Trade Creation and Diversion Revisited:
Accounting for Model Uncertainty and
Natural Trading Partner Effects
Christian Henn
Theo S. Eicher
Chris Papageorgiou
IMF
University of Washington
IMF
Road Map

Reexamine PTAs’ impacts on trade flows
 Introduce New Methodology:
 Address Model Uncertainty
 Allow for Natural Trading Partners
 Generate New Results:
 Identify PTAs’ tangible benefits
 Identify individual PTA’s trade creation/diversion
 Highlight relevance of comprehensive approach
2
A Brief History of Trade Flow
Determinants





Ricardo (Technology)
Heckscher / Ohlin (Factor Endowments)
Brander / Spencer / Krugman (IRS, Market Structure)
Rivera Batiz / Romer / Young (Endogenous Growth)
Melitz/Antras (Intra-Industry Cost Heterogeneity)
3
Trade Flows and Trade Restrictions

Trail Blazers: Static Models
 Kruger (1972), Findlay Wellisz (1982), Hillman (1982)

Second Generation: Endogenous Protection Models
 Grossman Helpman (1995) etc.

Preferential Trade Agreements (PTAs)
4
Trade Flows and Trade Restrictions

Preferential Trade Agreements (PTAs)
Viner (1950) (Static Trade Creation/Diversion)
Stepping Stones vs. Stumbling Blocks Krugman (1991, 1993)

PTAs cannot be trade creating in the absence of intercontinental transport costs.
 Prohibitive transport costs => PTA trade creation dominates
 Finite transport costs => PTA trade diversion dominates
Frankel,
Stein and Wei (1995, 6, 8) model continuum of transport costs.
 The more remote trading partners are from the rest of the world, the more
likely they are to form RTAs due to less potential trade diversion.
 The more “natural” trading partners are, the more likely an FTA will be
formed by the countries’ governments due to more potential trade creation.
5
Trade Flows and Trade Restrictions
 Stumbling
Blocks
 Lobbies as stumbling blocks, b/c common external tariff
(Panagranya & Findlay 1996)
 Diverting PTAs are politically more likely (GH 1995)
 Larger PTAs have monopoly power (Deardorff and Stern 1994)
 Given differences in factor endowments, trade with a few
countries is sufficient to maximize gains from trade. Deardorff
and Stern and Haveman (1994)
 PTA’s inhibit further multilateral tariff reduction, (Krishna 1998)
6
Trade Flows and Trade Restrictions

Stepping Stones (“complementary effect of pref. tariffs Bagwell Staiger
1998)
preferential tariff induces trade diversion, which is costly, so external tariffs
declines to compensate and shift imports back to their original source)
 PTAs allow lower external tariff b/c (tariff revenue competition, PTA free
trade benefits allow for lower tariff revenues, so external tariff falls. RTA may
promote external liberalization, Richardson 1993)
 Baldwin (1993, 1995, 1997) & Levy (1996)
 Non-member exporters lobby forces joining if diversion is large
 Larger PTAs shifts more power to the export lobby
 Larger PTA even more attractive to join (more trade creation)
 PTA induced harmonization, allows new revenues that overcome trade fixed
costs Freund (2000)
7
PTA Effects Can Have Many Sources
That Are Difficult to Disentangle

Possible Scenarios:




Among members: Trade Creation / Trade Diversion
Between members: Trade Diversion / Open Block
Between Non Members: Trade Diversion /Open Block
Which Theory is Empirically Relevant?


Which Effect(s) Dominate?
How do we sift through the many models for evidence??
8
PTA Effects Can Have Many Sources
That Are Difficult to Disentangle

Possible Scenarios:




Among members: Trade Creation / Trade Diversion
Between members: Trade Diversion / Open Block
Between non members: Trade Diversion /Open Block
Which Theory is Empirically Relevant?



Which effect(s) dominate?
How do we sift through the many models for evidence??
This is the DEFINITION of Models Uncertainty
9
Model Uncertainty: Why do we care?

Individual researchers typically emphasize a single
model as they seek support for a particular regressor


(the alternative is “Null Hypothesis”, or “no effect”)
Inferences procedures based on a single model
overstate the precision of the inferences
  procedures do not account for the additional
uncertainty surrounding the validity of model.

standard errors understate uncertainty
10
Trade Creation
Trade Diversion
Open Bloc
/
AFTAij
ANZCERTAij
APECij
APij
CACMij
CARICOMij
EEAij
EFTAij
EUij
LAIAij
MERCOSURij
NAFTAij
AFTAi
ANZCERTAi
APECj
APj
CACMj
CARICOMi
EEAi
EFTAi
EUi
LAIAi
MERCOSURi
NAFTAi
Est. relationship, past studies
Positive
None
3
2
1
3
3
2
4
2
3
9
4
2
1
2
1
2
1
Negative
5
9
2
3
3
1
1
1
2
2
2
1
1
2
2
2
2
2
1
Previous Approach To Model Uncertainty
In Trade Flow Empirics

Ghosh and Yamarik (2004, JIE)
 First (and to date only) attempt to account for
model uncertainty in trade flow / PTA estimation
 Use Extreme Bounds Analysis (EBA)
 (Leamer 1978, 1983, 1985)
13
Previous Approach To Model Uncertainty
In Trade Flow Theory

Ghosh and Yamarik (2004, JIE)
 First (and to date only) attempt to account for
model uncertainty in trade flow / PTA estimation
 Use Extreme Bounds Analysis (EBA)
 (Leamer 1978, 1983, 1985)

Conclusion: No evidence of trade creation or
diversion for any PTA! Relaxed extreme bounds pick
up only trade diverting PTAs!!
14
Extreme Bound Analysis
-- Assessment -
Lack of Statistical Theory Backbone
 Reduce model space artificially to avoid running all regressions. (Why?)
 Why has each model (no matter how terrible) equal weight…?
 What is the robustness criterion? (see growth theory)
 How relevant is the analysis if the model space holds billions of
models and the researcher searches only over less than 0.0000001% of
the model space?
 Severe omitted variable bias

Could be (and has been) considered data mining…
15
Bayesian Model Averaging (BMA)

BMA Intuition


Model Selection: estimating the performance of different
models in order to choose the best one.
Model Averaging: average predictions from different models
to achieve improved performance


Posterior Estimate: weighted average over all models, where
weights are given by model quality.
BMA is a) theory based, b) based on objective criteria, and c)
proof exists that BMA delivers best predictive performance
(Raftery 1995)
16
Bayesian Model Averaging:
Quick details for the math hungry
•
posterior probability for model Mk (the “weight”) is
p( D | M k ) p( M k )
p ( M k | D)  K
l 1 p(D | M l ) p(M l )
•
•
p( D | is
Mthe
, M k ) p( k (over
| M k )all
dregressors)
k ) integrated
k
 p( D |  k likelihood
Posterior Mean
E | D   k 0 E | D, M K  pM k | D 
K
D
Mk

is the data,
is a model in some model space M, K is number of models,
is a quantity of interest
Nice: priors largely wash out with 15000 observations
17
Econometric Framework

Basic Building Block: Gravity Equation
 Can be derived from a variety of trade models
Deardorff (1998)
 Successful in explaining implied trade flows (Frankel
Romer 1999)
log(T radeijt ) = a + b1 log(GDPit ×GDPjt ) + b2 log(Distanceij )
+ b 3Z ijt + b 4RT Aijt + b 5RT Ait + eijt
Zijt: proxies for trade costs / trade theory covariates
18
Definition of PTA Dummies
log(T radeijt ) = a + b1 log(GDPit ×GDPjt ) + b2 log(Distanceij )
+ b 3Z ijt + b 4RT Aijt + b 5RT Ait + eijt

Trade Creation dummies (PTAij)



1 only if both trade partners are members of a respective PTA
0 otherwise
Trade Diversion/Open Block Dummies (PTAi)


1 if one and only one trade partner is a member of a respective PTA
0 otherwise
19
Trade Theory Covariates (Zijt)

Geography


Historical Ties


Language, Common Colonizer, Colony
Exchange Rate / Trade Policy


Border, Remoteness, Landlocked, Island, Area
Sachs dummy, Currency Union, Floating FX rate, FX volatility
Factor Endowments / Development


GDP p.c.
Log Differences in



GDP p.c.
Education
Population Density
20
Data

Identical to “no effect” data of Ghosh and Yamarik (2004)
 Avg Bilateral Trade Flows, 186 countries, 14,522 observations,
3,420 bilateral trade pairs, five-year intervals (1970-1995)

12 major PTAs:
 Europe: EU, EFTA, EEA
 Pacific Rim: APEC, ASEAN, NAFTA, ANZCERTA
 Latin America: CACM, CARICOM, LAIA, AP (Andean
Pact), MERCOSUR
21
European PTA Members
Abbreviation
Name of RTA
Start
Member countries
EEA
European Economic
Area
1994
EFTA
European Free Trade
Arrangement
1960
EU
European Union
1958
Austria, Belgium, Denmark, Finland, France, Germany,
Greece, Luxembourg, Iceland, Italy, Ireland, Liechtenstein,
Netherlands, Norway, Portugal, Spain, Sweden, United
Kingdom.
Iceland, Liechtenstein (1991), Norway (1986), Switzerland
Former: Denmark (1960-72), United Kingdom (1960-72),
Portugal (1960-85), Austria (1960-94), Sweden (1960-94),
Finland (1986-94).
Austria (1995), Belgium, Denmark (1973), Finland (1995),
France, Germany, Greece (1981), Luxembourg, Ireland
(1973), Italy, Netherlands, Portugal (1986), Spain (1986),
Sweden (1995), United Kingdom (1973).
22
Results: Europe
Specification 1
Time Fixed Effects
Post.
Incl.
Prob.
Trade Creation
0.01
EEAijt
0.00
EFTAijt
0.00
EUijt
Trade Diversion Open Block
0.00
EEAit
1.00
EFTAit
1.00
EUit
Post.
Mean
Post.
Std.
Dev.
Post
Mean
in %
0.26
0.02
0.03
0.19
0.26
0.14
29%
2%
3%
0.01
0.35***
0.56***
0.08
0.05
0.04
1%
43%
75%
Clear, strong open Block Effects.
23
Pacific Rim PTA members
Abbreviation
Name of RTA
Start
Australia – New
Zealand Closer
Economic Relations
Trade Agreement
Asia Pacific Economic
Community
1983
Australia, New Zealand
1989
ASEAN
Association of South
East Asian Nations
1967
NAFTA
Canada-US Free Trade
Arrangement / North
America Free Trade
Agreement
1988
Australia, Brunei, Canada, China (1991), Chile (1994),
Taiwan (1991), Hong Kong (1991), Indonesia, Japan, South
Korea, Malaysia, Mexico (1993), New Zealand, Papua New
Guinea (1993), Peru (1998), Philippines, Singapore,
Thailand, United States, Vietnam (1998).
Brunei (1984), Cambodia (1998), Indonesia, Laos (1997),
Malaysia, Myanmar (1997), the Philippines, Singapore,
Thailand, Vietnam (1995).
Canada, United States, Mexico (1994).
ANZCERTA
APEC
Member countries
24
Results: Pacific Rim
Now significant trade
creation and diversion
and open block
Specification 1
Time Fixed Effects
Post.
Incl.
Prob.
Trade Creation
0.00
AFTAijt
0.01
ANZCERTAijt
1.00
APECijt
0.01
NAFTAijt
Trade Diversion, Open Bloc
0.03
AFTAit
1.00
ANZCERTAit
1.00
APECit
1.00
NAFTAit
Post.
Mean
Post.
Std.
Dev.
Post
Mean
in %
-0.22
0.89
1.48***
-0.89
0.54
0.96
0.15
0.84
-20%
144%
338%
-59%
0.17
-0.47***
0.55***
-0.63***
0.11
0.10
0.06
0.10
19%
-37%
73%
-47%
PTA’s matter a lot!
Puzzling high APEC…

Large Trade Diversion
for Nafta!



25
Latin American PTA Membership
Abbreviation
Name of RTA
Start
Andean Community /
Andean Pact
Central American
Common Market
Caribbean
Community/ Carifta
1969
LAIA
Latin America
Integration Agreement
1960
MERCOSUR
Mercado Comun Del
Sur
1991
AP
CACM
CARICOM
1960
1968
Member countries
Bolivia, Colombia, Ecuador, Peru, Venezuela (1973),
Former: Chile (1969-76)
Costa Rica (1963), El Salvador, Guatemala, Honduras,
Nicaragua.
Antigua and Barbuda, Bahamas (1983), Barbados, Belize
(1995), Dominica (1974), Guyana (1995), Grenada (1974),
Jamaica, Montserrat (1974), St. Kitts and Nevis, St. Lucia
(1974), St. Vincent and the Grenadines, Suriname (1995),
Trinidad and Tobago.
Argentina, Bolivia (1967), Brazil, Chile, Colombia (1961)
Ecuador (1961), Mexico, Paraguay, Peru, Uruguay,
Venezuela (1966).
Argentina, Brazil, Paraguay, Uruguay
26
Results: Latin America
Orthodox results of
Strong trade creation
And trade diversion
In Latin America
Time Fixed Effects
Post.
Incl.
Prob.
Trade Creation
0.01
APijt
1.00
CACMijt
1.00
CARICOMijt
0.91
LAIAijt
0.12
MERCOSURijt
Trade Diversion, Open Bloc
0.52
APit
0.85
CACMit
1.00
CARICOMit
1.00
LAIAit
0.79
MERCOSURit
Post.
Mean
Post.
Std.
Dev.
Post
Mean
in %
-0.05
2.25***
2.08***
0.46***
1.66
0.27
0.23
0.41
0.13
0.70
-5%
848%
702%
58%
424%



-0.19*
-0.18**
-0.74***
-0.40***
0.42**
0.06
0.05
0.07
0.07
0.12
-17%
-17%
-52%
-33%
52%



27
Summary part I

Using:
 identical data
 statistically sound approach to model uncertainty
 considers all possible models (including best model)
 Derive quality-weighted averaged estimates
 We overturn the bleak “PTAs don’t matter“
results and show that several have trade creating,
trade diverting and open block effects.
28
Results: Control Variables
Specification 1
Specification 2
Time Fixed Effects
Core Gravity
Economic
Policy
Development,
Factor
Endowment
Geography
History
log(DISTANCEij)
log(GDPit GDPjt)
log(gdpit gdpjt)
SACHSit+SACHSjt
CUijt
FLOATijt
VOLATILITYijt
abs(gdp_DIFF)
abs(DENS_DIFF)
abs(SCHOOL_DIFF)
BORDERij
REMOTEij
LANDLOCKij
log(AREAi AREAj)
ISLANDij
COMLANGij
COMCOLij
COLONYij
Post.
Incl.
Prob.
1.00
1.00
1.00
1.00
1.00
0.00
0.25
1.00
1.00
0.01
1.00
1.00
1.00
0.92
0.02
1.00
1.00
1.00
Post.
Mean
-1.19***
0.88***
0.55***
0.35***
1.40***
-0.01
0.006
0.18***
0.23***
0.02
0.53***
342***
-0.42***
-0.03**
-0.05
0.47***
0.77***
1.44***
Post.
Std.
Dev.
0.02
0.01
0.02
0.03
0.29
0.02
0.002
0.02
0.01
0.02
0.10
39.79
0.04
0.01
0.03
0.05
0.07
0.12
Post
Mean
in %
-1%
1%
1%
42%
307%
-1%
0%
0%
0%
0%
69%
342%
-34%
0%
-5%
60%
116%
320%
Time & Country-Pair Fixed
Effects
Post.
Post.
Post.
Post
Incl.
Mean
Std. Mean
Strong effects Dev.
from in
gravity
Prob.
%
0%
and
other country
specific
0.05
1.00 1.13***
1%
0.10
0.02
0.16
0%
Controls
1.00 0.13***
0.03
14%
0.01
-0.64
0.67
-47%
0.20
-0.05
0.02
The only variables that-5%
do not
0.01
-0.002 0.002
0%
are
1.00 receive
-0.31*** support
0.05
0%
0.75
0.25**
0.09
- FLOAT 0%
0.03
-0.06
0.04
0%
- ED DIFF
- ISLAND
29
The Importance of Fixed Effects
(Natural Trading Partners)


Hummels and Levinsohn (1995)
 Trade is largely specific to country-pairs
 We may not know why, or cannot account for it with our
covariates, no matter how many controls we include.
We introduce country pair fixed effects
 Isolate whether countries trade a lot with each other because
of PTA’s or because they are natural trading partners

Cheng and Wall, 1999 (EU, Nafta, Mercsur); Egger and
Pfaffermayr, 2003 (gravity only)
30
The Importance of Fixed Effects
(Natural Trading Partners)
Introduce Country Pair Fixed Effects
 Example: similarities in economic and social institutions,
such as corruption or rule of law, or simple economic
infrastructure such as telecommunications.
~
 log
AYconcrete
example
be France-Germany, with
   
Y   Z   PTA
  PTA would

excellent transport links (unobserved) and a PTA (observed),
vs. Afghanistan-Kazakhstan, with bad transport links
(unobserved) and no PTA (observed).

log Tijt
t
ij
1
it
jt
3
ijt
4
ijt
5
it
ijt
~
log Tijt   t   ij  1 log Yit Y jt   3 Z ijt   4 PTAijt   5 PTAit   ijt
ij
31
Europe
Trade Effects Purged of Natural Trading Partner Effects
Specification 1
Specification 2
Time Fixed Effects
Post.
Incl.
Prob.
Trade Creation
0.01
EEAijt
0.00
EFTAijt
0.00
EUijt
Trade Diversion Open Block
0.00
EEAit
1.00
EFTAit
1.00
EUit


Time & Country-Pair Fixed
Effects
Post.
Post.
Post.
Post
Incl.
Mean
Std. Mean
Prob.
Dev.
in %
Post.
Mean
Post.
Std.
Dev.
Post
Mean
in %
0.26
0.02
0.03
0.19
0.26
0.14
29%
2%
3%
0.00
0.03
1.00
0.06
0.52
0.66***
0.17
0.32
0.17
6%
69%
93%
0.01
0.35***
0.56***
0.08
0.05
0.04
1%
43%
75%
0.20
0.98
0.21
-0.14
0.26**
0.16
0.06
0.08
0.07
-13%
29%
17%


Now EU is Trade Creating
Strong Open Block Effect
32
Europe
PTA Flows Purged of Natural Trading Partner Effects

Only after controlling for natural trading pairs, the EU
is Trade Creating
 EU countries naturally under-trade relative to the
prediction of the standard gravity model. This
“missing trade” is a standard feature of the basic
gravity model, see e.g. Pollak (1996) or Rose (2004).

EU Open Block Trade Creation does not survive but is
simply due to similarities among trading partners
33
Latin America
PTA Flows Purged of Natural Trading Partner Effects
Time Fixed Effects
Post.
Incl.
Prob.
Trade Creation
0.01
APijt
1.00
CACMijt
1.00
CARICOMijt
0.91
LAIAijt
0.12
MERCOSURijt
Trade Diversion, Open Bloc
0.52
APit
0.85
CACMit
1.00
CARICOMit
1.00
LAIAit
0.79
MERCOSURit


Time & Country-Pair Fixed
Effects
Post.
Post.
Post.
Post
Incl.
Mean
Std. Mean
Prob.
Dev.
in %
Post.
Mean
Post.
Std.
Dev.
Post
Mean
in %
-0.05
2.25***
2.08***
0.46***
1.66
0.27
0.23
0.41
0.13
0.70
-5%
848%
702%
58%
424%
0.02
0.56
0.37
0.00
0.12
0.71
0.00
0.37
0.51
-0.19*
-0.18**
-0.74***
-0.40***
0.42**
0.06
0.05
0.07
0.07
0.12
-17%
-17%
-52%
-33%
52%
0.03
-0.17
0.10
0.39
-0.29
0.11
0.01
0.11
0.09
74%
0%
13%
0%
45%
-15%
0%
-25%
0%
11%
Counterintuitive Effects Disappear
All Effects are Natural Trading Partner Effects
34
Pacific Rim
PTA Flows Purged of Natural Trading Partner Effects
Time Fixed Effects
Post.
Incl.
Prob.
Trade Creation
0.00
AFTAijt
0.01
ANZCERTAijt
1.00
APECijt
0.01
NAFTAijt
Trade Diversion, Open Bloc
0.03
AFTAit
1.00
ANZCERTAit
1.00
APECit
1.00
NAFTAit
Time & Country-Pair Fixed
Effects
Post.
Post.
Post.
Post
Incl.
Mean
Std. Mean
Prob.
Dev.
in %
Post.
Mean
Post.
Std.
Dev.
Post
Mean
in %
-0.22
0.89
1.48***
-0.89
0.54
0.96
0.15
0.84
-20%
144%
338%
-59%
0.00
0.00
0.01
0.00
-0.22
0.30
0.14
0.60
0.39
0.92
0.12
0.65
-19%
35%
16%
81%
0.17
-0.47***
0.55***
-0.63***
0.11
0.10
0.06
0.10
19%
-37%
73%
-47%
1.00
0.10
1.00
0.99
0.40***
-0.18
0.23***
-0.31***
0.08
0.09
0.05
0.08
49%
-17%
26%
-26%
!



Counter Intuitive APEC Trade Creation vanishes
AFTA and APEC Trade Creation confirmed
NAFTA Trade Creation is maintained but smaller
35
Pacific Rim
PTA Flows Purged of Natural Trading Partner Effects

NAFTA 26% trade diversion?
36
Trade Creation/Diversion
Importance of Comprehensive Approach

Most PTA Studies estimate Trade Creation for one/a few PTAs

Marginal Effects (too diverse to average) vs Global Effects

Most studies don’t distinguish Trade Creation/Open Block

Example1: Two countries are in two different PTAs: what matters for
trade flows is the Open Block effect / Trade Diversion of BOTH PTAs
(70 countries, most high income countries are in PTAs)

Example2: All countries (PTA or not) get open block effects from all
other PTA’s. So: pairwise trade diversion may turn into multilateral trade
creation!
37
Trade Creation/Diversion
NAFTA/APEC ANZCERTA/APEC AFTA/APEC
NAFTA
0%
ANZCERTA
-26%
AFTA
10%
APEC
EU
EFTA
0%
49%
0%
0%
49%
=((1+49%)*(1-26%))-1
Rest of APEC
EU
-26%
-7%
0%
=((1+26%)*(1-26%))-1
20%
EFTA
Rest of the World
=((1+29%)*(1+26%)*
(1-26%))-1
-7%
=((1+26%)*(1-26%))-1


87%
0%
26%
93%
=((1+26%)*(1+49%))-1
63%
=((1+26%)*(1+29%))-1
26%
142%
63%
=((1+26%)*
=((1+26%)*(1+29%))-1
(1+49%)*(1+29%))-1
87%
26%
29%
0%
0%
29%
=((1+26%)*(1+49%))-1
Most PTA interactions are Trade Creating
EVEN NAFTA trade can be trade creating with Nafta non members
38
Conclusions




1) Address Model Uncertainty:
 obtain correct PTA’s effects: PTA’s do matter (!)
2) Purge Natural Trading Partner contamination.
 Eliminate Counterintuitive Results (!)
 Isolate actual PTA effects: (most PTA effects smaller)
3) Comprehensive Approach
 Identifies crucial interactions that can overcome “false trade
diversion” (US / ASIA)
Overwhelmingly, PTAs are stepping stones to freer trade
39