8344 0.6024 2.9741 1.4902 0.8659 0.6398 2.9361
Unpen. Synth. 1.4453 0.8016 0.5590 4.9819 1.4482 0.8199 0.5860 4.9819 1.4529 0.8387 0.6120 4.9819 1.4771 0.8983 0.6873 4.9819
Pure Synth. 1.4480 0.8012 0.5600 3.3532 1.4509 0.8178 0.5839 3.3532 1.4553 0.8347 0.6068 3.3532 1.4768 0.8877 0.6728 3.3532
Matching 1.7010 0.8992 0.6264 1.6989 0.8994 0.6240 1.6980 0.8996 0.6212 1.7033 0.9022 0.6133
Opt. Matching 1.5795 0.9682 0.7764 3.6150 1.5787 0.9698 0.7750 3.5250 1.5787 0.9709 0.7712 3.4870 1.5763 0.9789 0.7711 3.4050
Pen. Synth. (BC) 1.3051 0.6096 0.0078 1.3096 0.6132 0.0197 1.3180 0.6204 0.0317 1.3601 0.6493 0.0715
Unpen. Synth. (BC) 1.2897 0.6168 0.0151 1.2938 0.6211 0.0157 1.3024 0.6276 0.0167 1.3491 0.6574 0.0181
Pure Synth. (BC) 1.2927 0.6156 0.0141 1.2969 0.6200 0.0178 1.3050 0.6266 0.0219 1.3488 0.6558 0.0326

24
Matching (BC) 1.4455 0.6431 0.0045 1.4485 0.6468 0.0248 1.4556 0.6533 0.0540 1.4967 0.6859 0.1372
Opt. Matching (BC) 1.2815 0.6003 0.0132 1.2878 0.6038 0.0445 1.2990 0.6121 0.0767 1.3469 0.6527 0.1671
p = 8, average number of treated outside convex hull: 10.00
Pen. Synth. 1.8514 1.3112 1.1722 3.7048 1.8707 1.3385 1.2007 3.6512 1.8874 1.3615 1.2250 3.5926 1.9428 1.4326 1.2958 3.4703
Unpen Synth. 1.8275 1.3144 1.1826 4.3612 1.8494 1.3498 1.2216 4.3612 1.8717 1.3843 1.2589 4.3612 1.9426 1.4846 1.3650 4.3612
Pure Synth. 1.8275 1.3144 1.1826 4.3601 1.8494 1.3498 1.2216 4.3601 1.8717 1.3843 1.2589 4.3601 1.9426 1.4846 1.3650 4.3601
Matching 2.0989 1.3945 1.2228 2.1094 1.4080 1.2361 2.1204 1.4206 1.2480 2.1580 1.4569 1.2802
Opt. Matching 2.0103 1.4916 1.3637 3.2010 2.0191 1.5073 1.3789 3.0990 2.0299 1.5179 1.3874 2.9480 2.0575 1.5491 1.4146 2.7920
Pen. Synth. (BC) 1.5372 0.8046 0.0167 1.5440 0.8090 0.0405 1.5556 0.8176 0.0670 1.6139 0.8557 0.1408
Unpen. Synth. (BC) 1.5172 0.7980 0.0117 1.5213 0.8007 0.0283 1.5303 0.8062 0.0451 1.5824 0.8363 0.0926
Pure Synth. (BC) 1.5172 0.7980 0.0116 1.5213 0.8008 0.0282 1.5304 0.8062 0.0451 1.5824 0.8363 0.0926
Matching (BC) 1.6450 0.8274 0.0310 1.6501 0.8320 0.0665 1.6604 0.8402 0.1021 1.7179 0.8838 0.2046
Opt. Matching (BC) 1.5497 0.8108 0.0220 1.5548 0.8142 0.0564 1.5676 0.8247 0.0927 1.6244 0.8671 0.1958
p = 10, average number of treated outside convex hull: 10.00
Pen. Synth. 2.0497 1.5653 1.4361 3.9785 2.0759 1.6012 1.4736 3.9163 2.1005 1.6322 1.5060 3.8594 2.1721 1.7179 1.5894 3.6503
Unpen Synth. 2.0307 1.5705 1.4510 4.7351 2.0618 1.6144 1.4975 4.7351 2.0928 1.6568 1.5421 4.7351 2.1871 1.7789 1.6683 4.7351
Pure Synth. 2.0307 1.5705 1.4510 4.7351 2.0618 1.6144 1.4975 4.7351 2.0928 1.6568 1.5421 4.7351 2.1871 1.7789 1.6683 4.7351
Matching 2.2908 1.6316 1.4763 2.3075 1.6520 1.4956 2.3241 1.6710 1.5132 2.3760 1.7249 1.5610
Opt. Matching 2.2098 1.7203 1.5950 3.0870 2.2270 1.7413 1.6142 2.9670 2.2438 1.7601 1.6319 2.8680 2.2917 1.8147 1.6816 2.7000
Pen. Synth. (BC) 1.7845 1.0655 0.0166 1.7881 1.0679 0.0427 1.7961 1.0721 0.0714 1.8489 1.1045 0.1600
Unpen. Synth. (BC) 1.7705 1.0598 0.0172 1.7725 1.0604 0.0383 1.7794 1.0638 0.0596 1.8264 1.0888 0.1204
Pure Synth. (BC) 1.7705 1.0598 0.0172 1.7725 1.0604 0.0383 1.7794 1.0638 0.0596 1.8264 1.0888 0.1204
Matching (BC) 1.8667 1.0867 0.0245 1.8691 1.0881 0.0626 1.8769 1.0931 0.1008 1.9285 1.1272 0.2112
Opt. Matching (BC) 1.7970 1.0698 0.0192 1.8008 1.0724 0.0568 1.8103 1.0779 0.0940 1.8629 1.1130 0.2027
 Table 2: Monte-Carlo Simulations, n1 = 10, n0 = 40
r=1 r = 1.2 r = 1.4 r=2
RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density
indiv. aggreg. indiv. aggreg. indiv. aggreg. indiv. aggreg.
p = 2, average number of treated outside convex hull: 2.88
Pen. Synth. 1.2955 0.4948 0.1258 2.4458 1.2932 0.4934 0.1275 2.4432 1.2920 0.4945 0.1296 2.4466 1.2876 0.4940 0.1332 2.4393
Unpen. Synth. 1.1985 0.5144 0.0973 27.6213 1.1970 0.5211 0.1420 27.6213 1.1989 0.5322 0.1850 27.6213 1.2202 0.5835 0.3059 27.6213
Pure Synth. 1.2695 0.4923 0.1132 2.6839 1.2673 0.4927 0.1192 2.6839 1.2658 0.4933 0.1245 2.6839 1.2641 0.4963 0.1385 2.6839
Matching 1.4757 0.5442 0.1626 1.4715 0.5410 0.1574 1.4684 0.5382 0.1523 1.4636 0.5322 0.1389
Opt. Matching 1.2582 0.5516 0.3226 5.8850 1.2498 0.5483 0.3216 5.9820 1.2440 0.5433 0.3167 6.0020 1.2333 0.5350 0.3034 6.1120
Pen. Synth. (BC) 1.2707 0.4749 0.0103 1.2726 0.4757 0.0067 1.2754 0.4785 0.0036 1.2852 0.4853 0.0081
Unpen. Synth. (BC) 1.1772 0.5024 0.0083 1.1790 0.5035 0.0313 1.1847 0.5091 0.0693 1.2183 0.5449 0.1762
Pure Synth. (BC) 1.2494 0.4782 0.0076 1.2504 0.4791 0.0085 1.2524 0.4805 0.0089 1.2623 0.4865 0.0088
Matching (BC) 1.4237 0.5078 0.0224 1.4245 0.5077 0.0108 1.4266 0.5086 0.0006 1.4385 0.5148 0.0318
Opt. Matching (BC) 1.1637 0.4380 0.0071 1.1612 0.4376 0.0126 1.1648 0.4388 0.0322 1.1893 0.4570 0.0915
p = 4, average number of treated outside convex hull: 7.66
Pen. Synth. 1.3772 0.6611 0.4131 3.3715 1.3769 0.6693 0.4291 3.3603 1.3786 0.6797 0.4459 3.3563 1.3897 0.7059 0.4848 3.3040
Unpen. Synth. 1.3373 0.6500 0.4034 10.9692 1.3397 0.6704 0.4395 10.9692 1.3446 0.6920 0.4744 10.9692 1.3725 0.7628 0.5745 10.9692
Pure Synth. 1.3470 0.6496 0.4024 3.7334 1.3489 0.6646 0.4293 3.7334 1.3524 0.6801 0.4550 3.7334 1.3712 0.7297 0.5282 3.7334
Matching 1.6240 0.7609 0.4792 1.6202 0.7572 0.4749 1.6177 0.7537 0.4704 1.6183 0.7462 0.4584
Opt. Matching 1.4690 0.8333 0.6701 4.6600 1.4634 0.8299 0.6667 4.5690 1.4598 0.8283 0.6653 4.5340 1.4529 0.8283 0.6662 4.5050
Pen. Synth. (BC) 1.2462 0.5056 0.0031 1.2487 0.5058 0.0056 1.2527 0.5061 0.0066 1.2815 0.5177 0.0201
Unpen. Synth. (BC) 1.2107 0.5047 0.0018 1.2128 0.5056 0.0201 1.2189 0.5088 0.0377 1.2565 0.5296 0.0887
Pure Synth. (BC) 1.2214 0.5030 0.0008 1.2229 0.5032 0.0099 1.2274 0.5049 0.0183 1.2551 0.5166 0.0423

25
Matching (BC) 1.4267 0.5576 0.0096 1.4283 0.5573 0.0338 1.4330 0.5590 0.0579 1.4615 0.5739 0.1254
Opt. Matching (BC) 1.2074 0.4860 0.0061 1.2085 0.4868 0.0210 1.2132 0.4887 0.0480 1.2428 0.5106 0.1209
p = 8, average number of treated outside convex hull: 9.97
Pen. Synth. 1.6953 1.1269 0.9992 4.3286 1.7109 1.1534 1.0274 4.2726 1.7261 1.1782 1.0546 4.2336 1.7735 1.2441 1.1228 4.0561
Unpen Synth. 1.6654 1.1153 0.9917 5.0143 1.6846 1.1511 1.0323 5.0143 1.7046 1.1860 1.0713 5.0143 1.7706 1.2884 1.1827 5.0143
Pure Synth. 1.6655 1.1154 0.9918 4.9575 1.6846 1.1512 1.0323 4.9575 1.7047 1.1861 1.0713 4.9575 1.7706 1.2883 1.1825 4.9575
Matching 1.9804 1.2263 1.0740 1.9864 1.2345 1.0819 1.9932 1.2421 1.0887 2.0196 1.2650 1.1075
Opt. Matching 1.8812 1.3390 1.2200 3.6280 1.8845 1.3451 1.2254 3.4810 1.8894 1.3506 1.2284 3.3950 1.9083 1.3736 1.2457 3.2840
Pen. Synth. (BC) 1.2816 0.5653 0.0127 1.2875 0.5685 0.0259 1.2965 0.5735 0.0400 1.3449 0.6019 0.0909
Unpen. Synth. (BC) 1.2607 0.5649 0.0152 1.2633 0.5665 0.0176 1.2701 0.5697 0.0207 1.3108 0.5873 0.0288
Pure Synth. (BC) 1.2607 0.5649 0.0151 1.2634 0.5665 0.0176 1.2702 0.5697 0.0207 1.3108 0.5874 0.0290
Matching (BC) 1.4708 0.6122 0.0214 1.4741 0.6155 0.0556 1.4817 0.6221 0.0899 1.5267 0.6580 0.1880
Opt. Matching (BC) 1.3084 0.5818 0.0072 1.3128 0.5851 0.0416 1.3220 0.5910 0.0773 1.3733 0.6336 0.1805
p = 10, average number of treated outside convex hull: 10.00
Pen. Synth. 1.8726 1.3592 1.2501 4.6745 1.8981 1.3961 1.2893 4.5992 1.9215 1.4282 1.3231 4.5124 1.9877 1.5135 1.4104 4.2991
Unpen Synth. 1.8494 1.3567 1.2503 5.4268 1.8770 1.3995 1.2971 5.4268 1.9050 1.4410 1.3419 5.4268 1.9926 1.5616 1.4697 5.4268
Pure Synth. 1.8494 1.3567 1.2503 5.4268 1.8770 1.3995 1.2971 5.4268 1.9050 1.4410 1.3419 5.4268 1.9926 1.5616 1.4697 5.4268
Matching 2.1648 1.4732 1.3334 2.1759 1.4867 1.3470 2.1875 1.4993 1.3593 2.2263 1.5360 1.3934
Opt. Matching 2.0708 1.5651 1.4577 3.6240 2.0802 1.5762 1.4687 3.4790 2.0901 1.5882 1.4806 3.3780 2.1193 1.6236 1.5126 3.2200
Pen. Synth. (BC) 1.3179 0.6141 0.0064 1.3227 0.6167 0.0207 1.3316 0.6211 0.0387 1.3788 0.6485 0.0978
Unpen. Synth. (BC) 1.2966 0.6077 0.0085 1.2988 0.6077 0.0153 1.3054 0.6098 0.0226 1.3462 0.6260 0.0432
Pure Synth. (BC) 1.2966 0.6077 0.0085 1.2988 0.6077 0.0153 1.3054 0.6098 0.0226 1.3462 0.6260 0.0432
Matching (BC) 1.4854 0.6578 0.0057 1.4874 0.6582 0.0425 1.4942 0.6625 0.0795 1.5396 0.6953 0.1854
Opt. Matching (BC) 1.3534 0.6279 0.0009 1.3573 0.6291 0.0366 1.3659 0.6356 0.0730 1.4120 0.6696 0.1823
 Table 3: Monte-Carlo Simulations, n1 = 10, n0 = 100
r=1 r = 1.2 r = 1.4 r=2
RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density
indiv. aggreg. indiv. aggreg. indiv. aggreg. indiv. aggreg.
p = 2, average number of treated outside convex hull: 1.43
Pen. Synth. 1.2724 0.4375 0.0688 2.5804 1.2711 0.4367 0.0708 2.5783 1.2700 0.4373 0.0715 2.5806 1.2680 0.4373 0.0745 2.5795
Unpen. Synth. 1.1475 0.4517 0.0553 75.6452 1.1504 0.4633 0.1205 75.6452 1.1578 0.4826 0.1815 75.6452 1.1985 0.5668 0.3450 75.6452
Pure Synth. 1.2475 0.4281 0.0647 2.8435 1.2470 0.4285 0.0699 2.8435 1.2466 0.4290 0.0741 2.8435 1.2464 0.4303 0.0838 2.8435
Matching 1.4454 0.4905 0.0816 1.4430 0.4895 0.0786 1.4414 0.4887 0.0758 1.4390 0.4876 0.0687
Opt. Matching 1.1725 0.4575 0.2407 8.3450 1.1673 0.4538 0.2356 8.4790 1.1618 0.4517 0.2307 8.5130 1.1498 0.4460 0.2271 8.8670
Pen. Synth. (BC) 1.2648 0.4317 0.0236 1.2650 0.4313 0.0235 1.2654 0.4326 0.0226 1.2684 0.4341 0.0205
Unpen. Synth. (BC) 1.1399 0.4470 0.0169 1.1441 0.4547 0.0806 1.1528 0.4706 0.1400 1.1976 0.5463 0.2990
Pure Synth. (BC) 1.2405 0.4228 0.0264 1.2411 0.4233 0.0299 1.2420 0.4239 0.0326 1.2455 0.4259 0.0378
Matching (BC) 1.4264 0.4794 0.0164 1.4267 0.4798 0.0108 1.4277 0.4805 0.0054 1.4328 0.4835 0.0090
Opt. Matching (BC) 1.1176 0.3791 0.0249 1.1186 0.3782 0.0090 1.1203 0.3801 0.0054 1.1300 0.3863 0.0425
p = 4, average number of treated outside convex hull: 5.82
Pen. Synth. 1.2868 0.5276 0.2525 3.6931 1.2821 0.5322 0.2684 3.6955 1.2810 0.5377 0.2818 3.6891 1.2856 0.5562 0.3205 3.6450
Unpen. Synth. 1.2250 0.5073 0.2250 33.7853 1.2269 0.5291 0.2744 33.7853 1.2323 0.5541 0.3219 33.7853 1.2664 0.6410 0.4558 33.7853
Pure Synth. 1.2469 0.5113 0.2326 4.1437 1.2477 0.5232 0.2597 4.1437 1.2499 0.5358 0.2854 4.1437 1.2634 0.5771 0.3572 4.1437
Matching 1.5503 0.6203 0.3304 1.5460 0.6166 0.3248 1.5431 0.6134 0.3194 1.5425 0.6073 0.3067
Opt. Matching 1.3286 0.6638 0.4936 5.7240 1.3232 0.6602 0.4892 5.6550 1.3171 0.6570 0.4831 5.6230 1.3114 0.6563 0.4770 5.7180
Pen. Synth. (BC) 1.2196 0.4541 0.0202 1.2181 0.4531 0.0152 1.2212 0.4542 0.0125 1.2410 0.4621 0.0072
Unpen. Synth. (BC) 1.1604 0.4474 0.0238 1.1630 0.4487 0.0156 1.1700 0.4545 0.0533 1.2116 0.4919 0.1596
Pure Synth. (BC) 1.1835 0.4479 0.0162 1.1849 0.4485 0.0009 1.1885 0.4505 0.0169 1.2085 0.4620 0.0610

26
Matching (BC) 1.4308 0.5160 0.0220 1.4332 0.5181 0.0412 1.4378 0.5214 0.0598 1.4618 0.5365 0.1103
Opt. Matching (BC) 1.1562 0.4231 0.0299 1.1587 0.4282 0.0527 1.1641 0.4345 0.0773 1.1957 0.4608 0.1436
p = 8, average number of treated outside convex hull: 9.79
Pen. Synth. 1.5478 0.9490 0.8178 5.0248 1.5631 0.9779 0.8521 4.9698 1.5787 1.0064 0.8834 4.8960 1.6196 1.0743 0.9562 4.6851
Unpen Synth. 1.5221 0.9349 0.8024 6.6565 1.5394 0.9741 0.8489 6.6565 1.5581 1.0126 0.8935 6.6565 1.6231 1.1262 1.0207 6.6565
Pure Synth. 1.5222 0.9352 0.8027 5.6929 1.5395 0.9738 0.8485 5.6929 1.5581 1.0118 0.8924 5.6929 1.6223 1.1236 1.0176 5.6929
Matching 1.8664 1.0824 0.9207 1.8668 1.0849 0.9225 1.8684 1.0873 0.9238 1.8810 1.0964 0.9285
Opt. Matching 1.7385 1.1782 1.0613 4.3520 1.7385 1.1810 1.0642 4.2450 1.7376 1.1838 1.0656 4.2330 1.7405 1.1895 1.0683 4.1050
Pen. Synth. (BC) 1.1927 0.4670 0.0120 1.1979 0.4686 0.0120 1.2047 0.4731 0.0098 1.2414 0.4932 0.0122
Unpen. Synth. (BC) 1.1703 0.4655 0.0145 1.1722 0.4671 0.0293 1.1780 0.4705 0.0431 1.2146 0.4888 0.0826
Pure Synth. (BC) 1.1705 0.4655 0.0147 1.1723 0.4670 0.0289 1.1780 0.4702 0.0420 1.2135 0.4877 0.0795
Matching (BC) 1.4286 0.5245 0.0046 1.4297 0.5267 0.0282 1.4346 0.5319 0.0608 1.4686 0.5622 0.1524
Opt. Matching (BC) 1.2186 0.4753 0.0069 1.2228 0.4767 0.0272 1.2276 0.4819 0.0620 1.2664 0.5178 0.1645
p = 10, average number of treated outside convex hull: 9.97
Pen. Synth. 1.6937 1.1442 1.0379 5.5262 1.7173 1.1818 1.0793 5.4363 1.7390 1.2146 1.1155 5.3607 1.7991 1.2992 1.2051 5.0745
Unpen Synth. 1.6639 1.1232 1.0205 6.3829 1.6884 1.1678 1.0704 6.3829 1.7138 1.2113 1.1184 6.3829 1.7960 1.3377 1.2551 6.3829
Pure Synth. 1.6640 1.1232 1.0204 6.2551 1.6884 1.1677 1.0703 6.2551 1.7138 1.2111 1.1181 6.2551 1.7959 1.3373 1.2547 6.2551
Matching 2.0378 1.3018 1.1663 2.0431 1.3094 1.1741 2.0490 1.3165 1.1811 2.0731 1.3383 1.2013
Opt. Matching 1.9132 1.3796 1.2830 3.7810 1.9166 1.3863 1.2898 3.6950 1.9208 1.3930 1.2961 3.6600 1.9408 1.4155 1.3174 3.5530
Pen. Synth. (BC) 1.2011 0.4775 0.0042 1.2059 0.4784 0.0050 1.2129 0.4811 0.0104 1.2540 0.5012 0.0401
Unpen. Synth. (BC) 1.1796 0.4655 0.0088 1.1812 0.4655 0.0006 1.1865 0.4671 0.0090 1.2215 0.4794 0.0332
Pure Synth. (BC) 1.1796 0.4656 0.0089 1.1812 0.4656 0.0004 1.1865 0.4672 0.0087 1.2213 0.4795 0.0328
Matching (BC) 1.4430 0.5533 0.0046 1.4447 0.5550 0.0398 1.4507 0.5600 0.0750 1.4896 0.5917 0.1747
Opt. Matching (BC) 1.2477 0.4965 0.0026 1.2497 0.5001 0.0345 1.2560 0.5062 0.0706 1.2981 0.5381 0.1748
 Table 4: Monte-Carlo Simulations, n1 = 100, n0 = 400
r=1 r = 1.2 r = 1.4 r=2
RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density RMSE RMSE |Bias| Density
indiv. aggreg. indiv. aggreg. indiv. aggreg. indiv. aggreg.
p = 2, average number of treated outside convex hull: 1.43
Pen. Synth. 1.2359 0.1521 0.0004 2.9157 1.2359 0.1519 0.0016 2.9155 1.2358 0.1519 0.0029 2.9156 1.2356 0.1520 0.0051 2.9156
Unpen. Synth. 1.0819 0.1965 0.0030 189.1356 1.0838 0.2073 0.0656 189.1356 1.0898 0.2350 0.1276 189.1356 1.1226 0.3489 0.2859 189.1356
Pure Synth. 1.2322 0.1510 0.0017 2.9586 1.2321 0.1509 0.0006 2.9586 1.2320 0.1509 0.0024 2.9586 1.2319 0.1509 0.0055 2.9586
Matching 1.4241 0.1625 0.0089 1.4235 0.1623 0.0076 1.4231 0.1622 0.0065 1.4227 0.1620 0.0036
Opt. Matching 1.0623 0.1651 0.1056 13.7820 1.0588 0.1630 0.1025 14.0870 1.0557 0.1607 0.0989 14.3150 1.0494 0.1547 0.0903 14.9920
Pen. Synth. (BC) 1.2352 0.1519 0.0076 1.2353 0.1516 0.0058 1.2354 0.1517 0.0048 1.2359 0.1519 0.0034
Unpen. Synth. (BC) 1.0812 0.1966 0.0093 1.0833 0.2056 0.0590 1.0895 0.2318 0.1207 1.1229 0.3435 0.2784
Pure Synth. (BC) 1.2316 0.1509 0.0080 1.2316 0.1508 0.0059 1.2317 0.1508 0.0044 1.2322 0.1508 0.0020
Matching (BC) 1.4199 0.1625 0.0104 1.4201 0.1627 0.0124 1.4205 0.1629 0.0143 1.4222 0.1637 0.0192
Opt. Matching (BC) 1.0442 0.1271 0.0070 1.0438 0.1285 0.0164 1.0441 0.1302 0.0260 1.0484 0.1387 0.0522
p = 4, average nb. treated outside convex hull: 31.64
Pen. Synth. 1.1939 0.2045 0.1116 4.4127 1.1941 0.2140 0.1296 4.4056 1.1953 0.2243 0.1470 4.3925 1.2022 0.2536 0.1899 4.3451
Unpen. Synth. 1.1396 0.2255 0.1021 82.2579 1.1431 0.2628 0.1706 82.2579 1.1515 0.3083 0.2351 82.2579 1.1983 0.4568 0.4106 82.2579
Pure Synth. 1.1862 0.2020 0.1012 4.5760 1.1872 0.2148 0.1252 4.5760 1.1891 0.2283 0.1472 4.5760 1.1980 0.2696 0.2053 4.5760
Matching 1.4817 0.2698 0.2012 1.4788 0.2662 0.1969 1.4770 0.2630 0.1929 1.4771 0.2564 0.1842
Opt. Matching 1.1872 0.4041 0.3773 8.2250 1.1805 0.4018 0.3750 8.3720 1.1744 0.3995 0.3726 8.5480 1.1620 0.3962 0.3695 9.0720
Pen. Synth. (BC) 1.1701 0.1668 0.0088 1.1708 0.1681 0.0226 1.1726 0.1712 0.0352 1.1825 0.1829 0.0633
Unpen. Synth. (BC) 1.1158 0.1977 0.0104 1.1194 0.2117 0.0754 1.1281 0.2411 0.1364 1.1768 0.3647 0.3022
Pure Synth. (BC) 1.1634 0.1714 0.0095 1.1644 0.1741 0.0300 1.1664 0.1788 0.0485 1.1765 0.1996 0.0969

27
Matching (BC) 1.4182 0.1745 0.0057 1.4189 0.1745 0.0060 1.4210 0.1754 0.0172 1.4324 0.1817 0.0463
Opt. Matching (BC) 1.0731 0.1397 0.0078 1.0727 0.1401 0.0107 1.0736 0.1428 0.0293 1.0851 0.1632 0.0812
p = 8, average nb. treated outside convex hull: 89.06
Pen. Synth. 1.3329 0.5593 0.5220 6.3437 1.3467 0.6031 0.5698 6.2703 1.3624 0.6433 0.6130 6.1805 1.4094 0.7375 0.7112 5.8088
Unpen Synth. 1.3220 0.5492 0.5093 12.4059 1.3370 0.6053 0.5699 12.4059 1.3551 0.6600 0.6281 12.4059 1.4250 0.8174 0.7921 12.4059
Pure Synth. 1.3240 0.5486 0.5092 6.6374 1.3387 0.6000 0.5647 6.6374 1.3557 0.6499 0.6178 6.6374 1.4190 0.7932 0.7674 6.6374
Matching 1.7224 0.7350 0.7026 1.7202 0.7335 0.7009 1.7196 0.7320 0.6991 1.7268 0.7305 0.6967
Opt. Matching 1.5189 0.8892 0.8681 4.8290 1.5147 0.8907 0.8698 4.8240 1.5118 0.8918 0.8708 4.8260 1.5077 0.8961 0.8746 4.8420
Pen. Synth. (BC) 1.1361 0.1927 0.0028 1.1390 0.1938 0.0272 1.1458 0.1982 0.0459 1.1776 0.2157 0.0652
Unpen. Synth. (BC) 1.1274 0.1986 0.0049 1.1298 0.2040 0.0464 1.1370 0.2173 0.0859 1.1811 0.2844 0.1967
Pure Synth. (BC) 1.1298 0.1969 0.0048 1.1318 0.2012 0.0412 1.1377 0.2117 0.0756 1.1739 0.2660 0.1720
Matching (BC) 1.4188 0.2025 0.0010 1.4201 0.2040 0.0268 1.4245 0.2096 0.0541 1.4525 0.2423 0.1293
Opt. Matching (BC) 1.1259 0.1770 0.0008 1.1267 0.1791 0.0294 1.1311 0.1869 0.0599 1.1589 0.2319 0.1465
p = 10, average nb. treated outside convex hull: 97.20
Pen. Synth. 1.4496 0.7692 0.7434 7.0704 1.4722 0.8209 0.7972 6.9771 1.4959 0.8685 0.8466 6.8578 1.5611 0.9738 0.9544 6.3061
Unpen Synth. 1.4396 0.7574 0.7308 8.4734 1.4631 0.8157 0.7914 8.4734 1.4886 0.8720 0.8495 8.4734 1.5762 1.0330 1.0141 8.4734
Pure Synth. 1.4401 0.7572 0.7306 7.3541 1.4635 0.8144 0.7900 7.3541 1.4888 0.8696 0.8470 7.3541 1.5747 1.0274 1.0084 7.3541
Matching 1.8549 0.9581 0.9342 1.8562 0.9606 0.9365 1.8587 0.9627 0.9384 1.8744 0.9708 0.9454
Opt. Matching 1.6860 1.1074 1.0909 4.2200 1.6866 1.1129 1.0964 4.1840 1.6871 1.1154 1.0987 4.1180 1.6924 1.1281 1.1108 4.0750
Pen. Synth. (BC) 1.1277 0.1949 0.0017 1.1315 0.1964 0.0203 1.1391 0.2010 0.0362 1.1785 0.2177 0.0358
Unpen. Synth. (BC) 1.1213 0.1956 0.0026 1.1238 0.1983 0.0304 1.1307 0.2064 0.0615 1.1720 0.2521 0.1494
Pure Synth. (BC) 1.1219 0.1954 0.0027 1.1244 0.1979 0.0290 1.1309 0.2055 0.0590 1.1701 0.2484 0.1437
Matching (BC) 1.4190 0.2073 0.0057 1.4214 0.2113 0.0381 1.4275 0.2202 0.0702 1.4629 0.2666 0.1596
Opt. Matching (BC) 1.1430 0.1809 0.0040 1.1455 0.1855 0.0382 1.1528 0.1963 0.0730 1.1874 0.2529 0.1708
 6. Empirical Application

Starting with LaLonde (1986), many studies have used data from the National Supported
Work Demonstration (NSW) to demonstrate the applicability and performance of alterna-
tive estimators of treatment effects (see, e.g., Dehejia and Wahba, 2002; Smith and Todd,
2005; Abadie and Imbens, 2011). This section provides an empirical application of penalized
synthetic control estimators using NSW data. The NSW program was aimed at improving
employment opportunities for individuals at the margins of the labor market by providing
them with temporary subsidized jobs. It targeted individuals with low levels of education
or criminal records, former drug addicts, and mothers who received welfare benefits for sev-
eral years. In this application, the quantity of interest is the impact of the participation in
the program on 1978 yearly earnings in dollars for this specific population. In the original
experiment, individuals from the targeted population were randomly split between a treat-
ment arm (n1 = 185) and a control arm (n0 = 260). On that sample, the ATET estimate
is $1,794, which provides an experimental benchmark. A second control group, extracted
from the Panel Study of Income Dynamics (PSID, n0 = 2, 490), has been made available to
study estimators based on observational data (LaLonde, 1986; Dehejia and Wahba, 1999).
We use this second sample to illustrate the penalized synthetic control estimator. We in-
clude in Xi the 10 covariates of the dataset (age, education, black, hispanic, married, no
degree, income in 1974, income in 1975, no earnings in 1974, no earnings in 1975). Before
computing estimates, we divide each of the predictors in Xi by their standard deviations in
the treated sample. For the 1974 and 1975 income variables, which feature long right tails
in their distributions, we first discard values above the .9 quantiles before computing the
standard deviation that we use to rescale these variables.
We compare four penalized synthetic control estimators: one with a fixed (and small)
value of λ (λ = .1), the pure synthetic control (λ → 0) and two with data-driven penalties
– one that minimizes individual RMSE, and one that minimizes the bias. The computation
of the pure synthetic control estimator is based on the result in Theorem 2 and discus-
sion thereafter. We also report the results from nearest neighbor matching estimators: a

28
one-match-with-replacement procedure, and one where the optimal number of neighbors
minimizes the individual RMSE. For the latter estimator, the number of matches is selected
over the set of positive integers not greater than 30. In order to select the penalty level λ
(and the number of neighbors in the matching procedure), we adapt the strategy proposed
in Section 4.1. In particular, to reduce the computational burden, we select 170 untreated
units that are close to the treated units, and for each one of them construct a synthetic unit
using all the other (2,489) untreated units. The goal is to select a penalty level in a setting as
close as possible to the one where the estimator is going to be applied. The set of these 170
“placebo-treated” untreated units is constructed as the union of all four nearest neighbors
of each treated unit. The selected penalty level λ
b is then used to compute the synthetic

units for each of the 185 treated, using the 2,490 untreated. In the PSID dataset, several un-
treated individuals are identical in terms of covariates, which can create ties in matches and
non-unique solutions in penalized synthetic control weights. To make the solution unique,
we consolidate identical untreated individuals into a single individual whose outcome is an
average of the outcomes of its components before computing the synthetic control weights.
Table 5 reports the results. All penalized synthetic control estimators, as well as the
one-match nearest neighbor matching, in columns (1) to (5), come fairly close to the treated
sample in terms of average values of the predictors. These four estimators yield treatment
effects ranging from $1,881 to $2,171, in the ballpark of the experimental benchmark, $1,794.
The matching procedure in column (6) selects 23 neighbors, and yields a matched sample
that is substantially different than the sample of treated units in the average value of the
predictors. The estimated treatment effect for this matching estimator is $983, markedly
smaller than the experimental benchmark. The synthetic control estimators in columns (1),
(3) and (4), which employ positive values of λ, are very sparse. In contrast, the synthetic
control estimator in column (2) employs λ = 0 and produces some highly dense estimates,
as evidenced by the value of its maximal density.
Figure 3 displays the results of permutation tests described in Section 3 for an estimator
with a fixed level λ = .1 – column (1) in Table 5. We consider two tests statistics: the sum

29
 of ranks and the aggregate treatment effect. p-values are computed using one-sided tests (no
effect vs. positive effect) and 1,000 permutations. In both cases, the effect lies at the right
tail of the distribution and is significant at the 1% level for the sum-of-rank statistics and
at 5% level for the aggregate treatment effect statistics.
Figure 3: Permutation tests for the NSW data

100

90

75

60

50

30
25

0 0

1.5e+07 1.6e+07 1.7e+07 1.8e+07 1.9e+07 −2000 0 2000

Sum of ranks Aggregate Treatment Effect

Note: Results are obtained for 1,000 permutations, using a fixed value of λ = .1. The left panel displays the
histogram of the sum-of-ranks statistics. The right panel displays the histogram of the aggregate treatment
effect. The red doted line is the value of the statistics for the observed assignment. p-values for the one
sided tests are 0.003 and 0.013 respectively.

7. Conclusions

In this article, we propose a penalized synthetic control estimator that trades-off pairwise
matching discrepancies with respect to the characteristics of each unit in the synthetic control
against matching discrepancies with respect to the characteristics of the synthetic control
unit as a whole. We study the properties of this estimator and propose data-driven choices
of the penalization parameter. We show that the penalized synthetic control estimator is
unique and sparse, which makes it particularly convenient for empirical applications with
multiple treated units, where the focus of the analysis may be on average treatment effects.
We propose a bias-correction for the penalized synthetic control estimator, and extend the
inferential methods for synthetic controls in Abadie et al. (2010) to settings with multiple

30
 Table 5: LaLonde (1986) dataset, Results

Treated Untreated Untreated Pen. Synth. Pen. Synth. Pen. Synth. Pure Synth. Matching Matching
(Experimental) (PSID) λf ixed λ
bRM SE λ
bbias λ→0 M =1 M
cRM SE
(1) (2) (3) (4) (5) (6)
Married 0.19 0.15 0.87 0.20 0.20 0.21 0.20 0.21 0.44
Black 0.84 0.83 0.25 0.82 0.81 0.83 0.81 0.84 0.68
Hispanic 0.06 0.11 0.03 0.05 0.05 0.05 0.05 0.05 0.02
No Degree 0.71 0.83 0.31 0.67 0.66 0.67 0.66 0.68 0.61
No Earnings 1974 0.71 0.75 0.09 0.67 0.68 0.67 0.68 0.69 0.54
No Earnings 1975 0.60 0.68 0.10 0.63 0.63 0.63 0.63 0.63 0.62
Age 25.8 25.1 34.9 27.2 27.4 26.6 27.4 26.3 29.5
Education 10.4 10.1 12.1 10.4 10.3 10.4 10.3 10.3 10.6
Earnings 1974 2,095.6 2,107.0 19,428.8 2,225.1 2,256.6 2,235.6 2,256.6 2,275.4 2,421.4

31
Earnings 1975 1,532.1 1,266.9 19,063.3 1,602.4 1,619.6 1,582.4 1,619.6 1,598.7 1,733.2
Treatment Effect 1,794.3 1,977.3 2,171.4 1,881.4 2,171.5 2,138.8 982.9
λ 0.1 0 0.95 →0
Min. Density 1 2 1 1 1 23
Median Density 4 6 2 6 1 23
Max. Density 8 1,021 6 11 2 26
Active Units 260 2490 193 1,664 124 257 67 511
Note : The upper part of the table displays the average characteristics of the individuals in the sample. For the columns “Pen. Synth.” (resp.
“Matching”), it is an average weighted by the synthetic control (resp. matching) weights. The lower part of the table displays the resulting treatment
effect, the corresponding value of λ and statistics regarding the weights. Here, density counts the number of non-zero elements in a vector of synthetic
weights, Wi∗ (λ). The median, min. and max. density rows report the median, minimal and maximal densities observed for a synthetic unit. “Active
units” refers to untreated units who receive a non-zero weight in at least one synthetic unit. Notice that for matching estimators the density can
differ from the chosen number of neighbors when there are ties.
treated units. We show that the penalized synthetic control estimator performs well in
simulations. The inferential methods proposed in this article are conditional on the sample
and do not rely on the sampling mechanism. Sampling-based inference, which requires an
approximation to sampling distribution of the penalized synthetic control estimator, is an
interesting avenue for future research.

32
 Appendix

Notation
For any real matrix X, let CH(X) and DT (X) be the convex hull and the Delaunay tesselation of
the columns of X, respectively. We recall that DT (X) is a partition of CH(X).
Proof of Lemma 1
Notice that if the first result in Lemma 1 does not hold, then Wi∗ (λ) cannot be a solution to the
problem in equation (5) since the nearest neighbor will result in a better fit. We start by proving
the upper bound in the second inequality. Since Wi∗ (λ) minimizes (5), it follows that

(Xi − X0 Wi∗ (λ))0 (Xi − X0 Wi∗ (λ)) + λ∆0i Wi∗ (λ) ≤ (Xi − XNNi )0 (Xi − XNNi ) + λ∆N
i
N
.

Therefore,
λ∆0i Wi∗ (λ) ≤ (1 + λ)∆N
i
N
,
and the result follows from λ > 0. The lower bound follows from the definition of ∆N
i
N. 
Proof of Theorem 1
Without loss of generality, consider the case with only one treated, n1 = 1. The penalized synthetic
control estimator is calculated as the vector of weights that solves

min fλ (W ) = (X1 − X0 W )0 (X1 − X0 W ) + λW 0 ∆1 ,

W
s.t. W ∈ W, (A.1)

where W = {W ∈ [0, 1]n0 | W 0 1n0 = 1}. It is easy to check that the feasible set, W, is convex
and compact. Because fλ is continuous and W is compact, it follows that the function attains a
minimum on W. Moreover, X00 X0 is positive semi-definite, so fλ is convex.
Suppose that more than one solution exists. In particular, assume that W1 and W2 are solutions,
with fλ (W1 ) = fλ (W2 ) = fλ∗ . Then, for any a ∈ (0, 1) we have that aW1 + (1 − a)W2 ∈ W. Because
fλ is convex, we obtain

fλ (aW1 + (1 − a)W2 ) ≤ afλ (W1 ) + (1 − a)fλ (W2 ) = fλ∗ .

This implies that the problem has either a unique solution or infinitely many. In addition, if there
are multiple solutions, they all produce the same fitted values X0 W . To prove this, suppose there
are two solutions W1 and W2 such that X0 W1 6= X0 W2 . Then, because kx − ck2 is strictly convex
in c, for a ∈ (0, 1) we obtain

fλ (aW1 + (1 − a)W2 ) = kX1 − X0 (aW1 + (1 − a)W2 )k2 + λ(aW1 + (1 − a)W2 )0 ∆1

< akX1 − X0 W1 k2 + (1 − a)kX1 − X0 W2 k2 + λ(aW1 + (1 − a)W2 )0 ∆1
= afλ∗ + (1 − a)fλ∗
= fλ∗ ,

which contradicts that W1 and W2 are solutions. As a result, if W1 and W2 are solutions, then
X0 W1 = X0 W2 . Moreover, λ > 0 implies W10 ∆1 = W20 ∆1 . Let A = [X00 1n0 ∆1 ]. It follows that, if
W1 and W2 are solutions, then A0 (W1 − W2 ) = 0 p+2 (where 0 p+2 is a (p + 2) × 1 vector of zeros).

33
Karush-Kuhn-Tucker conditions imply:
λ
Xj0 (X1 − X0 W ) − ∆1,j = π − γj
2
Wj ≥ 0, W 0 1n0 = 1, γj ≥ 0, γj Wj = 0.

Stacking the first n0 conditions above and pre-multiplying by W 0 , we obtain

λ 0
W 0 X00 (X1 − X0 W ) − W ∆1 = π.
2
From this equation, it follows that the value of π is unique across solutions, because X00 W and
W 0 ∆1 are unique across solutions. Given that π is unique, the equations
λ
Xj0 (X1 − X0 W ) − ∆1,j = π − γj .
2
imply that the γj ’s are unique across solutions. Let X e0 be the submatrix of X0 formed by the
columns associated with zero γj ’s, and define W f, ∆e 1 , and 1ne analogously, where n
0
e0 is the number
of columns of X0 . Then,
e
e00 (X1 − X
X f) = λ ∆
e0 W e 1 + π1ne . (A.2)
0
2
Notice that if λ > 0, then kX1 − X0 W k = 0 implies that ∆ e 1 is a constant vector. We therefore
obtain that if λ > 0 and ∆1 is not constant, then it must be the case that kX1 − X0 W k > 0.
e
Let Ae = [Xe 0 1ne ∆ e 1 ]. Consider the case ne0 ≥ p + 2. In this case A e has full column rank, which
0 0
implies that equation (A.2) cannot hold if λ > 0. As a result, when λ > 0, the solution to (A.1)
has p + 1 non-zero components at most.
Consider now the case n e0 ≤ p + 1. For this case A
e has full row rank. Moreover, if W f1 and W f2 are
solutions, it must be the case that A 0 f1 − W
e (W f2 ) = 0 p+2 . However, because A e has full row rank
0
the system A z = 0 p+2 admits only the trivial solution, z = 0 ne0 , which implies that the solution
e
to (A.1) is unique. 

Lemma A.1 (Optimality of Delaunay for the Compound Discrepancy, Rajan, 1994) Let
Z ∈ CH(X0 ). Consider a solution W
f = (W fn )0 of the problem
fn +1 , . . . , W
1

n
X
min Wj kXj − Zk2 , (A.3)
W ∈[0,1]n0
j=n1 +1
n
X
s.t. X0 W = Z, Wj = 1. (A.4)
j=n1 +1

Then, non-zero values of W

fj occur only among the vertices of the Delaunay simplex containing Z.

We restate the proof of Lemma 10 in Rajan (1994) for clarity and note that it does not rely on
general quadratic position of the set of points.
Proof of Lemma A.1

34
For a point X ∈ Rp , consider the transformation φ : X → (X, kXk2 ). The images under φ P of points
in R belong to the paraboloid of revolution P with vertical axis and equation xp+1 = pi=1 x2i .
p

By Theorem 17.3.1 in Boissonnat and Yvinec (1998), the Delaunay tessellation of the convex hull
of the n0 points Xn1 +1 , . . . , Xn in Rp are obtained by projecting onto Rp the faces of the lower
envelope of the convex hull of the n0 points φ(Xn1 +1 ), . . . , φ(Xn ) obtained by lifting the Xj ’s onto
the paraboloid P.
P 
n Pn 2
Now consider points j=n1 +1 Wj Xj , j=n1 +1 Wj kXj k subject to the constraints in (A.4).
 P 
n
These points are equal to Z, j=n1 +1 Wj kXj − Zk2 + kZk2 and belong to the convex hull of
φ(Xn1 +1 ), . . . , φ(Xn ). Hence, a solution of (A.3) for a fixed Z is given by such a point with the lowest
(p + 1)-th coordinate. It is a point on the lower envelope of the convex hull of φ(Xn1 +1 ), . . . , φ(Xn ),
so Z belongs to a p-simplex of the Delaunay tessellation. As a consequence, non-zero entries of
W
f occur only among the vertices of the face of the Delaunay tessellation of the columns of X0
containing Z. 
Proof of Theorem 2
It is enough to prove that the result holds for one treated unit, so we consider the case n1 = 1
and drop the treated units subscripts from the notation. We proceed by contradiction. Suppose
that the synthetic control weights are given by the vector W ∗ (λ) = (W2∗ (λ), . . . , Wn∗ (λ))0 , and that
Wj∗ (λ) > 0 for j which is not a vertex of the Delaunay simplex in DT (X0 ) containing X0 W ∗ (λ).
Because X0 W ∗ (λ) ∈ CH(X0 ), it follows from Lemma A.1 that we can always choose a n0 -vector
of weights W f = X0 W ∗ (λ), (ii) Pn W
f ∈ [0, 1]n0 , such that (i) X0 W fj = 1, (iii) W fj = 0 for any
j=2
j that is not a vertex of the Delaunay simplex containing X0 W ∗ (λ), and (iv) W f induces a lower
compound discrepancy than W ∗ (λ) relative to X0 W f = X0 W ∗ (λ),

n
X n
X
f (λ)k2 <
fj kXj − X0 W
W Wj∗ (λ)kXj − X0 W ∗ (λ)k2 . (A.5)
j=2 j=2

For any W ∈ [0, 1]n0 it can be easily seen that

n
X n
X
Wj kXj − X1 k2 = Wj kXj − X0 W k2 + kX1 − X0 W k2 . (A.6)
j=2 j=2

f k2 = kX1 − X0 W ∗ (λ)k2 , we
Combining equations (A.5) and (A.6) with the fact that kX1 − X0 W
obtain
Xn Xn
2
Wj kXj − X1 k <
f Wj∗ (λ)kXj − X1 k2 .
j=2 j=2

As a result
n
X n
X
f k2 + λ
kX1 − X0 W fj kXj − X1 k2 < kX1 − X0 W ∗ (λ)k2 + λ
W Wj∗ (λ)kXj − X1 k2 ,
j=2 j=2

which contradicts the premise that W ∗ (λ) is a solution to (5). 

Proof of Theorem 3

35
Since the columns of [X1 X0 ] are in general quadratic position, the augmented Delaunay tri-
angulation, DT ([X1 : X0 ]), exists and is unique. Without loss of generality, consider the case
of a single treated unit, and normalize X1 to be at the origin. Let UT (X1 , X0 ) be the union
of the Delaunay simplices that have X1 as a vertex in DT ([X1 : X0 ]). Consider a point z ∈
CH([X1 : X0 ])\UT (X1 , X0 ). We will first show that z cannot be equal to X0 W ∗ (λ). Because z
does not belong to UT (X1 , X0 ) and because the set CH([X1 : X0 ]) is convex, it is always possible
to find a point v ∈ CH([X1 : X0 ])\UT (X1 , X0 ) on the line segment that connects z and X1 such
that kX1 − vk < kX1 − zk (or, equivalently, kvk < kzk). For any point inPx ∈ CH([X1 : X0 ])
n
consider
Pn the set of non-negative weights, w1 (x), . . . , wn (x), such that: (i) i=1 wi (x) = 1, (ii)
i=1 w i (x)X i = x, and (iii) if Xi is not a vertex of the Delaunay simplex containing x, then
wi (x) = 0. If x ∈ CH([X1 : X0 ])\UT (X1 , X0 ), then the Delaunay simplex containing x in
DT ([X1 : X0 ]) is the same as the the Delaunay simplex contain x in DT (X0 ) (Devillers and Teil-
laud, 2003; Boissonnat et al., 2009). Therefore, by Theorem 2, if x ∈ CH([X1P : X0 ])\UT (X1 , X0 )
∗ ∗ 0 n 2
and
Pn X0 W (λ) = x, 2then W (λ) = (w2 (x), . . . , wn (x)) . Now, let g(x) = i=1 wi (x)kXi k =
i=1 wi (x)kX1 − Xi k . This function is convex because it is the lower boundary of the convex hull
of {(X1 , kX1 k2 ), . . . , (Xn , kXn k2 )} (Rajan, 1994), and is minimized at x = X1 . As we move from z
to v, we travel in the direction of the minimum of g(x). Because g(x) is a convex function, it follows
that g(v) < g(z). Because kvk < kzk and g(v) < g(z), it follows that X0 W ∗ (λ) 6= z, regardless of
the value of λ. This implies that X0 W ∗ (λ) must belong to UT (X1 , X0 ) and the result follows from
Theorem 2. 

36
 References
Abadie, A. (2020). Using synthetic controls: Feasibility, data requirements, and methodolog-
ical aspects. Journal of Economic Literature. Forthcoming. Available online at https:
//www.aeaweb.org/articles/pdf/doi/10.1257/jel.20191450.

Abadie, A., Diamond, A., and Hainmueller, J. (2010). Synthetic control methods for comparative
case studies: Estimating the effect of California’s tobacco control program. Journal of the
American Statistical Association, 105(490):493–505.

Abadie, A., Diamond, A., and Hainmueller, J. (2015). Comparative politics and the synthetic
control method. American Journal of Political Science, 59(2):495–510.

Abadie, A. and Gardeazabal, J. (2003). The Economic Costs of Conflict: A Case Study of the
Basque Country. American Economic Review, 93(1):113–132.

Abadie, A. and Imbens, G. W. (2006). Large sample properties of matching estimators for average
treatment effects. Econometrica, 74(1):235–267.

Abadie, A. and Imbens, G. W. (2011). Bias-corrected matching estimators for average treatment
effects. Journal of Business & Economic Statistics, 29(1):1–11.

Acemoglu, D., Johnson, S., Kermani, A., Kwak, J., and Mitton, T. (2016). The value of connections
in turbulent times: Evidence from the united states. Journal of Financial Economics, 121:368–
391.

Amjad, M., Shah, D., and Shen, D. (2018). Robust synthetic control. Journal of Machine Learning
Research, 19(22):1–51.

Angrist, J. D. and Pischke, J.-S. (2008). Mostly Harmless Econometrics: An Empiricist’s Com-
panion. Princeton University Press, Princeton, NJ.

Arkhangelsky, D., Athey, S., Hirshberg, D. A., Imbens, G. W., and Wager, S. (2018). Synthetic
Difference in Differences. arXiv:1812.09970.

Athey, S., Bayati, M., Doudchenko, N., Imbens, G., and Khosravi, K. (2021). Matrix completion
methods for causal panel data models. Journal of the American Statistical Association, 0(0):1–15.

Athey, S. and Imbens, G. (2016). Recursive partitioning for heterogeneous causal effects. Proceed-
ings of the National Academy of Sciences, 113(7):7353–7360.

Athey, S. and Imbens, G. W. (2021). Design-based analysis in difference-in-differences settings with

staggered adoption. Journal of Econometrics.

Ben-Michael, E., Feller, A., and Rothstein, J. (2019). Synthetic controls and weighted event studies
with staggered adoption.

Ben-Michael, E., Feller, A., and Rothstein, J. (2021). The augmented synthetic control method.
Journal of the American Statistical Association, 0(ja):1–34.

37
Bohn, S., Lofstrom, M., and Raphael, S. (2014). Did the 2007 Legal Arizona Workers Act reduce the
state’s unauthorized immigrant population? Review of Economics and Statistics, 96(2):258–269.

Boissonnat, J.-D., Devillers, O., and Hornus, S. (2009). Incremental construction of the Delaunay
graph in medium dimension. In Annual Symposium on Computational Geometry, pages 208–216,
Aarhus, Denmark.

Boissonnat, J.-D. and Yvinec, M. (1998). Algorithmic Geometry. Cambridge University Press, New
York, NY, USA.

Cattaneo, M. D., Feng, Y., and Titiunik, R. (2019). Prediction intervals for synthetic control
methods. arXiv:1912.07120.

Chernozhukov, V., Wüthrich, K., and Zhu, Y. (2021). An exact and robust conformal inference
method for counterfactual and synthetic controls. Journal of the American Statistical Associa-
tion, 0(ja):1–44.

Cunningham, S. and Shah, M. (2018). Decriminalizing indoor prostitution: Implications for sexual
violence and public health. The Review of Economic Studies, 85(3):1683–1715.

Dehejia, R. and Wahba, S. (2002). Propensity score-matching methods for nonexperimental causal
studies. The Review of Economics and Statistics, 84(1):151–161.

Dehejia, R. H. and Wahba, S. (1999). Causal effects in nonexperimental studies: Reevaluating the
evaluation of training programs. Journal of the American Statistical Association, 94(448):1053–
1062.

Devillers, O. and Teillaud, M. (2003). Perturbations and vertex removal in a 3D Delaunay tri-
angulation. In 14th ACM-Siam Symposium on Discrete Algorithms (SODA), pages 313–319,
Baltimore, MA, United States.

Doudchenko, N. and Imbens, G. W. (2016). Balancing, regression, difference-in-differences and

synthetic control methods: A synthesis. NBER Working Papers, 22791.

Dube, A. and Zipperer, B. (2015). Pooling Multiple Case Studies Using Synthetic Controls: An
Application to Minimum Wage Policies. IZA Discussion Papers 8944, Institute for the Study of
Labor (IZA).

Firpo, S. and Possebom, V. (2018). Synthetic Control Method: Inference, Sensitivity Analysis and
Confidence Sets. Journal of Causal Inference, 6(2):1–26.

Gobillon, L. and Magnac, T. (2016). Regional policy evaluation: Interactive fixed effects and
synthetic controls. Review of Economics and Statistics, 98(3):535–551.

Hackmann, M. B., Kolstad, J. T., and Kowalski, A. E. (2015). Adverse selection and an individual
mandate: When theory meets practice. American Economic Review, 105(3):1030–1066.

Hirano, K., Imbens, G. W., and Ridder, G. (2003). Efficient estimation of average treatment effects
using the estimated propensity score. Econometrica, 71(4):1161–1189.

38
Imbens, G. W. and Rubin, D. B. (2015). Causal Inference for Statistics, Social, and Biomedical
Sciences: An Introduction. Cambridge University Press, New York, NY, USA.

Kellogg, M., Mogstad, M., Pouliot, G. A., and Torgovitsky, A. (2020). Combining matching and
synthetic control to trade off biases from extrapolation and interpolation. Available online at
https://a-torgovitsky.github.io/.

Kleven, H. J., Landais, C., and Saez, E. (2013). Taxation and international migration of superstars:
Evidence from the European football market. American Economic Review, 103(5):1892–1924.

Kreif, N., Grieve, R., Hangartner, D., Turner, A. J., Nikolova, S., and Sutton, M. (2016). Exami-
nation of the synthetic control method for evaluating health policies with multiple treated units.
Health Economics, 25(12):1514–1528.

LaLonde, R. J. (1986). Evaluating the econometric evaluations of training programs with experi-
mental data. The American Economic Review, 76(4):604–620.

Okabe, A., Boots, B., Sugihara, K., and Chiu, S. N. (2000). Spatial Tessellations: Concepts and
Applications of Voronoi Diagrams. Series in Probability and Statistics. John Wiley and Sons,
Inc.

Rajan, V. T. (1994). Optimality of the Delaunay triangulation in Rd . Discrete & Computational

Geometry, 12(2):189–202.

Rosenbaum, P. and Rubin, D. B. (1983). The central role of the propensity score in observational
studies for causal effects. Biometrika, 70(1):41–55.

Rubin, D. B. (1973). The use of matched sampling and regression adjustment to remove bias in
observational studies. Biometrics, 29(1):185–203.

Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized

studies. Journal of Educational Psychology, 66(5):688.

Smith, J. and Todd, P. (2005). Does matching overcome LaLonde’s critique of nonexperimental
estimators? Journal of Econometrics, 125(1-2):305–353.

39