The Quarterly Review of Economics and Finance 81 (2021) 70–81

Contents lists available at ScienceDirect

The Quarterly Review of Economics and Finance

journal homepage: www.elsevier.com/locate/qref

An empirical comparison between a regression framework and the

Synthetic Control Method
Orkideh Gharehgozli
Department of Economics, Feliciano School of Business, Montclair State University, 1 Normal Ave, New Jersey, NJ, United States

a r t i c l e i n f o a b s t r a c t

Article history: The Synthetic Control Method has been used in comparative case studies in which the existence of a
Received 9 August 2020 counterfactual unit with a high level of similarities and comparability is crucial. On the other hand, while
Received in revised form 28 April 2021 many methods have been developed to enhance our estimation power, not many studies have explored
Accepted 8 May 2021
the prediction power of the traditional regression frameworks in such comparative case studies. In this
Available online 11 May 2021
paper, we empirically compare the Synthetic Control Method with a Dynamic Panel Data Regression
Framework. We compare the estimation result and the prediction power of the predicted unit driven
JEL classification:
from the Dynamic Panel Data model and the counterfactual unit from the SCM. To apply the idea, we
C1
C23
employ the recent sanctions on Iran as a suitable case of policy intervention and a comparative case study.
C5 © 2021 Board of Trustees of the University of Illinois. Published by Elsevier Inc. All rights reserved.
F5
F4

Keywords:
Synthetic Control Method
Dynamic Panel Data
Treatment effect
Counterfactual
Comparative case study

1. Introduction variables, their performance is less evaluated in comparative case

studies when there exist an intervention and irregular behavior
In case of policy intervention, treatment, or an event, ideally, the of variables. In this paper, we apply and compare the Synthetic
treatment effect would be the difference of the outcome variable Control Method with a Dynamic Panel Data Regression Framework
values between the treated units and the same units under no treat- (DPRF) to estimate the effect of an intervention in a comparative
ment (Rubin, 1974, 1977, 1978). But “the fundamental problem case study: sanctions on Iran. A Dynamic Panel Data model would
of causal inference” (Holland, 1986) is the unknown and unob- be a suitable model to study the effect of policy interventions on the
served values of the outcome variable for the treated units under macroeconomic variables such as economic growth in aggregate
no treatment. The Synthetic Control Method (SCM) uses a weighted entities such as countries.
average of potential control units to provide a Synthetic Control In randomized experiments, the random assignment assures
unit that behaves more closely to the affected unit in terms of that the treated units and controls are “balanced” or “matched” in
related predictors and the outcome variable of interest. SCM has the covariates and assures that the characteristics of the units are
been very popular recently and has worked as a practical tool to not affecting the probability of treatment assignment. Therefore, a
answer important and interesting research questions. However, comparison between the treated and control units are very infor-
there are some caveats to be noted with this method such as the fact mative and can be interpreted as the treatment effect. In contrast,
that it only provides a point estimate and it is unable to approxi- in an observational study, random assignment is not an option.
mate the sampling distributions of test statistics (Abadie, 2019). In the absence of randomization, we need to assume “some
On the other hand, although traditional regression frameworks are form of exogeneity” (Imbens, 2002) of the treatment assignment. It
widely used to explore the relationship between macroeconomic is only under this assumption that we can compare the outcome
variable between the treated units and controls, and meaning-
fully attribute the difference to the treatment. As Imbens (2002)
summarizes, different studies in the literature refer to this notion
E-mail address: [email protected]

https://doi.org/10.1016/j.qref.2021.05.002
1062-9769/© 2021 Board of Trustees of the University of Illinois. Published by Elsevier Inc. All rights reserved.
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

with different names (such as unconfoundedness (Rosenbaum and relax some of the assumptions in the Hsiao et al. (2012) method
Rubin, 1983), the selection on observable (Barnow et al., 1980; and discuss the asymptotic properties of their suggested estima-
Fitzgerald et al., 1998), or conditional independence (Lechner, tor.
1999), and many studies explore unbiased estimators for the In this paper, the question under study is whether a more fun-
treatment effect. Proposed methods can be summarized into few damental regression framework (such as a Dynamic Panel Data
categories such as traditional regression of the outcome of interest Regression Framework) performs as well as the Synthetic Control
on the covariates (Rubin, 1977), matching on covariates (Abadie Method in creating the predicted unit and estimating the effects.
et al., 2010, 2015a; Card and Krueger, 1994), propensity score If that is the case, we believe the regression framework could be
(Dehejia and Wahbe, 2002; Dehejia and Wahba, 1999; Heckman more applicable because not only does it provide the estimates of
et al., 1997), and a combination of these approaches (Abadie and the effects, it also reveals more information about the point esti-
Imbens, 2002). mates. We have more coherent assumptions about the asymptotic
Matching methods are one strand of the methods developed distribution of the estimators; we have estimates for the standard
to provide the essential similarities between treated and control errors, and the method is more sensitive to the choice of variables
units (Rosenbaum, 2005). In large-sample studies, matching meth- and is less arbitrary to the choice of the donor units. In Section 2,
ods aim to equate (or “balance”) the distribution of covariates in we describe the data and resources. Section 3 provides empirical
the treated and control units. In small-sample social comparative analysis, including a revisit of the Synthetic Control Method esti-
studies, where the interventions affect aggregate entities such as mation, and the Dynamic Panel Data model estimation of the effect
countries or states, it is often very difficult to find suitable con- of multinational sanctions on Iran’s GDP. We compare the “fit” of
trols that are unaffected by the intervention and also have similar the Synthetic Control Method with the prediction power of the esti-
characteristics to those of the affected unit in the pre-intervention mated unit driven from the DPRF at the end. Section 3.4 provides
period (Lijphart, 1971; Collier, 1993; Abadie et al., 2010). Despite unit-root and co-integration tests, and Section 4 concludes.
this fact, for example, in a well-known study, Card and Krueger
(1994) use “diff-in-diff” method to explore the effect of the 1992 2. Sample and data
change in minimum wages on the unemployment rate in New
Jersey. They surveyed 410 fast-food restaurants in New Jersey and Our empirical analysis is based on annual country-level panel
eastern Pennsylvania (which shows high similarities to New Jersey) data for the period 1980–2015 (in 2015, the multinational agree-
before and after the rise. Comparisons of employment growth at ment to remove the sanctions (Iran Deal) was signed setting in
stores in New Jersey and Pennsylvania provide simple estimates of motion the loosening of sanctions). The international sanctions
the effect of the higher minimum wage. against Iran were imposed in 2011 which yields a pre-intervention
Instead of using a single control unit, the Synthetic Control period of more than 30 years in our analysis. Our control pool (donor
Method (Abadie and Gardeazabal, 2003; Abadie et al., 2010, 2015a) pool in the synthetic Control Method) includes eight OPEC mem-
uses a weighted average of potential control units. For example, ber countries: Algeria, Ecuador, Kuwait, Libya, Nigeria, Qatar, Saudi
Abadie and Gardeazabal (2003) develop a synthetic unit control as Arabia, and the United Arab Emirates.1 Also, to increase the size of
a weighted average of two other Spanish regions to find the effect the pool, we add countries from major non-OPEC oil producer coun-
of an outbreak of terrorism in Basque Country in the late 1960s, tries (i.e. Canada and China) as well as the rest of non-OPEC Iran’s
on the economic growth of this region. Abadie et al. (2010) study neighbors with close economic similarities (i.e. Oman, Bahrain,
the effect of California’s tobacco control program (Proposition 99) and Turkey). The variables used in our analysis are listed in the
on cigarette consumption using the Synthetic Control Method. In Data Appendix along with descriptions and data sources. The out-
another study, Abadie et al. (2015a) implement the Synthetic Con- come variable of interest, Yjt is the log of real GDP for country j at
trol Method to study the effect of German reunification in 1990 time t. We also use GDP growth as the outcome variable of inter-
on west Germany’s GDP. Some other recent applications of the est in some of the estimations. GDP is Purchasing Power Parity
SCM can be found in Donohue et al. (2019), Cunningham and Shah (PPP)-adjusted and measured in constant 2011 international dol-
(2018), Peri and Yasenov (2019), Allegretto et al. (2017) and Borjas lars. Because our control countries are heavily dependent on rents
(2017). from natural resources, for the pre-sanction predictors, we rely on
The Synthetic Control Method is more of an optimization prob- a standard set of economic growth indicators for these countries.
lem that aims to minimize the discrepancy between the synthetic
counterfactual unit and the actual treated unit in terms of the out- 3. Empirical analysis: synthetic control method, and
come variable over a specific period. It may potentially ignore the dynamic panel data
economic principles of causal relationships. In contrast to a regres-
sion framework, the Synthetic Control Method does not impose 3.1. The Synthetic Control Method
fundamental restrictions on the choice of the variables, or the
donor units. Hsiao et al. (2012) propose a Panel Data model to Suppose there are unit i = 1, . . ., J + 1 and t = 1, . . ., T0 , . . ., T ,
assess the impact of a policy intervention and demonstrate that the where T0 is the last period before the intervention. yit N is the out-
dependence among cross-sectional units (through a Factor Model come that would be observed for unit i at time t in the absence of
framework) can be utilized to construct the counterfactual. They the intervention, while yit I is the outcome that would be observed
suggest if the outcome variable is generated from a Factor Model for unit i = 1, . . ., J + 1 and t = 1, . . ., T0 , . . ., T . So, for t = 1, . . ., T0
with unobservable factors and factor loadings, other units’ outcome and all i = 1, . . ., J + 1, we have that yit N = yit I .
variables can be used in lieu of the unobservable factors and factor Following Abadie et al. (2010), let ˛it = yitI − yitN be the effect of
loadings. In a more recent study, Wan et al. (2018) study and com- the intervention for unit i at time t. We assume unit i = 1 is exposed
pare the 2 methods in their underlying assumptions and through a to the intervention for t = T0 + 1, . . ., T , therefore we want to esti-
series of simulations and they show that the Panel Data approach mate ˛1t = y1tI − yN for t = T + 1, . . ., T . Note that yI is observed,
1t 0 1t
dominates SCM in a majority of cases. In contrast, Gardeazabal
and Vega-Bayo (2017) compare the SCM and Panel Data approach
working with real data and conclude that when more preinter- 1
We left Venezuela and Iraq out of the pool due to economic fluctuations in these
vention periods and covariates are available and we have a good countries during the period of the analysis. We also left Angola out due to data
match in SCM, it performs better than the PDA. Li and Bell (2017) limitations.

71
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

Fig. 1. Real GDP: Iran vs. Synthetic Iran.

Table 1 Table 2
Donor pool countries and share of each in the construction of the Synthetic Iran. SCM estimation result.

County Weight Country Weight Year Actual Value SC Value The Effect

Algeria 0.138 Nigeria 0.002 GDP $B

Bahrain 0.000 Oman 0.001 2012 1260.32 1432.03 −171.71 (12%)
Canada 0.268 Qatar 0.002 2013 1236.23 1497.28 −261.05 (17.4%)
China 0.027 Saudi Arabia 0.112 2014 1289.92 1561.23 −271.31 (17.3%)
Ecuador 0.000 Turkey 0.194 2015 1270.57 1683.48 −412.91 (24.5%)
Kuwait 0.001 UAE 0.254 Growth %
Libya 0.001 2012 −7.44 4.38 −11.83
2013 −0.19 4.81 −5.00
2014 4.60 4.19 0.42
so we just need to estimate y1t N . If we assume X and X to be a 2015 −1.5 3.89 −5.39
1 0
vector and matrix collecting pre-intervention values of the pre- Note that “The Effect” is calculated as the gap/discrepancy between the actual values
dictors for unit 1 and the rest, then the Synthetic Control Method and the synthetic control values. The Effect = Actual Value − SC Value.
N
suggests y1T can be measured by a weighted average of the
+1
J+10 N
donors: wj yjT where the weights are obtained by minimiz-
 j=2 0 +1 Because we also include Iran’s growth rate (in addition to level-
ing (X1 − X0 W )V (X1 − X0 W ); and V is a positive semi-definite GDP) in our DPRF later in this section, here we expand Gharehgozli
matrix weighting the predictors. (2017) and perform the SCM for Iran’s growth rate as well. Iran’s
In simple words, the result of the SCM; if the weights are suc- annual GDP growth rates are positive for more than a decade before
cessfully found; is a control unit with values for the outcome the sanctions. However, after the sanctions of 2011, the growth rate
variable (GDP in this case) that would “match” the actual values falls to −7.44% in 2012 and −0.19% in 2013, it goes up to 4.6% in
of the outcome variable very well and for a long period pre- 2014 and falls again to −1.5 in 2015. We run the Synthetic Control
intervention. Therefore, using the same weights we can assume Method to estimate the figures for Iran’s growth rate for 2012, 2013,
the post-intervention behavior of the outcome variable would fol- 2014, 2015. The actual values and the synthetic values are reported
low what the synthetic unit suggests. First, we revisit the result of in Table 2. The actual growth rate of Iran in 2012 drops to −7.44%,
the Synthetic Control Method regarding the case under study by which according to our result this rate is −11.82% lower than what
Gharehgozli (2017). The study uses SCM to estimate the effect of it would be if there had been no sanctions. It means under no sanc-
intensification of sanctions in 2011 on Iran’s GDP. tions scenario, Iran’s growth rate would have been 4.38% in 2012.
Table 1 provides the list of the donor countries and the share of In 2013, the actual growth suffers a drop of −0.19%, while accord-
each in the construction of Synthetic Iran. Iran’s counterfactual is ing to Synthetic Iran, the growth would be 4.81% had there been
best reproduced by a weighted average of Canada, United Arab Emi- no sanctions. Therefore, we estimate the growth to be −5.00 lower
rates, Turkey, Algeria, Saudi Arabia, and China. The share of other than the no sanction scenario. In 2014 the growth would be 0.42%
countries in the pool is either zero or very small. Canada has the lower than the actual value of 4.6%. Iran’s growth in 2015 was −1.5%
highest weight followed by UAE. Fig. 1 displays the paths of the real which we estimate to be 5.39% lower than what it would have been
GDP of Iran and Synthetic Iran from 1995 to 2015. Synthetic Iran under no sanctions. It means the growth rate in 2015 would have
closely resembles Iran’s GDP over the pre-sanction period. The esti- been 3.89%.
mate of the effect of international sanctions imposed in 2011 is the In a recent study, Abadie (2019) summarizes the feasibility,
difference between the GDP of actual Iran and Synthetic Iran from data requirements, contextual requirements, and methodological
2011 to 2015 period. The discrepancy between the two after 2011, issues related to the empirical application of synthetic controls.
suggests a large negative effect of the sanctions on the country’s The paper provides a review discussion regarding inference with
GDP. SCM. It reviews the potential “placebo studies” exercises (and
Fig. 1 shows that while the GDP of Synthetic Iran grows, the GDP how we can use the distribution of the post/pre error ratio to
of actual Iran drops notably after 2011 with the gap between the strengthen such exercises). Initially, Abadie et al. (2010) proposed
two growing in magnitude. As summarized in Table 2, Iran’s GDP in this mode of inference for the Synthetic Control Method that is
2015 was 1270.57 billion dollars, which is estimated to be 412.91 based on permutation methods. The effect on the intervention is
billion dollars less than the value it would have been had there been estimated separately for each of the units in the sample. Then,
no sanctions imposed in or after 2011. This is equal to a 24.5% drop the effect of the treatment on the unit affected by the interven-
in GDP over the course of four years of heavy sanctions. Relative to tion is deemed to be significant when its magnitude is an outlier
the Synthetic Iran bench-mark, Iran’s GDP was reduced by 12.0% in relative to the permutation distribution (for Iran’s case, placebo
the first year after the sanctions (2012). studies can be found in Gharehgozli (2017)). One important caveat

72
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

Table 3
Dynamic Panel Data model estimation result.

(FE) lgdp (FD) lgdp (AB) lgdp (AB) lgdp (BB) lgdp (BB) lgdp

Iran 2012 −0.127*** (0.0180) −0.0948*** (0.0146) −0.128*** (0.0152) −0.102*** (0.00682) −0.128*** (0.0140) −0.113***(0.00952)

Iran 2013 −0.0726*** (0.0172) −0.056* (0.0315) −0.0736*** (0.0155) −0.0530*** (0.00826) −0.0691*** (0.00915) −0.0646*** (0.00913)

Iran 2014 −0.00879 (0.0174) −0.003 (0.043) −0.00998 (0.0161) .00414 (0.00877) −0.00295 (0.00880) −0.00200 (0.00920)

Iran 2015 −0.0163 (0.0478) −0.0103 (0.0642) −0.0174 (0.0546) −0.0395*** (0.0128) −0.0287 (0.0457) −0.0615*** (0.0100)

L.lgdp .956*** (0.0791) .829 *** (0.223) .948*** (0.0190) .970*** (0.0199) 1.004*** (0.00798) 1.005*** (0.00744)

Log Pop .00806 (0.0368) −0.018 (0.157) .0102 (0.0259) .0476 (0.0384) −0.000692 (0.00463) .000233 (0.00273)

Rents .0140 (0.143) 0.049 (0.166) .00785 (0.0762) 0.157* (0.0868) −0.0557** (0.0232) −0.0208 (0.0245)

Trade −0.0192 (0.0557) 0.034 (0.095) −0.0193 (0.0343) −0.0245 (0.0514) −0.00747 (0.0146) 0.00950 (0.0116)

Agriculture −0.029 (0.256) −0.049 (0.187) −0.0605 (0.173) −0.219 (0.171) 0.102* (0.0560) 0.0610 (0.0591)

N 291 277 284 284 299 299

T-FE Yes Yes Yes No Yes No
AR(2) . . 0.112 0.159 0.172 0.195
Overid . . 1.000 1.000 1.000 1.000

Robust standard errors in parentheses. The dependent variable is log GDP.

Variables in the FD model are first-differenced.
T-FE stands for Time Fixed Effects.
AB AR(2) (Pr > z), and Over-id (Pr > z) are reported in the bottom 2 rows.
*
p < 0.10.
**
p < 0.05.
***
p < 0.01.

with this method is the inability to approximate the sampling dis- where GDP ¨ it = GDP it − GDP it−1 and so on. In this setup, the first
tributions of test statistics. As mentioned in Abadie (2019) “in a difference cancels out the fixed effects, however, the transformed
comparative case study framework, however, the sampling-based lagged dependent variable is still correlated with the error term.
inference is complicated-sometimes because of the absence of a Anderson and Hsiao suggest using level instruments GDP it−2 or the
well-defined sampling mechanism or data generating process, and lagged difference GDP it−2 − GDP it−3 as the instruments for the dif-
sometimes because the sample is the same as the population.” In a ferences lagged endogenous regressor GDP it − GDP it−1 . However,
more recent study though, Abadie et al. (2020) study the interpreta- Arellano, 1989 suggest using levels due to lower variance and no
tion of standard errors in regression analysis when the assumption points of singularities where a condition of  = 0 invalidates the
that the sample is drawn randomly from a much larger population instruments. We utilize the methods mentioned above to estimate
of interest is not appropriate. Equation 6 and we provide the result in Table 3.

3.2. Dynamic Panel Data: fixed effect and first difference model
3.3. Dynamic Panel Data: GMM and system GMM estimation
In the Dynamic Panel Data model for country i ∈ 1, . . ., J + 1 and
for t ∈ 1, . . ., T0 , . . ., T we have in generic terms:
The Arellano and Bond (1991), Arellano and Bover (1995),
yitN = yit−1
N
+ Zit   + Xj   + ıt + i + it (1) Blundell and Bond (1998) Dynamic Panel Data estimators are
very popular. They are general estimators designed for situa-
and in our specific application: tions with (1) “small T, large N” panels, meaning few periods and
GDP it = GDP it−1 + Xit  ˇ + ıt + i + it (2) many individuals; (2) a linear functional relationship; (3) a single
left-hand-side variable that is dynamic, depending on its past real-
where GDP in this set up depends on the lagged value of itself, izations; (4) independent variables that are not strictly exogenous,
parameter  captures the effect of the lagged GDP, Xit is the set of meaning correlated with past and possibly current realizations of
predictors same as those used in the SCM. Now, to cancel out the the error; (5) fixed individual effects, and (6) heteroskedasticity
countries fixed effects i we subtract the average of each country. and autocorrelation within individuals but not across them (e.g.
 see Roodman (2006)).
˜ it = GDP˜it−1 + X˜it ˇ + ı̃t + ˜it
GDP (3)
We provide the details of these approaches in Appendix C. The
˜ it = GDP it − GDP i and the GDP i is the average of GDP results of the estimations of Eq. (2) working with the approaches
where GDP
mentioned above are summarized in Table 3.
˜ it = GDP it − 1 (GDP i1 +
for country i over the validation period (GDP T The first four regressors are dummies representing Iran under-
GDP i2 + . . . + GDP iT )). Due to the correlation between ˜it and going sanctions as of 2012, 2013, 2014, and 2015. For the predictors,
GDP˜it−1 (Nickell, 1981; Bond, 2002), we use one period lagged val- similar to SCM, we rely on a set of growth predictors that are
ues of the explanatory variables Xit−1 as instruments. standard in the analysis of countries heavily dependent on natu-
Another approach to eliminating the fixed effects is suggested ral resources rent. Specifically, we include population (log), rents,
by Anderson and Hsiao (1981) in the form of the first difference trade, and agriculture all three as value-added as a percentage of
approach: GDP. For the Fixed Effect and First Difference model, we rely on the
standard set of instruments for these models (discussed in Section
¨ it = GDP¨ it−1 + X¨it  ˇ + ı̈t + ¨it
GDP (4) 3.2). For the Arellano-Bond and Blundell-Bond model, we treat the

73
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

lagged of log GDP as endogenous and other predictors as predeter- Table 4

Unit root tests results.
mined.
The effect of the first year of sanctions on the country’s GDP Im-Pesaran-Shin test: Fisher type test:
is a negative effect from 9.5% to 12.8%. Note that while report- p-Value p-Value
ing the cumulative effect for each year after the intervention, we Log GDP 1.0000 1.0000
should consider the dynamic nature of our model which includes Log Population 0.0009 0.0000
the lagged values of GDP; therefore, the effect for each year includes Total natural rent (% of GDP) 0.0000 0.0319
Trade (% of GDP) 0.0950 0.2359
the lagged effect (multiplied by the coefficient of the lagged term).
Agriculture (% of GDP) 0.0000 0.0003
⎧
According to the fixed effect model:
For the Im-Pesaran-Shin test: H 0 : all panels contain unit roots and H a : some panels
⎪
⎨ lGDP 2012 = −0.127
are stationary. For the Fisher type test, we have H 0 : all panels contain unit roots and
lGDP 2013 = −0.0726 + 0.956 ∗ lGDP 2012 = −0.194 H a : at least one panel is stationary.
⎪
⎩ lGDP 2014 = −0.00879 + 0.956 ∗ lGDP 2013 = −0.194
lGDP 2015 = −0.0163 + 0.956 ∗ lGDP 2014 = −0.202.
⎧
Following the same calculation for the first difference model: yit = i yit−1  + Zit  i + it (5)
⎪
⎨ lGDP 2012 = −0.0948
where i indexes the units; t = 1, . . ., T indexes the time; yit is the
lGDP 2013 = −0.056 + 0.829 ∗ lGDP 2012 = −0.1346
variable being tested for unit root; i is a vector of coefficients on
⎪
⎩ lGDP 2014 = −0.003 + 0.829 ∗ lGDP 2013 = −0.114 Zit ; the deterministic terms that control for panel-specific effects
lGDP 2015 = −0.0103 + 0.829 ∗ lGDP 2014 = −0.1053.
and linear time trends. Dependent on the test and specifications,
⎧
For the Arellano Bond with time fixed effects:
Zit can be = (1, t) where Zit  i will provide the unit fixed effects,
⎪ lGDP 2012 = −0.128
⎨
lGDP 2013 = −0.0736 + 0.948 ∗ lGDP 2012 = −0.1949 unit-specific linear time trends. it is the error term. The test of
the unit-root is to test the null hypothesis H 0 : i = 1 for all units
⎪
⎩ lGDP 2014 = −0.00998 + 0.948 ∗ lGDP 2013 = −0.1948
lGDP 2015 = −0.0174 + 0.948 ∗ lGDP 2014 = −0.202 i against the alternative H a : i < 1. In some tests, the alternative
⎧
For the Arellano Bond with no time fixed effects: may hold for one unit, for some it holds for a fraction of all units or
⎪
⎨ lGDP 2012 = −0.102 all. For the details of the provided tests please refer to Appendix D.
lGDP 2013 = −0.053 + 0.970 ∗ lGDP 2012 = −0.1519 Table 4 provides the result.
⎪
⎩ lGDP 2014 = 0.00414 + 0.970 ∗ lGDP 2013 = −0.1432
lGDP 2015 = −0.0395 + 0.970 ∗ lGDP 2014 = −0.1784. 3.4.2. Co-integration tests
⎧
For the Blundell Bond with time fixed effects: A non-stationary process wanders over time because it has time-
⎪ lGDP 2012 = −0.128
⎨ varying first two moments (mean, variance, or both). Our variables
lGDP 2013 = −0.0691 + 1.004 ∗ lGDP 2012 = −0.1976 are integrated of order one, meaning that we have stationary first-
⎪
⎩ lGDP 2014 = −0.00295 + 1.004 ∗ lGDP 2013 = 0.2014 differences processes (I(1)), and we want to check if the linear
lGDP 2015 = −0.0287 + 1.004 ∗ lGDP 2014 = −0.2309. combination of our several I(1) series is stationary and the series
⎧
And finally for the Blundell Bond with no time fixed effects: are co-integrated (Engle and Granger, 1987). If we find from our test
⎪ lGDP 2012 = −0.113
⎨ that our series are co-integrated, it means in long run, they move
lGDP 2013 = −0.0646 + 1.005 ∗ lGDP 2012 = −0.178
together although they can each wander arbitrarily.
⎪
⎩ lGDP 2014 = −0.002 + 1.005 ∗ lGDP 2013 = −0.181
The Panel Data Model considered for all the co-integration tests
lGDP 2015 = −0.0615 + 1.005 ∗ lGDP 2014 = −0.2435.
is as follows and for this section, we stick to the same notation:
The Arellano Bond test of AR(2) of it is reported in Table 3. If the
idiosyncratic part of the error term it is serially correlated of order yit = Xit  ˇi + Zit  i + eit (6)
1, then yit−2 will be endogenous to it−1 and therefore would not
serve as a good instrument. Because in Arellano Bond estimation For each unit i the predictors Xit is considered to be an I(1) process.
we take first difference first, then of course having an AR(1) process ˇi denotes the co-integrating vector, which may vary across units. i
in the it is normal, so we look at the AR(2) test result of the is a vector of coefficients on Zit , the deterministic terms that control
differenced error term (the it includes fixed effects, therefore, is for panel-specific effects and linear time trends. Dependent on the
serially correlated). The result does not support the AR(2) process test and specifications, Zit can be = (1, t) where Zit  i will provide
in the differenced error term both for AB and BB estimation. the unit fixed effects, unit-specific linear time trends. eit is the error
The result of the estimation is similar to those found by the Syn- term.
thetic Control Method. Moreover, in addition to providing similar The common null hypothesis in the tests is that the dependent
estimates, PDRF provides standard errors and an estimated coef- variable and predictors are not co-integrated. The tests check the
ficient for other covariates included in the model. The strong suit co-integration by testing if eit is non-stationary. If the null of the
of the SCM is the strong resemblance of the counterfactual unit. In non-stationary error term is rejected, it means Yit and Xit are co-
Section 3.5 we provide the “fit” of the predicted unit driven from integrated.
the PDRF estimations to compare. We will provide the result of three co-integration tests: The Kao
Before that, the estimation results in Table 3 raise a cautionary test, the Pedroni test, and the Westerlund test. For further details
flag: the estimate of , which is the slope of the lagged dependent of the tests please refer to Appendix D.
variable, is close to 1 in every regression. This suggests that the Table 5 provides the result of the co-integration tests. The series
dependent variable (log GDP) may well have a unit root. In Section included in the tests are log GDP and trade, the two non-stationary
3.4, we examine the issue of unit roots and co-integration in the series in our analysis. As we can see in the table, all test statis-
data. tics reject the null hypothesis of no co-integration in favor of the
alternative hypothesis of the existence of a co-integrating relation
among the series (Westerlund test rejects the null at the 10% sig-
3.4. Unit root and co-integration test nificance level).
In principle, a regression that is based on non-stationary panel
3.4.1. Unit root tests variables may prove spurious as in the case of time-series. How-
We consider a panel-data model with a first-order auto- ever, in the case of panel data, Kao (1999) showed that estimates of
regressive component for our series: the structural parameter binding two independent non-stationary

74
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

Table 5 Table 7
Co-integration tests results. Root Mean Squared Prediction Errors.

Statistics p-Value Methods

RMSPE
Kao tests
Modified Dickey–Fuller t 2.8071 0.0025 SCM 0.0217
Dickey–Fuller t 3.0129 0.0013 DPD FE 0.0242
Augmented Dickey–Fuller t 3.2682 0.0005 DPD FD 0.0321
Unadjusted modified Dickey–Fuller t 2.5224 0.0058 DPD AB 0.0239
Unadjusted Dickey–Fuller t 2.4996 0.0062 DPD BB 0.0247
Pedroni tests (same AR parameter)
RMSPE for all methods is calculated for the period 1995 to 2011.
Modified variance ratio −3.8702 0.0001
Modified Phillips–Perron t 1.6607 0.0484
Phillips–Perron t 2.0511 0.0201
Augmented Dickey–Fuller t 1.8093 0.0352 Table 1) according to all the regressions in the table the negative
Westerlund tests (against some panels) effect of sanctions is reported to be higher than that at least for the
Variance ratio 1.5636 0.0590 first year. The result of the AR test does not show a dependency
of the idiosyncratic error term to the lagged values of the growth
which are being used as instruments, and the result of the over-
variables converge to zero, whereas in the case of time series it is a identification test shows the independence of the instruments to
random variable. This means that the point estimations of the value the included predictors.
of parameters are consistent. We also have co-integrated series in
our data so the previous results in Table 3 will hold. However, we 3.5. Estimated units
will provide the same Panel Data analysis for the growth of GDP
as well; in this analysis, we also consider the growth of trade as a Because the highlight of the Synthetic Control Method is to pro-
predictor instead of levels, and the result is provided in Table 6. duce a well-fitted control unit, we look at the “fit” of the predicted
Same as Table 3, the first column provides the result of a Fixed unit created by the DPRF. The fitted values of the models estimated
Effect model, the second column provides the result of the First above are presented in Figs. 2–5, and we observe an almost per-
Differenced model, and column 3 to 6 report the result of the Arel- fect fit similar to the Synthetic Control unit. We find the BB method
lano Bond and Blundell Bond models with and without time fixed with no time fixed effects to perform better throughout the paper,
effects. The result of the table is in line with the previous result and here we find higher prediction power observed in Fig. 5.
we obtained earlier in this paper. As we can see, the coefficient Lastly, to extend the comparison of the fit, in Table 7 below we
of the lagged term is not close to one. The effect of sanctions on provide the Root Mean Squared Prediction Errors for the Synthetic
Iran’s growth is reported to be a large significant negative effect for Control Method as well as the Dynamic Panel Data approaches. Note
2012 in all specifications. For 2013, the coefficients are negative for that
most of the regressions. For 2014, all of the models report a pos- 
T0
itive coefficient. As mentioned in Table 1, the actual growth of is 1 
J+1

Iran is reported to suffer a −7.4% drop in 2012, followed by a −0.2% RMSPE =

(y1t − wj yij )
T0
drop in 2013, a change of a positive 4.6% for 2014, and a drop of t=1 j=2
1.5% for 2015. Similar to our findings from the SCM (provided in

Table 6
Dynamic panel data model estimation result.

(FE) Growth (FD) Growth (AB) Growth (AB) Growth (BB) Growth (BB) Growth

Iran 2012 −10.9*** (2.154) −9.38*** (1.737) −12.20*** (0.999) −10.95*** (0.185) −11.14*** (0.931) −10.01*** (0.358)

Iran 2013 0.264 (4.401) −6.678*** (2.543) −3.553** (1.395) −2.138*** (0.613) −0.510 (1.339) 0.927 (1.063)

Iran 2014 1.899 (2.170) 2.589 (2.157) 0.229 (1.142) 1.020*** (0.282) 2.095*** (0.568) 2.871*** (0.286)

Iran 2015 −5.78*** (1.596) 0.873 (2.289) −5.788*** (1.291) −4.695*** (0.586) −3.316*** (0.650) −4.508*** (0.524)

L.Growth 0.413 (0.277) −0.607*** (0.079) 0.133*** (0.0396) 0.168*** (0.0322) 0.305*** (0.105) 0.390*** (0.0982)

Pop Growth 17.48 (17.97) 16.93 (23.83) 24.39 (24.06) 10.66 (27.71) 26.76 (21.35) 34.59** (17.62)

Rents 1.327 (4.451) 11.58 (7.71) 3.016** (1.432) 11.21*** (1.924) −3.722* (1.964) 2.063 (1.897)

Trade Growth −8.000 (7.573) −8.056 (5.433) −7.411 (5.167) −2.678 (6.797) −6.611 (5.105) 0.193 (6.429)

Agriculture 18.38* (10.62) −10.99 (20.06) 21.17** (9.164) −0.721 (3.528) 7.661* (4.482) 8.727* (4.649)

N 292 283 291 291 306 306

T-FE Yes Yes Yes No Yes No
AR(2) . . 0.436 0.298 0.225 0.15
Overid . . 1.000 1.000 1.000 1.000

Robust standard errors in parentheses. The dependent variable is GDP Annual Growth.
Variables in the FD model are first-differenced.
T-FE stands for Time Fixed Effects.
AB AR(2) (Pr > z), and Over-id (Pr > z) are reported in the bottom 2 rows.
*
p < 0.10.
**
p < 0.05.
***
p < 0.01.

75
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

Fig. 2. Log GDP vs. FE model fitted values.

Fig. 3. Log GDP (differenced) vs. FD model fitted values. Note that with the FD model, predicted values are plotted against first-differenced values of log GDP.

Fig. 4. Log GDP vs. AB (with no time fixed effect) model fitted values.

Fig. 5. Log GDP vs. BB (with no time fixed effect) model fitted values.

measures the discrepancy between the values of the outcome vari- 3.6. Discussion
able of interest between the actual unit and the counterfactual unit.
In other words, it measures the discrepancy between the two paths SCM has a linear underlying model (both in the structural model
(actual values and the fitted values) for each method for the pre- and in the objective function of the weights). On the other hand,
intervention period. Note that because the validation period for a standard DID model also assumes linearity in time trends and
the SCM is 1995–2011, we use the same period for RMSPE calcu- unobservable characteristics. Athey and Imbens (2006) develop
lations for the Dynamic Panel Data approaches as well. As we can and alternative approach (CIC) to DID in which they reevaluate the
see, RMSPEs for the Dynamic Panel Data approaches; specifically for “functional form” of the underlying models and they reassess the
the BB approach which also visually provided a well fitted predicted heterogeneity in the treatment effect based on unobservable char-
unit; are very small and insubstantially different than the SCM. acteristics of the units. Rather than the average treatment effect

76
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

of the outcome, we can consider the entire distribution for the Therefore, the regression estimated counter-factual of the affected
control units and how it changes over time, and this enables us unit is Y0 W reg where W reg = X0 (X0 X0 )−1 X1 . (Abadie et al., 2015b;
to estimate the entire counterfactual distribution of the outcomes Wan et al., 2018). In fact, Abadie et al. (2015b) show that simi-
for the treated units (assuming they would have followed the same lar to the SCM, the regression estimator is a weighting estimator
counterfactual distributional changes as the controls in case of no with weights that sum to one, and the counter-factual unit in a
intervention). The distributional estimation also enables us to eval- regression is also a linear combination of the control units, and the
uate the effect of a policy intervention in a mean-variance trade-off weights on the controls sum to one. But unlike the synthetic control
way. We can consider a nonlinear functional form in which we method, in the regression, weights are not restricted to the [0, 1]
allow the unobservable characteristics of the units to also affect interval.
the outcome. This strand of research is based on identification and To conclude, in our paper with N = 13 and T0 = 30, we use a tra-
estimation for the case with many individuals in each of two groups ditional SCM which assumes A1, A2, and A3 (weights summarized
and two time periods. In our analysis, we only have one treated unit in Table 1); we work with the traditional underlying linear model;
and multiple periods. The assignment to the treatment (sanctions) and we work with weights driven from the minimum distance
also can be considered to be orthogonal to the characteristic of the approach. In the DPRF which the assumptions are A2, as well as
single treated unit that we have (Iran). “no inference” assumption discussed below. The weights are driven
I is observed, we need to esti-
As mentioned in Section 3, while y1t from the regression as explained above. The underlying models of
N which can be estimated by yN =  +
J+1 the two methods are:
mate y1t 1t
w yI (objective
j=2 j jt
SCM:
function of weights). In this equation the unobserved values for the
treated unit are approximated by a linear weighted combination of yitN = ıt + t Zi + t i + it
the control units.
Doudchenko and Imbens (2016) summarize critical assump- PDRF:
tions (listed below) and different approaches that current methods
in the literature take to estimate the parameter of Eq. y1tN =+ yitN = yit−1
N
+ Zit   + Xj   + ıt + i + it
J+1
j=2
wj yjtI . Moreover, they propose a new framework that nests Gharehgozli (2021) shows SCM is an unbiased estimator if the
many of the existing approaches and allows researchers to con- outcomes follows PDRF model above. Because the comparison units
trast the critical assumptions underlying the previously proposed are meant to approximate the counterfactual of the case of interest
methods (for example allowing the negative weights). The critical without the intervention, we shall restrict the donor pool to units
assumptions in the literature are: with outcomes that are thought to be driven by the same structural
process as for the unit representing the case of interest. Moreover,
Abadie et al. (2015b) discuss that it is important to choose the
A1: No Intercept:  = 0
J+1 donors that were not subject to structural shocks to the outcome
A2: Adding-up: w =1
j=2 j variable during the sample period of the study. Abadie (2019) refers
A3: No Negative Weight: for all j wj  0 to this point as the “no interference” assumption which implies
A4: Constant Weights: for all j wj = w: that there is no interference across units. That is, units’ outcomes
Traditional SCM assumes A1, A2, and A3 while a DID assumes are invariant to other units’ treatments. The outcome variable of
A2, A3, A4. interest in our case is GDP and therefore we abide by a standard
set of relevant predictors and characteristics, while for the donor
On the same note, in another paper Wan et al. (2018), the assump- pool we try to stick to donors that have an economic structure sim-
tions and performance of the PDRF and SCM are compared through ilar to Iran, mainly being dependent on natural resources rents.
simulations. PDRF does not assume A1, and additionally assumes Gharehgozli (2020) studies the sensitivity of the methods in this
that controls’ outcome variable must not be affected by the treat- regard.
ment, and assumes no restriction to weights such as those in SCM. To avoid interference, in the data we made sure none of the
While SCM has the same 3 assumption mentioned above. They con- donors’ values of the GDP is affected by Iran’s treatment. The main
clude that PDRF performs better (lower prediction error) compared concern is about the outcome variable GDP, but moreover to avoid
to SCM as long as TN is not large (N being the number of controls, “anticipation” of the pre-treatment characteristics we incorporate
0
and T0 being the number of pretreatment periods). the pre-treatment characteristics as the ratio to the GDP. Trade, for
Moreover, regarding N and T0 , Doudchenko and Imbens (2016) example, is used in terms of percentage of GDP. Note that the base
summarize that if T0  N meaning too little pre-treatment peri- of our donor pool is constructed from OPEC countries that a have
ods and too many controls, matching methods work better because very similar economic structure to Iran, however, we needed to
it is easier to find a potential matching control unit among many expand our donor pool to have a better match while working with
that fit the treated unit for the few periods of pre-treatment. If the SCM. In our choice of donors, we made sure:
N  T0 meaning too many pre-treatment periods and too little con-
trols, predictions based on time-series forecasting may be a better 1. No country is affected by the sanction in terms of GDP (none of
approach. And if N ≈ T0 then the strategy to choose is more critical the countries actual GDP is responsive to Iran’s sanctions).
and maybe working with linear combination of control is a good 2. And again, because we work with ratios of the variable as a
approach. percentage of GDP none of the predictors are affected by the
Going back to the assumptions and our current research, in the intervention substantially.
PDRF the linearity of the unobservable characteristics (i.e. i ) is
assumed, and the PDRF’s counter-factual unit for post-intervention In the end, in contrast to a regression framework, the SCM does not
ˆ  X1 , where ˇ
period is a ((t − T0 ) × 1) vector of ˇ ˆ is the matrix of impose fundamental restrictions on the choice of the variables, or
regression coefficients of Y0 on X0 . Y0 is the matrix of the val- the donor units. As concluded by Wan et al. (2018) “PDRF rules out
ues of the outcome variable for control units, X1 is (K × 1) vector control units that could be affected by the treatment and places no
containing the values of predictors of the treated unit that we restriction on the predictors while SCM does not place restrictions
aim to match as closely as possible and X0 is the (K × J) matrix on the control units but constrains the weights to be non-negative
containing the values of the same variables for the donor units. and must add up to one.” As we mention in the second paragraph of

77
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

the Introduction Section on Page 3: the SCM is more of an optimiza- years of sanctions), the Dynamic Panel Data framework reports a
tion problem that aims to minimize the discrepancy between the drop of 24.35% (according to the Blundell Bond approach). Not only
synthetic counterfactual unit and the actual treated unit in terms of does the regression framework provides similar estimates to the
the outcome variable over a specific pre-intervention period. It may Synthetic Control Method, but it also has the advantage of estima-
potentially ignore the economic principles of causal relationships. tion of the standard errors. Moreover, because the highlight of the
One could extend the current comparison, possibly through Synthetic Control Method is to produce a well-fitted control unit,
simulations, to allow the effect of such treatment to vary by unob- we look at the “fit” of the predicted unit created by the Dynamic
servable characteristics of the units, and to allow the distribution Panel Data Regression Framework, and we observe an almost per-
of those unobservables to vary across units. That is, the treatment fect fit similar to the Synthetic Control unit.
group may differ from the control group not just in terms of the dis-
tribution of outcomes in the absence of the treatment but also in Conflict of interest
the effects of the treatment (Doudchenko and Imbens, 2016). This
could open the door to evaluate a range of policy related questions, None declared.
such as questions about mean-variance trade-offs or which parts
of the distribution benefit most from a policy, while maintaining a Acknowledgements
single, internally consistent economic model of how outcomes are
generated. Once could also consider the possibility of unstable con- I wish to thank Professor Wim Vijverberg who has been a con-
trol and treated units over time. But points mentioned above are tinuous source of guidance and help throughout this study. I am
not addressed in this paper, and we consider them as limitations grateful for his valuable comments and suggestions. I thank the
of the current study due to the nature of the observational data we editor and two anonymous reviewers whose comments helped
have, the unequal number of treated units (1) and controls (13), and improve this manuscript greatly.
due to the fact that the treatment (sanctions) can be considered to
be orthogonal to the unobservables of the treated unit (Iran), and Appendix A. Data appendix
even an exogenous shock to the outcome.
The data source employed for the analysis are as follows:
4. Conclusion
–Gross Domestic Production (PPP, Constant 2011 International
In comparative case studies, we need similarities in the covari-
Dollars). Source: World Bank, International Comparison Program
ates between the treated unit and the controls. The Synthetic
Database, WEI, 2020. Second Source: IMF, World Economic Out-
Control Method creates a control unit that resembles the values
look Databases (WEO) 2020.
of the outcome variable almost perfectly. However, the product of
–Gross Domestic Production per Capita (PPP, Constant 2011 Inter-
the method is a point estimate of the effect of the intervention. The
national Dollars). Source: World Bank, International Comparison
placebo studies have been developed in pursuit of ways to test the
Program Database, WEI, 2020. Second Source: IMF, World Eco-
significance of the point estimates, however, there is no assumption
nomic Outlook Databases (WEO) 2020.
about the asymptotic distribution of the estimator. In the Synthetic
–GDP Deflator (100 in 2011). Source: World Bank National
Control Method, the choice of the variables and the donors are less
Accounts Data, WEI, 2020. Second Source: IMF, World Economic
definitive, less restricted, and less determinant of the result com-
Outlook Databases (WEO) 2020.
pared to a traditional regression framework. While in a regression
–PPP Conversion Factor, GDP (LCU per International Dollar).
framework, the causal relationship between macroeconomic vari-
Source: World Bank, International Comparison Program Database,
ables is more fundamentally defined and the point estimates are
WEI, 2020.
followed by estimation of the standard errors and the tests of the
–Total Population. Source: OPEC Annual Statistical Bulletin, 2020.
significance. Hence in this paper, we aim to empirically compare
–Total Natural Resources Rents (Percentage of GDP). Source:
the Synthetic Control Method with a Dynamic Panel Data Regres-
World Bank National Accounts Data, WEI, 2020.
sion Framework. The question under study is whether the Dynamic
–Agriculture, Value Added (Constant 2010 USD). Source: World
Panel Data model performs as well in comparative case studies in
Bank National Accounts Data, WEI, 2020.
the case of an intervention in which the outcome variable of inter-
–Services, etc., Value Added (Constant 2010 USD). Source: World
est suffers a shock and behaves irregularly. Specifically, we study
Bank National Accounts Data, WEI, 2020.
the effect of international sanctions against Iran as an important
–Industry, Value Added (Percentage of GDP). Source: World Bank
case of policy intervention and a good example of a comparative
National Accounts Data, WEI, 2020.
case study.
We observe that the Dynamic Panel Data approach results in
Appendix B. Descriptive statistics
similar estimates of the effects. We conclude that close to the Syn-
thetic Control Method (which estimates the effect of sanctions on
Table 8 provides descriptive statistics:
Iran’s GDP to be a negative 24.5% in 2015 after four consecutive

Table 8
Descriptive statistics.

Variable Mean Std.Dev. Min Max Observations

Year Overall 1997 10.39 1980 2015 N = 504

Between . 0 1997 1997 n = 14
Within . 10.11 1980 2015 T = 36
ID Overall 7.5 4.04 1 14 N = 504
Between . 4.18 1 14 n = 14
Within . 0 7.5 7.5 T = 36
GDP Overall 796.89 2013.84 0 17406.24 N = 499
Between . 1485.91 31.15 5478.43 n = 14
Within . 1407.77 −4307.85 12692.13 T = 35.64

78
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

Table 8 (Continued)

Variable Mean Std.Dev. Min Max Observations

Pop. Overall +08

1.12e 3.09e +08
223715 1.37e +09
N = 501
Between . 3.18e+08 738145 1.21e+09 n = 14
Within . 3.34e+07 −1.16e+08 2.74e+08 T = 35.8
Rents Overall 19.69 15.42 0.12 66.48 N = 491
Between . 13.56 0.55 38.79 n = 14
Within . 8.36 -11.91 57.51 T = 35.1
Trade Overall 72.64 38.61 12.42 251.14 N = 455
Between . 37.19 37.18 160.28 n = 14
Within . 16.39 22.89 163.5 T = 32.5
Ind. Overall 46.42 23.74 20.16 213.69 N = 304
Between . 23.59 28.76 118.24 n = 13
Within . 13.38 −14.37 141.88 T = 23.38
Agr. Overall 11.53 10.39 0.09 48.57 N = 337
Between . 9.65 0.17 32.68 n = 14
Within . 4.35 −0.92 27.41 T = 20.1
Growth Overall 4.09 7.04 −62.08 33.99 N = 456
Between . 3.12 −1.606 10.99 n = 14
Within . 6.58 −56.38 34.57 T = 32.6
Ser. Overall 43.72 11.35 19.49 71.31 N = 337
Between . 11.73 22.97 68.74 n = 14
Within . 6.68 25.43 71.37 T = 20.1

Appendix C. Dynamic Panel Data model by (M⊥ ⊗ IN ), where the matrix IN is the identity matrix of
order N and M⊥ is as below. Because lagged observations of a vari-
Arellano-Bond estimation starts by transforming all regres- able do not enter the formula for the transformation, they remain
sors, usually by differencing, and uses the Generalized Method of orthogonal to the transformed errors and available as instruments.
Moments (Hansen, 1982), and so is called Difference GMM. The On balanced panels, GMM estimators based on the two trans-
Arellano Bover/Blundell-Bond estimator augments Arellano-Bond forms return numerically identical coefficient estimates, holding
by making an additional assumption, that the first differences of the instrument set fixed Arellano and Bover (1995)2 :
instrument variables are uncorrelated with the fixed effects. This ⎡ ⎤
allows the introduction of more instruments, and can dramatically T −1 1 1
⎢ − − . . .⎥
improve efficiency. It builds a system of two equations – the origi- ⎢ T ⎥
nal equation as well as the transformed one – and is known as the ⎢ T (T − 1) T (T − 1)
⎥
⎢ T −1 1 ⎥
System GMM estimator. ⎢ − . . .⎥
⎢ T −2 (T − 1)(T − 2) ⎥
To further open the approaches mentioned above, let us go back M⊥ = ⎢
⎢
 ⎥
⎥
⎢ T −2
. . .⎥
to the general Dynamic Panel Data model we consider:
⎢ T −3 ⎥
GDP it = ˇ0 + GDP it−1 + Xit  ˇ + ıt + i + it (7) ⎢ ⎥
⎢ . ⎥
⎣ ⎦
With the Arellano-Bond estimation, to improve efficiency, we can .
take the Anderson-Hsiao approach further, using longer lags of the .
dependent variable as additional instruments. So Arellano-Bond A potential weakness in the Arellano-Bond estimator was
estimation starts with the first-difference transform (just like the revealed in later work by Arellano and Bover (1995) and Blundell
Anderson Hsiao estimation discussed earlier) followed by a GMM and Bond (1998). As Blundell and Bond (1998) demonstrate in sim-
estimator. If we assume a matrix M with the orthogonal of −1’s ulations, if yit is close to a random walk, then Difference GMM
and sub-diagonal of 1’s just to the right, and the matrix IN the iden- performs poorly because past levels convey little information about
tity matrix of order N then the transformation is to multiply the future changes, so that un-transformed lags (levels) are weak
model by (M ⊗ IN ). This takes us to Eq. (4): instruments for transformed variables.
⎡ ⎤ Note that we can also write the model above as:
−1 1 0 0 . .
⎢ 0 −1 1 0 0 . ⎥ GDP it = ( − 1)GDP it−1 + Xit  ˇ + ıt + i + it (8)
⎢ 0 0 −1 ⎥
M = ⎢ ⎢ 1 0 . ⎥
. .⎥ So the model can equally be thought of as being for the level or
⎣ ⎦ increase of y. Therefore:
. .
0 . . 0 −1 1 
GDP it−1 = ( − 1)GDP it−2 + Xit−1 ˇ + ıt−1 + i + it−1 (9)
The fixed effects are gone, but as mentioned earlier, the lagged This will introduce more options for instruments as long as 
dependent variable is still potentially endogenous as well as any does not equal to 1. So where Arellano-Bond instruments endoge-
predetermined variables in the regressors set that are not strictly nous first differences or orthogonal deviations with lagged levels,
exogenous become potentially endogenous. But longer lags of the Blundell-Bond instruments levels with differences. For variables
regressors remain orthogonal to the error and available as instru- that follow a random walk process, past changes may indeed be
ments. more predictive of current levels than past levels are of current
Another possible transformation can be the “orthogonal devia- changes, so that the new instruments are more relevant (Blundell
tions transform” proposed by Arellano and Bover (1995), in which
rather than subtracting the previous observation, we subtract the
average of all available future observations. Like first-differencing, 2
Our result is robust to the choice of the transformation (FD or FOD), to avoid too
taking orthogonal deviations removes fixed effects. In a balanced many reported regressions we will not provide the result of this transformation in
panel, the transformation can be done by multiplication of the the tables.

79
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

and Bond, 1998; Roodman, 2006). Blundell-Bond estimation builds For the PP t statistic reported by Pedroni test, first, we estimate
a stacked data set with twice the observations to use new moment Eq. (6) using ordinary least square, and then we fit the DF regression
conditions while retaining the Arellano Bond moments. The matrix model below which is slightly different from above:
M∗+ is as follows:
  êit = i êi,t−1 + it (11)
M∗ In this model, we have unit-specific , and a PP t tests if i = 1,
M∗+ =
I and a modified pp t tests if i − 1 = 0.
The Augmented DF t uses additional lags of the error terms to
where M∗ can be M or M⊥ 3 . incorporate serial correlation. We still check if  = 1 but the ADF
regression model is:
Appendix D. Unit roots and co-integration tests

p

êit = i êi,t−1 + ij êi,t−1 + it∗ (12)

One of the key differences between different unit root tests
j=1
is how they treat the auto-regressive parameter i Out of the 6
unit root tests available, the Levin-Lin-Chu test (Levin et al., 2002),
where êi,t−1 is the jth lag of the first difference of êit and j =
Harris-Tsavalis test (Harris and Tzavalis, 1999), and Breitung test
1, . . ., p where p is the number of lag differences.
(Breitung and Das, 2005; Breitung, 2001) assume that all the units
To be more clear, the Kao test (Kao, 1999) assumes ˇi = ˇ mean-
share the same auto-regressive parameter meaning i =  for all i.
ing that there is a co-integration vector that is the same across all
However, if as an example, the variable being tested is the economic
the units, and there is a unit fixed effect and no linear time trend
growth rate, and we are testing whether economic growth rates of
(Zit = (1, 0)). As mentioned above, the null hypothesis of the Kao
countries converge to a long-run value, this restrictive assumption
test is that there is no co-integration among the series. So, the
implies that the rate of convergence would be the same for all the
co-integration relationship is:
countries (Maddala and Wu, 1999).
Another key difference is whether these tests assume N or T is yit = Xit  ˇ + i + eit (13)
fixed, and if not at what rate they tend to infinity. Hlouskova and
The tests derived by Pedroni (1999, 2004) on the other hand,
Wagner (2006) provide an overview of these assumptions and com-
allow for unit-specific co-integrating vectors ˇi . Also, the DF
pare the performance of these tests using Monte Carlo simulation.
regression model incorporates heterogeneous unit-specific AR
Another difference across these test is whether they perform with
coefficients (i ), although the tests allow restricting the AR coeffi-
balanced or unbalanced data. The first three tests mentioned above
cient to be the same across units. Considering a linear time trend is
only perform with the balanced data. The Im-Pesaran-Shin test (Im
also possible with this test. Same to the Kao test, the null hypothesis
et al., 2003) assumes T and N are fixed or N can be infinite, assumes
is no co-integration between the dependent variable and covari-
auto-regressive parameter is unit specific, and it performs well with
ates.
unbalanced data. The Fisher type test (Choi, 2001) assumes T is
In the Westerlund tests (Westerlund, 2005) similar to Pedroni’s
not finite and N can be both finite and infinite. The test assumes
tests, the AR parameter is considered to be unit-specific (i ). In
unit-specific auto-regressive parameter and it performs well with
contrast to the Kao and Pedroni tests that check the null hypothe-
unbalanced data. The last test, the Hadri LM test (Hadri, 2000) also
sis against an alternative of “all series are co-integrated”, this test
assumes infinite T and N and it requires balanced data. The most
checks the null against the alternative that some of the series are
appropriate test for our case of study seems to be the Im-Pesaran-
co-integrated. However, the test allows for the alternative of “all
Shin test. However, we also provide the result of the Fisher type
series are co-integrated” under the restriction i = . So the West-
test in Table 4. The Hardi LM test is not feasible due to the missing
erlund derived two Variance Ratio statistics one for unit-specific
values (the test requires strongly balanced data). We also will not
AR parameter and one for the same AR parameter across all units.
consider the first three tests due to the reasons mentioned above.
The series included in the tests are log GDP, log population, trade,
References
natural rents, and agriculture. As we can see in this table, log GDP
and trade are two series that follow a unit root process. Abadie, A. (2019). Using synthetic controls: Feasibility, data requirements, and
The details of the co-integration tests are as follows: methodological aspects. Journal of Economic Literature.
We provided the result of three co-integration tests: The Kao Abadie, A., & Gardeazabal, J. (2003). The economic costs of conflict: A case study of
the Basque country. American Economic Review, 93(1), 112–132.
test, the Pedroni test, and the Westerlund test. The different DF Abadie, A., & Imbens, G. (2002). Simple and bias-corrected matching estimators for
statistics that are reported by the first two tests use different regres- average treatment effects. NBER Technical Working Paper (283).
sion frameworks to incorporate the serial correlation in eit . The VA Abadie, A., Diamond, A., & Hainmueller, J. (2010). Synthetic Control Methods for
comparative case studies: Estimating the effect of California’s Tobacco Control
test reported by the latter two, do not require any accommoda- Program. American Statistical Association, 105(490), 493–505.
tions for serial correlation. All different Dicky–Fuller t-test statistics Abadie, A., Diamond, A., & Hainmueller, J. (2015a]). Comparative politics and the
that are reported with these tests are constructed by fitting a ver- Synthetic Control Method. American Journal of Political Science, 59(2), 495–510.
Abadie, A., Diamond, A., & Hainmueller, J. (2015b]). Comparative politics and the
sion of Eq. (6) using ordinary least squares, obtaining the predicted
synthetic control method. American Journal of Political Science, 59(2), 495–510.
residuals, and then fitting the DF regression model: Abadie, A., Athey, S., Imbens, G. W., & Wooldridge, J. M. (2020). Sampling-based
versus design-based uncertainty in regression analysis. Econometrica, 88(1),
êit = êi,t−1 + ˜it it (10) 265–296.
Allegretto, S., Dube, A., Reich, M., & Zipperer, B. (2017). Credible research designs
for minimum wage studies: A response to Neumark, Salas, and Wascher. ILR
where  is the AR parameter and it is a stationary error term. The Review, 70(3), 559–592.
DF t and un-adjusted DF t check if  = 1. On the other hand, the Anderson, T. W., & Hsiao, C. (1981). Estimation of dynamic models with error
Modified DF t and un-adjusted Modified DF t check if  − 1 = 0, components. Journal of the American Statistical Association, 76(375), 598–606.
Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte
and non-stationarity under the null hypothesis causes these two to
Carlo evidence and an application to employment equations. Review of
be different. Economic Studies, 58, 277–297.
Arellano, M., & Bover, O. (1995). Another look at the instrumental variable
estimation of error-components models. Journal of Econometrics, 68, 29–51.
Arellano, M. (1989). A note on the Anderson Hsiao estimator for panel data.
3
The notation for matrix M is borrowed from Roodman (2006). Economics letters, 31, 337–341.

80
O. Gharehgozli The Quarterly Review of Economics and Finance 81 (2021) 70–81

Athey, S., & Imbens, G. W. (2006). Identification and inference in nonlinear Heckman, J., Ichimura, H., & Todd, P. (1997). Matching as an econometric
difference-in-differences models. Econometrica, 74(2), 431–497. evaluation estimator: Evidence from evaluating a job training program. Review
Barnow, B. S., Cain, G. G., & Goldberger, A. S. (1980). Issues in the analysis of of Economic Studies, 64, 605–654.
selectivity bias. In E. Stromsdorfer, & G. Farkas (Eds.), Evaluation studies (p. 5). Hlouskova, J., & Wagner, M. (2006). The performance of panel unit root and
Blundell, R., & Bond, S. (1998). Initial conditions and moment restrictions in stationarity tests: Results from a large scale simulation study. Econometric
Dynamic Panel Data models. Journal of Econometrics, 87, 115–143. Reviews, 25(1), 85–116.
Bond, S. R. (2002). Dynamic Panel Data models: A guide to micro data methods and Hsiao, C., Steve Ching, H., & Ki Wan, S. (2012). A Panel data approach for program
practice. Portuguese Economic Journal, 1(2), 141–162. evaluation: Measuring the benefits of political and economic integration of
Borjas, G. J. (2017). The wage impact of the marielitos: A reappraisal. ILR Review, Hong Kong with mainland China. Journal of Applied Econometrics, 27(5),
70(5), 1077–1110. 705–740.
Breitung, J. (2001). The local power of some unit root tests for panel data. Im, K. S., Pesaran, M. H., & Shin, Y. (2003). Testing for unit roots in heterogeneous
Breitung, J., & Das, S. (2005). Panel unit root tests under cross-sectional panels. Journal of Econometrics, 115(1), 53–74.
dependence. Statistica Neerlandica, 59(4), 414–433. Imbens, G. W. (2002). Non-parametric estimation of average treatment effects
Card, D., & Krueger, A. (1994). Minimum wages and employment: A case study of under exogeneity: A review. Review of Economics and Statistics, 86(1),
the fast-food industry in New Jersey and Pennsylvania. American Economic 4–29.
Review, 84(4), 772–793. Kao, C. (1999). Spurious regression and residual-based tests for cointegration in
Choi, I. (2001). Unit root tests for panel data. Journal of International Money and panel data. Journal of Econometrics, 90(1), 1–44.
Finance, 20(2), 249–272. Lechner, M. (1999). Earnings and employment effects of continuous off-the-job
Collier, D. (1993). The Comparative Method. In W. Ada, & Finifter (Eds.), Political training in East Germany after unification. Journal of Business and Economic
science: The state of the discipline II (pp. 105–119). Washington, DC: American Statistics, 17(1), 74–90.
Political Science Association. Levin, A., Lin, C.-F., & Chu, C.-S. J. (2002). Unit root tests in panel data: Asymptotic
Cunningham, S., & Shah, M. (2018). Decriminalizing indoor prostitution: and finite-sample properties. Journal of Econometrics, 108(1), 1–24.
Implications for sexual violence and public health. The Review of Economic Li, K. T., & Bell, D. R. (2017). Estimation of average treatment effects with panel
Studies, 85(3), 1683–1715. data: Asymptotic theory and implementation. Journal of Econometrics, 197(1),
Dehejia, R. H., & Wahba, S. (1999). Casual effects in non-experimental studies: 65–75.
Reevaluating the evaluation of training programs. Journal of American Statistics Lijphart, A. (1971). Comparative politics and the comparative method. American
Association, 94(448), 1053–1062. Political Science Review, 65(3), 682–693.
Dehejia, R. H., & Wahbe, S. (2002). Propensity Score-Matching Methods for Maddala, G. S., & Wu, S. (1999). A comparative study of unit root tests with panel
non-experimental causal studies. Review of Economics and Statistics, 84(1), data and a new simple test. Oxford Bulletin of Economics and Statistics, 61(S1),
151–161. 631–652.
Donohue, J. J., Aneja, A., & Weber, K. D. (2019). Right-to-carry laws and violent Nickell, S. (1981). Biases in dynamic models with fixed effects. Econometrica, 49(6),
crime: A comprehensive assessment using panel data and a state-level 1417–1426.
Synthetic Control analysis. Journal of Empirical Legal Studies, 16(2), Holland, p. (1986). Statistics and causal inference. Journal of the American Statistical
198–247. Association, 81, 945–970.
Doudchenko, N., & Imbens, G. W. (2016). Balancing, regression, Pedroni, P. (1999). Critical values for cointegration tests in heterogeneous panels
difference-in-differences and synthetic control methods: A synthesis. National with multiple regressors. Oxford Bulletin of Economics and Statistics, 61(S1),
Bureau of Economic Research. Technical report. 653–670.
Engle, R. F., & Granger, C. W. (1987). Co-integration and error correction: Pedroni, P. (2004). Panel cointegration: Asymptotic and finite sample properties of
Representation, estimation, and testing. Econometrica, 251–276. pooled time series tests with an application to the ppp hypothesis. Econometric
Fitzgerald, J., Gottschalk, P., & Moffitt, R. (1998). An analysis of sample attrition in Theory, 20(3), 597–625.
panel data: The Michigan panel study of income dynamics. Journal of Human Peri, G., & Yasenov, V. (2019). The labor market effects of a refugee wave synthetic
Resources, 33, 251–299. control method meets the mariel boatlift. Journal of Human Resources, 54(2),
Gardeazabal, J., & Vega-Bayo, A. (2017). An empirical comparison between the 267–309.
synthetic control method and Hsiao et al.’s panel data approach to program Roodman, D. (2006). How to do xtabond2: An introduction to “Difference” and
evaluation. Journal of Applied Econometrics, 32(5), 983–1002. “System” GMM in Stata. Working Paper. The Center for Global Development
Gharehgozli, O. (2017). An estimation of the economic cost of recent sanctions on (103).
Iran using the Synthetic Control Method. Economics Letters, 157, 141–144. Rosenbaum, P., & Rubin, D. (1983). The central role of the Propensity Score in
Gharehgozli, O. (2020). A monte carlo analysis of robustness of the synthetic observational studies for causal effects. Biometrica, 70, 41–55.
control method and dynamic panel estimation: a comparative case study of a Rosenbaum, P. R. (2005). Observational study. Encyclopedia of Statistics in
policy intervention. Journal of Statistical and Econometric Methods, 9(1), 67–87. Behavioral Science, 3, 1451–1462.
In press. http://www.scienpress.com/Upload/JSEM%2FVol%209 1 4.pdf Rubin, D. (1974). Estimating causal effects of treatments in randomized and
Gharehgozli, O. (2021). Panel data model and the synthetic control method: a note non-randomized studies. Journal of Educational Psychology, 66, 688–701.
on the unbiasedness of the synthetic control estimator. Econometrics Letters, Rubin, D. (1977). Assignment to treatment group on the basis of a covariate.
7(2), 1–11. http://easyfinancetechnology.com/wp-content/uploads/2021/05/ Journal of Educational Statistics, 2(1), 1–26.
Orkideh-Gharehgozli.pdf Rubin, D. (1978). Bayesian inference for causal effects: The role of randomization.
Hadri, K. (2000). Testing for stationarity in heterogeneous panel data. The Annals of Statistics, 6, 34–58.
Econometrics Journal, 3(2), 148–161. Wan, S.-K., Xie, Y., & Hsiao, C. (2018). Panel data approach vs synthetic control
Hansen, L. P. (1982). Large sample properties of Generalized Method of Moments method. Economics Letters, 164, 121–123.
estimators. Econometrica, 1029–1054. Westerlund, J. (2005). New simple tests for panel cointegration. Econometric
Harris, R. D., & Tzavalis, E. (1999). Inference for unit roots in dynamic panels where Reviews, 24(3), 297–316.
the time dimension is fixed. Journal of Econometrics, 91(2), 201–226.

81