European Integration and
Economic Growth:
A Counterfactual Analysis
Nauro F Campos
Brunel University
Fabrizio Coricelli
Luigi Moretti
Paris School of Economics
University of Padova
Conference on “Transition Economics Meets New Structural Economics”
London, SSEES/UCL, June 2013
Motivation
• Are the countries that joined the European Integration
project better-off?
• Direct costs of EU membership (ok), indirect costs (???),
and benefits (??)
• Voluminous literature on effects of single market, Euro,
enlargements, trade and growth
• Range of estimates from Eichengreen-Boltho to
Badinger: without Integration, pci Europe 5-20% lower
Counterfactuals are key
• Counterfactuals and causality
• Wide use of counterfactuals: “EU average” and
“compared to France” (“75% of EU average”)
• Can we improve upon these counterfactuals?
Research Question and Method
• What would have been the growth rates of per capita
GDP and productivity in EU countries if they had not
become full-fledged EU members?
• Synthetic control methods for causal inference in
comparative case studies
or
“synthetic counterfactuals”
• Abadie et al: AER 2003, JASA 2009, mimeo 2012
Method: Synthetic counterfactuals
• A recent development in econometrics of program
evaluation (Imbens and Wooldridge JEL 2009)
• “artificial control group” (JEL 2009, p. 79)
• It estimates the effect of a given intervention by
comparing the evolution of an aggregate outcome
variable for a country “treated” to its evolution for
a synthetic control group
Synthetic counterfactuals (con’t)
• Researcher specifies: (1) treatment (what and
when), (2) matching covariates, and (3) “donor
pool” (to synthetic/artificial control group)
• Method minimizes the pre-treatment distance
(mean squared error of pre-treatment
outcomes) between the vector of treated
country’s characteristics and the vector of
potential synthetic control characteristics
What is a SYNTHETIC COUNTERFACTUAL?
More formally:
Be Y an outcome variable (eg. GDP per capita).
 it  Yit  Yit
I
C
t . T0
where Y itC is unknow for
Given N+1 the observed countries, with i=1 the treated country and i =2,…, N+1 the
control/donor countries, Abadie et al. (AER 2003, JASA 2010) show that:
N 1
ˆit  Y   w i *Yit
for t  T.0
I
it
i2
N 1
The set of weights is
W  ( w 2 ,..., w N 1 ) with
w
i
and
1
w.i  0
i2
Thus pre-treatment:
N 1
w
N 1
i
* Yit  Y1t
i2
where Z is a set of covariates/predictors of Y.
w
i2
i
* Z i  Z1
Original Example: Basque GDP & ETA
SYNTHETIC COUNTERFACTUAL: Assumptions
Assumptions:
1.
Z should contain variables that help the approximation of Y1t pre-treatment, but
should not include variables which anticipate the effect.
2. Donor countries (i=2,…,N+1) should not be affected by the treatment.
If assumptions (1) and (2) do not hold, it's likely that the estimation of the posttreatment effect is downward biased.
Advantages:
•
•
It allows the study of the dynamic effects.
It is designed for case-study, so it can allow the evaluation of treatment
independently from: i) the number of treated units; ii) the number of control units;
iii) the timing of the treatment.
Disadvantages:
•
It does not allow the assessment the significance of the results using standard
(large-sample) inferential techniques: only permutation tests on the donor sample
(placebo experiment).
What did we do?
• Synthetic counterfactuals method
• Estimate growth and productivity payoffs
• EU membership
• All enlargements: 1973, 1980s, 1995, 2004
Three key issues
1. Year treatment starts (EU membership)
– 1973: IRL, DK, UK; 1980s: Greece, SP, Port; 1995: Austria,
Fin, Sweden; 2004: Poland CZ etc
2. Matching over which covariates?
–
Similar to Abadie AER 2003: investment, labour force,
population, share of agriculture in GDP, level of secondary
and tertiary education, etc
3. Donor pool: used a range from whole world to
neighbours, but report upper middle income
Main Results
15000
10000
5000
rgdpch
20000
Portugal
1970
1980
1990
year
PRT
2000
synthetic PRT
2010
Main Sensitivity analysis:
2004 Enlargement and Anticipation
Not shown today:
different GDP measures, of labour productivity,
changes in covariate sets, regional evidence,
Full range of placebo tests
Statistical significance
DID estimates show most
results are statistically
significant
Interpretation
Summary and main findings
• Strong tendency for the growth and productivity
effects from EU membership to be positive
• Yet considerable heterogeneity across countries
• GDP/productivity significantly increase: Denmark,
Ireland, UK, Portugal, Spain, Austria, Finland,
Estonia, Poland, Latvia and Lithuania
• Growth effects tend to be smaller: Sweden, Czech
Republic, Slovakia, Slovenia and Hungary
• Greece is the only exception
• Magnitude of aggregate, average effect: 10 percent
Thank you