1. travel
2. tourist destinations
3. mexico and central america

Not all Free Trade Agreements have the same Advantages ∗
Thomas Zylkin†
November 1, 2014
Abstract
Using NAFTA as an illustrating example, I show that heterogeneity in the effects of free trade
agreements (FTAs) both within and across agreements is not well understood. Not only has
NAFTA reduced trade frictions significantly more than other FTAs, but each NAFTA member
country has enjoyed larger gains in access to the other two markets in some industries over
others. Most notably, I observe a crucial role for reductions in unobservable non-tariff barriers
in determining these patterns. I then show these overlooked sources of heterogeneity have important first-order implications for both prices and welfare in general equilibrium. Using tariffs
to project welfare for example not only greatly underestimates the overall welfare gains for all
three countries—Mexico’s in particular—but also overstates the benefits for U.S. producers. I
also show, using a novel measure of the gains from specialization of factors of production, that
these first-order differences can imply even larger biases in more intricate welfare calculations.
JEL Classification Codes: F13, F14, F15
Keywords: Free Trade Agreements, International Trade, Gravity
∗ Special thanks first and foremost to the members of my dissertation committee (Constantinos Syropoulos, Yoto V.
Yotov, Vibhas Madan, Eric Bond, and Irina Murtazashvili) for their help in bringing this paper to fruition and for being
valuable resources for me throughout my career. I also wish to give acknowledgments to the following people who have
also been very generous with both their time and comments: Delina Agnosteva, Ingo Borchert, Lorenzo Caliendo, Peter
Eppinger, Mario Larch, Philip Luck, Roger McCain, Lorenzo Rotunno, Nicholas Sly, Matthew Weinberg, and Benjamin
Zissimos. I also thank seminar participants at the 2014 Institutions, Trade, and Development (InsTED) Workshop in
Eugene, Oregon as well as at the Drexel University School of Economics Seminar Series in Philadelphia, Pennsylvania.
All errors are mine.
† Contact information: Department of Economics, Drexel University.
1
1
Introduction
We do not really know what it is free trade agreements (FTAs) do for trade. How much will the proposed Trans-Atlantic Trade and Investment Partnership (TTIP) between the U.S. and E.U. increase
trade between relative to other FTAs? How will the consequences be different for individuals living in Bulgaria as opposed to those living in the U.K. or in the U.S.? Despite the longstanding,
consistent interest in the trade literature in the general equilibrium welfare effects of trade integration,1 the answers to these questions remain elusive. There is no consistent way of characterizing
what effects other FTAs with similar provisions to TTIP have had on trade barriers, let alone what
each individual FTA may mean for different nations.
This paper works towards addressing these issues by offering a tractable methodology by
which “heterogeneity” in the effects of FTAs, both within and across agreements, can be easily
identified and analyzed. Using a combination of panel methods and structural “gravity” estimation techniques,2 I am able to infer directly from the trade data, ex post, what effects a particular
FTA has had on trade barriers for each of its member countries. Similar ex post approaches, starting with seminal empirical work by Baier and Bergstrand (2007), have illustrated many new facts
about the “average treatment effect” of FTAs in recent years. These contributions include studies
on the stability of FTA effects over time (Baier and Bergstrand, 2009), differences in FTA effects
across industries (Anderson and Yotov, 2014), and differences between FTAs and other types of
Economic Integration Agreements (Baier et al, 2013).3 However this literature has not yet considered the substantial deviations from the average that may occur both within and across individual
FTAs. Using NAFTA as my illustrating example,4 I show that this heterogeneity is not well understood. Different FTAs can have significantly different effects on trade. And even within FTAs there
exist significant asymmetries in trade barrier reductions which have not been studied empirically.
Most notably, I demonstrate that NAFTA has had very different effects on trade than would
be implied by its associated tariff reductions. With few exceptions,5 ex post analyses of specific
trade deals, and of NAFTA in particular, continue to rely on tariffs to summarize all potential
heterogeneity in changes in trade barriers. If two partner countries have relatively high initial
tariffs, the reasoning goes, they will experience larger increases in trade from an agreement than
1 Viner
(1950) originates the current literature on the welfare analysis of trade integration. This literature is not
small. Baghawati, Krishna, and Panagariya (1999) summarize many of the major theoretical developments since Viner.
Current trends in quantitative welfare analysis have roots in two major papers, Eaton and Kortum (2002) and Anderson
and van Wincoop (2003). Anderson and Yotov (2014) and Caliendo and Parro (2014) have recently added important
methodological innovations. Head and Mayer (2014) and Costinot and Rodriguez-Clare (2014) survey other recent
developments in quantitative methods.
2 “Gravity” is a general empirical model for estimating trade frictions that is ubiquitous in both the empirical trade
literature and in current general equilibrium analysis of “gains from trade”. Head and Mayer (2014) provide a comprehensive background.
3 The broader empirical literature on FTAs dates back to Tinbergen (1962). More recent papers include Frankel et. al
(1995), Magee (2003), Ghosh and Yamarik (2004), and an earlier series of papers by Baier and Bergstrand (2002, 2004).
Frankel (1997) and Baier and Bergstrand (2007) review the literature.
4 The North American Free Trade Agreement (“NAFTA”) entered into force between the U.S., Mexico, and Canada
on January 1st, 1994.
5 Shikher (2012) is an important exception, as noted below.
2
other FTA pairs. Likewise, if one partner has higher tariffs on a particular good, its imports of
that good should increase more than its exports. This logic is not wrong, rather incomplete. FTAs
have many provisions besides tariff agreements that believed to affect trade (provisions promoting
foreign investment for example). Unlike tariff reductions, the effects of these other provisions on
trade are not easily observable and quantifiable ex ante. It is only by ex post analysis of changes in
trade patterns that the total impact of an FTA on trade barriers can be fully appreciated.
My estimation methodology builds on Baier and Bergstrand’s (2007) original panel fixed effects techniques for identifying the effects of FTAs. The key mechanism in this framework is the
use of pair-specific fixed effects to control for each pair of countries’ inherent propensity to trade
and form trade agreements. I follow this method but introduce two key twists. Whereas Baier and
Bergstrand use only a single FTA dummy (intended to capture the “average treatment effect” of
FTAs), I separate out this dummy as follows: First, I allow NAFTA to affect trade differently than
all other FTAs; Second, I allow for NAFTA to affect trade barriers in a given industry asymmetrically for each different trading pair. I thus allow for the possibility that, for instance, NAFTA may
have lowered trade barriers for Canadian Food producers trying to sell in Mexico more than it did
for Mexican sellers trying to sell Food products in Canada. In this way, I am able to document
significant new empirical facts about a trade agreement that is now 20 years old and has been
frequently studied in the trade literature.
For example, I find strong evidence that NAFTA has promoted trade significantly more than
other FTAs have.6 In addition, I confirm that NAFTA has had strongly asymmetric effects on
trade barriers within individual industries. Mexico’s Metals producers for instance have received
substantially more access to Canadian and U.S. import markets than vice versa; the opposite has
occurred for Food trade however. Lastly, I document that these asymmetries differ significantly
from what would have been expected to happen based on observed tariff reductions; it follows that
so-called “non-tariff barriers” have played a large, underappreciated role in determining NAFTA’s
specific influence on trade.
This last finding is particularly notable in light of current trade debates because other work
that has tried to evaluate each country’s benefits from NAFTA has tended to place too much emphasis the pre-NAFTA tariff differences between the U.S. and Mexico. Whereas projections based
on tariffs alone would have suggested that U.S. producers should have gained much more access
to Mexican import markets than they received in return, the actual changes in trade frictions I
identify tell a different story. Overall, NAFTA coincided with sizeable 63.1% decrease in barriers
to U.S. import markets for Mexico’s exporters, compared with a 45.9% decrease in barriers against
U.S. products headed in the other direction. This underappreciated growth in imports from Mexico may be a leading cause of the current skepticism towards TTIP in the U.S. “In all the time I’ve
been in Congress, I’ve never seen a trade bill that benefits the American producer or the American worker,” U.S. Congresswoman Louise Slaughter (D-NY) recently declared. “People are sick
and tired of the one way trade deal.”7 My results do not support these complaints of “one way
6 Cipollina
7 “Specter
and Salviticci (2010) document a similar finding via meta-analysis.
of NAFTA hurts Obama’s trade dreams”, Politico, February 19th, 2014.
3
trade” per se, but they do confirm that the realization of NAFTA has differed substantially from the
prevailing narratives available at the time.
An important feature of my empirical approach is that it is derived from a structural general
equilibrium gravity model, of the type detailed in Head and Mayer (2014). This approach enhances
the quality of the empirics in two appreciable ways. First, the general equilibrium structure informs the identification by providing flexible ways to account for all other tendencies toward trade
not directly related to trade costs (e.g. comparative advantage). Implicitly, the logic of structural
gravity assumes that trade costs are the driving force behind relative export decisions, i.e. why
origin A ships more goods to destination B as opposed to an otherwise-similar destination C. Time
variation in relative trade flows then identifies the effects of a free trade agreement on trade costs.8
Second, changes in all other key moments of the model (e.g. prices, production, and expenditure) can be solved for using the changes in trade costs obtained from the estimation. I am thus
able to follow Anderson and Yotov (2014) in translating my panel estimates for FTA effects directly
into implied price and welfare effects of NAFTA for every country in the data (including NAFTA’s
effect on non-NAFTA countries). These price and welfare effects based on my estimates can then
be compared directly with alternative welfare calculations based on either an “average” FTA effect
or changes in tariffs.
My welfare analysis shows the difference between NAFTA’s actual effects on trade patterns
and the implied effects based on existing methods is not trivial. Using either an average FTA
effect or observed tariff changes significantly underestimates the magnitude of welfare gains for
all three countries. Furthermore, using tariffs also fails to identify how these how these gains are
distributed, both within and across its member countries. For demonstration purposes, I use a
simple multi-sector endowment economy model to test NAFTA’s first-order effects on prices and
welfare for each economy, under different assumptions about its effects on trade barriers. These
initial price effects in turn have direct implications for more elaborate models commonly used to
quantify “gains from trade”.9 The endowment economy setting reveals the following: i) implied
welfare gains for all three NAFTA members are 50% to 100% as large as would be implied by an
average FTA effect and three times as large as would be found using tariff reductions; ii) the size
of Mexico’s implied gains from NAFTA in particular (2.37% vs. 8.87%)10 are underestimated using
tariffs; iii) the nature of the U.S.’s gains from NAFTA tend to be mischaracterized: the U.S. has
8 This
approach admittedly requires taking the (expansive) view that “trade costs” have a large “conceptual” component related to information asymmetries and/or network biases. For more on the surprising size of these conceptual
barriers to trade, see Head and Mayer (2013).
9 Table 1 in Head and Mayer (2014) provides a (non-exhaustive) overview of these “more elaborate” models. These
frameworks include allowances for monopolistic competition and increasing returns (Krugman, 1980; Wei, 1996), endogenous firm-level export entry (Melitz, 2003; Helpman et al., 2008), trade in intermediates (Eaton and Kortum, 2002;
Caliendo and Parro, 2014), and variable firm mark-ups (Melitz and Ottaviano, 2008; Arkolakis et al., 2012a; Behrens
et al, 2014). Costinot and Rodriguez-Clare (2014) demonstrate how many of these elements can be combined within a
single framework.
10 As is often discussed in the trade literature (see Arkolakis et al., 2012b), the magnitudes obtained for “gains from
trade” are highly sensitive to the assumptions used for the “trade elasticity” (i.e. the senstivity of trade to changes
in trade costs). I use values from Broda et al. (2006) for these elasticies. Using other elasticity values would affect
absolute magnitudes of welfare gains but would not significantly affect relative magnitudes across the different ways
of measuring NAFTA’s effects. I stand on the latter as my main result.
4
benefited mostly from an increased ability to buy goods from Canada and Mexico, not from being
better able to sell its own goods to the other two markets.
In addition, in order to demonstrate more specifically how these first-order price effects impact
higher-order welfare channels in other trade models, I develop a novel measure of the “gains from
specialization” and apply it to the case of NAFTA. The construction of this measure is straightforward. I assume the presence of a single factor of production (labor), which is allocated across several different monopolistically competitive industries in each country. Comparing welfare results
from the endowment economy model (where the allocation of labor is held fixed) with an alternate model where labor is allowed to move across industries (in response to changes in relative
prices, as in a Ricardian model) directly reveals additional gains due to the resulting re-allocation.
These “specialization gains” too are inherent to a wide class of trade models (see Costinot and
Rodriguez-Clare, 2014, Section 3.3), but have not been studied specifically. Evidence from NAFTA
shows that gains from specialization are very sensitive to the patterns of heterogeneity identified in
my empirical results. For example, I find that wholly 40% of Mexico’s implied gains from NAFTA
are due to specialization when I allow for general asymmetries in NAFTA’s effects. This ratio then
dwindles to only 9% when I strictly consider the effects of tariffs.
Due to my focus on NAFTA in particular, these welfare findings resonate with the recent literature that has tried to re-evaluate the policy impact of NAFTA using progressively more updated
tools for analyzing gains from trade. Caliendo and Parro (2014) for example simulate the effects of
NAFTA in a calibrated model with cross-sectoral input linkages that explains a significant portion
of the aggregate changes in country-level exports and imports that occurred post-NAFTA. Their
work builds on the analyses performed by Krueger (1999), Anderson and van Wincoop (2002), Romalis (2007), Shikher (2012), and others to try to characterize NAFTA’s imprint on each member’s
welfare. These papers generally do not look at how bilateral frictions between members may have
been affected along other dimensions besides tariffs.11
By contrast, my empirical methodology flexibly accounts for all sources in how NAFTA affected the pattern of trade ex post between each pair of countries across different sectors. This
flexibility is particularly important for characterizing the impact of modern FTAs because, since
NAFTA, FTAs have been becoming increasingly expansive with regards to hard-to-quantify nontariff provisions. An appraisal of “NAFTA at 20” earlier this year by the U.S. Congressional
Research Service (Villarreal and Ferguson, 2014) reserves special praise for the lasting influence
NAFTA has left on a “new generation of trade agreements” as a model for how to incorporate,
among other things, guarantees on freedom of investment, intellectual property rights, and investment property rights, and on the harmonization of product standards and customs procedures. In these terms, TTIP will likely mark an even bolder landmark. CEPII’s recent study on
TTIP (Fontagn´e et al., 2013) notes that average tariffs between the U.S. and E.U. are currently only
11 Shikher
(2012) is an important exception in that he attempts to infer tariff equivalents of non-tariff barriers using
the Nicita and Olareagga (2007) “Trade, Production, and Protection” database. However his estimates still show much
higher initial liberalization for Mexican imports than for Mexican exports, which is contrary to the results I obtain for
the actual trade impact of NAFTA.
5
around 2%-3% to start with—clearly both entities envision considerable gains from reducing “unnecessary regulatory differences”, “restrictive licensing”, and other non-tariff barriers to trade.12
Anticipating what these stated goals may mean for trade requires more focused tools for analyzing trade integration: ongoing work in the same vein as this paper (Yotov and Zylkin, 2014) will
explore how heterogeneity in FTA effects may depend on both the wording of agreements as well
as the characteristics of the countries who sign them.
The following section will briefly lay out my empirical model and then introduce and discuss
my main results. In subsequent sections, I offer some discussion of possible explanations for the
patterns I observe, present some potentially valuable avenues for future work, and demonstrate
the implications for welfare analysis. The last section adds concluding remarks.
2
Estimation Approach
2.1
Methodology
The “structural gravity” equation, as generalized by Head and Mayer (2014), naturally motivates a
panel fixed effects estimation strategy for identifying the impact of free trade agreements on trade.
Fixed effects gravity models of this type have been widely used in the literature for estimating the
effects of FTAs. The general framework assumes an R country world with K sectors and costly
trade in differentiated goods within each sector. Exports from i to j in sector k (Xijk ) then can be
expressed via the following gravity equation, which will explicitly motivate the estimation that
follows:
Xijk =
k
Yik Ej
·
· φijk .
Ωik Φkj
(1)
Ekj and Yik here are, respectively, j’s expenditure on industry k and the value of i’s production
in k. The longstanding logic of “gravity”, which dates back to Tinbergen (1962), is that trade is
increasing in the size of the two countries (i.e. E and Y) and decreasing in the trade costs between
them, which here would be reflected in the bilateral parameter φijk . To complete the analogy to
Newtonian gravity, φijk can be said to vary inversely with how far apart i and j are geographically.
Also playing an important role in (1) however are Φkj and Ωik , which themselves have the following structural interpretations:
Φkj
=∑
i
Ωik
=∑
j
12 See
φijk
Ωik
φijk
Φkj
· Yik
(2)
· Ekj .
(3)
USTR (2014) for a full list of these and other stated objectives for the U.S.
6
Intuitively, these so-called “multilateral resistance” terms index the total incidence of trade costs
on an individual country’s ability to access world markets, both on the buyer side (in the case of
Φkj ) and on the seller side (in the case of Ωik ). The more easily a producer in i is able to sell to world
markets in general, the less inclined he will be to sell to any one particular destination j for any
given level of bilateral trade costs φijk . A similar logic applies for buyers: better access to sellers
around the world all else equal makes them less likely to buy from any one particular exporter i.
These structural terms were originally introduced in Anderson and van Wincoop (2003), but are
common to a surprisingly wide class of models that fall under the heading “structural gravity”
and have different structural interpretations in each case.13
The advantage of presenting the trade model in this particular way then is that it is very general; thus it will me to make claims about both identification and welfare implications that will generalize across many different theoretical settings. In particular I use a multi-sector “Dixit-StiglitzKrugman” monopolistic competition model in order to simulate general equilibrium.14 However
the system defined by (1)-(3) can also be used to describe other gravity models founded on (for instance) within-industry comparative advantage (Eaton and Kortum, 2002), national-level product
differentiation (Anderson and van Wincoop, 2003), or variable firm productivity and endogenous
export entry (Helpman, Melitz, and Rubinstein, 2008).
For empirical purposes, my main parameter of interest is φijk , the parameter reflecting how trade
costs directly affect trade between i and j. Without loss of generality φijk ∈ (0, 1) can be thought of
as the amount of “market access” that sellers in i enjoy when attempting to sell their variety of
good k to import market j: when trade integration lowers trade barriers, φijk increases and trade
in turn increases proportionately with the change in market access, all else equal. The empirical
question I am looking to examine is how bilateral market access depends on the presence of a free
trade agreement between i and j, and how these market access effects may vary within the same
agreement. I will refer continually in this paper to %∆φijk − 1 as the “amount by which market
access increased”
in
industry k as the result of an agreement. Without loss of generality, I can also
call 1 − %∆ 1/φijk as the “amount by which barriers to market access fell”.
An important point about estimating values for φijk in this setting is that the combined system
(1)-(3) is “modular” (or “separable”). That is, even though the values for production and expenditure for each sector Yik and Ekj depend on what occurs across all sectors in general equilibrium, if I
simply take these terms as given, it follows from (1)-(3) that I can treat the φijk ’s in each individual
industry k as an independent set of parameters to be estimated separately, with no cross-equation
restrictions across industries. That is not to say that I am in any way restricting values for Y and E
to be fixed, only that the modularity of these structural gravity models allows for production and
budgeting decisions to be made at an “upper level” of the model, such that decisions over where
expression for Ωkj , (3), is usually not shown in the presentation of these other models. Nonetheless, Head and
Mayer (2014) show it is a general result that follows from an accounting identify for any model where (1) and (3) already
hold.
14 Head and Mayer (2014) credit Wei (1996) as being one of the earliest authors to derive this particular flavor of
gravity, which has its roots in Krugman’s (1980) original monopolistic competition framework.
13 The
7
and how to source varieties within an industry can take those values as given. I explain how the
use of time-varying exporter- and importer- fixed effects allows me to focus on this “lower level”
of the purchasing problem for each individual industry in my development of the econometric
specification below.
Following Baier and Bergstrand (2007), I assume φijk can be specified in the following manner
for each sector:
φijk = e βFTAij +δ
kZ
ij
,
(4)
where FTAij,t is an indicator variable (or set of indicator variables) reflecting whether i and j have
an FTA at time t and Zij is a set of controls for inherent bilateral characteristics assumed to have
some effect on trade (e.g. the distance between i and j, whether they share a common language,
whether they have a prior colonial relationship, presence of a common border, etc.) Together, (1)
and (4) specify my baseline estimating equation for trade in each industry k:
k
k
k
Xij,t
= exp ξ i,t
+ ψkj,t + ηijk + βk FTAij,t + eij,t
.
(5)
k and ψk are exporter and importer fixed effects that strip away the effects of all marketξ i,t
j,t
level variables that affect trade in industry k through channels other than through the amount of
direct bilateral market access. Note that these terms need to be time-varying: as noted, the Φkj and
Ωik terms in the structural model depend endogenously on the system of φijk terms in equilibrium
via (2)-(3). Furthermore, since these terms reflect buyer and seller prices, they in turn will have
implications for Ekj and Yik as well.15 It is important then that changes in φijk are identified not
just by changes in bilateral trade flows, but rather by changes in bilateral trade flows relative to
changes in each country’s unilateral tendency towards exporting and importing goods of type
k. When a country shifts resources towards production of good k for example, that shift should
generate increased exports of k to all destinations, not to any one destination in particular.
The (symmetric) pair-specific fixed effect term ηijk is meant to capture all time-invariant bilateral
relationships between i and j that influence trade (effectively absorbing δk Zijk in (4)). In panel estimation terms, due to the presence of ηijk , βk essentially serves as a “within” fixed effects estimator
for the effect of an FTA on exports from i to j for goods of type k.16 As Baier and Bergstrand (2007)
demonstrate—adapting the panel identification methods discussed in Wooldridge (2002)—the use
of pair-specific fixed effects in a panel gravity setting is a simple-to-apply procedure for identifying
the average treatment effect of FTAs and this approach has become standard in the literature. In
accordance with Santos Silva and Tenreyro (2006)’s recommendations for minimizing bias in gravity estimations, I will use the Poisson Pseudo-Maximum Likelihood (PPML) estimator to estimate
(5).17
15 I
characterize these linkages in more detail when I discuss welfare implications later in the paper.
construction, these pair fixed effects are symmetric. I relax this restriction later when I introduce directionspecific effects.
17 A more natural approach, also common in the prior literature, would be to estimate (5) via OLS. Santos Silva and
16 By
8
Breaking with the Baier and Bergstrand approach however, for my main results I will allow
FTAij,t to vary by agreement and, subsequently, by the direction of trade flows. Specifically, I focus
on NAFTA as a suitable example to show that FTAs may affect market in one direction differently
than they affect market access in the opposite direction. To do this, I split the single FTAij,t term in
(5) into a set of variables, as shown below:
k
k
k
Xij,t
= exp ξ i,t
+ ψkj,t + ηijk + β0,k FTA0ij,t + β N,k N AFTAij,t + eij,t
.
(6)
Here, the superscript “0” on FTA0ij,t is meant to indicate that β0,k is now measuring the average
effect of all FTAs aside from NAFTA. β N,k is then measuring all FTA effects that are associated
specifically with NAFTA. I can then split the N AFTAij,t variable even further in order to isolate
directional effects. For example, I will allow N AFTACAN MEX,t to be a single dummy for postNAFTA industry k exports from Canada to Mexico and N AFTA MEXCAN,t to be a separate dummy
for post-NAFTA flows in the other direction. NAFTA is a three country agreement, so there will
be 6 directional NAFTA effects to measure in all for each sector, plus the FTA0ij,t term to control for
the average effect of all other FTAs in effect.
In this last case, I need to be careful. The pair-wise fixed effect ηijk is intended to identify the
average effect of an FTA on average trade barriers in industry k for a given pair. However when
I allow NAFTA’s effects to be directional I am now for example interested in the specific effect
of NAFTA on trade frictions for Canadian Food producers trying to sell in US markets, and vice
versa. If trade barriers for Canadian Food producers selling in the US are different than those faced
N,Food
N,Food
will in part reflect this initial
by US producers trying to sell in Canada, then βUSCAN
and β CANUS
difference in trade barriers, rather than identifying the differences in how the FTA played out.
As such, I write down this last empirical model as follows,
k
k
k
~ ij,t + eij,t
Xij,t
= exp ξ i,t
+ ψkj,t + ηijk + ~ηijk + β0,k FTA0ij,t + ~β N,k N AFTA
,
(7)
where the “arrow” superscript indicates a set of effects that is allowed to vary by direction. The
additional effects ~η k are only in play for flows between the US, Canada, and Mexico. ~β N,k , the set
ij
of directional NAFTA effects, by definition also varies with the direction of flows.
This same simple procedure could be easily repeated for any FTA or set of FTAs in order to
identify direction-specific effects. NAFTA is an especially useful illustrating example for my purposes however, not just because of its continuing notoriety in current trade policy debates, but
also because its three country structure will offer the opportunity to make unique inferences about
the observable patterns of FTA effects. Furthermore, limiting the analysis to just NAFTA will also
allow me to test the more general hypothesis that some individual FTAs may promote trade more
than others.
Tenreyro however show that log-linearizing (5) introduces an important source of bias due to heteroskedasticity in eij
and measurement error in trade flows, which PPML estimation helps to minimize.
9
2.2
Data
The data used here builds on the data set used in Anderson and Yotov (2014). This data set spans
the period 1990 until 2002 for a sample of 40 individual countries plus an aggregate “Rest of the
World”, for a total of 41 trading regions in all. The main source for trade flows is the CEPII “TradeProd” data base, supplemented with data on exports from UN COMTRADE, which can be accessed using the WITS World Bank trade service. The original data uses observed trade flows from
every 4 years only—that is, 1990, 1994, 1998, and 2002. I then add additional data—referring to the
original sources and construction methods, and using the Anderson and Yotov data to interpolate
missing and unreasonable values where possible—for the years 1992, 1996, and 2000, such that
the full data set is for every two years. The reason why I do not include every year is because,
as Cheng and Wall (2002) point out, performing fixed effects gravity estimations over consecutive
years may fail to address the fact that trade patterns may not adjust right away to changes in trade
costs. The number of observations for each of the main results shown then is 11,765—i.e. 1,681
trading pairs (the square of the number of regions) times the 7, number of years in the data.
The level of aggregation for the sectoral results is the ISIC (Revision 2) 2 digit level, which is
comprised of 9 2 digit manufacturing industry classifications: 31. Food and Beverages, 32. Textiles,
33. Wood Products, 34. Paper Products, 35. Chemicals, 36. Minerals, 37. Metals, 38. Machinery,
and 39. Other Manufacturing. However since some countries report some Machinery products
under Other Manufacturing and vice versa in their output data, these two sectors are combined
into a single “Manufacturing” category in the final trade data.
A key feature of this data is the inclusion of internal trade flows. The inclusion of internal
trade values is crucial for my purposes because one cannot perform a true general equilibrium
analysis without some form of accounting for how domestic sales will respond to changes in trade
costs. These flows are constructed as the difference between total sectoral output and total sectoral
exports to all trading partners. Because exports are measured on a “gross” (rather than valueadded) basis, the data likewise uses gross output data for these purposes. Like the trade data, the
output data is mainly taken from TradeProd and then supplemented with another source, in this
case the United Nations UNIDO Industrial Statistics (“IndStat”) database. Missing internal trade
values have been extrapolated by comparing the share of internal trade with respect to output in
non-missing sectors. Each country’s total expenditure on a given industry, which plays a role in
the welfare analysis, can then be calculated by adding together internal trade and total imports.
The data on FTAs is based mainly on the original Baier and Bergstrand (2007) data set, which
is updated with some data on additional agreements taken from the WTO’s web site. Because the
trade data begins in 1990, only FTAs that entered into effect after that year are coded. NAFTA,
which went into effect in 1994, is obviously included, but the Canada-U.S. Free Trade Agreement
of 1987 (which preceded NAFTA) is not.18 Overall, there are 252 country pairs in the data that
entered into either a free frade agreement or customs union during the period under study.19
18 See
Table 1 in Anderson and Yotov (2014) for details on the FTAs included.
Unions are included as FTAs for these purposes.
19 Customs
10
Table 1: Industry-Level Results: NAFTA vs. Other FTAs
Food
Textile
Wood
Paper
Chemicals
Minerals
Metals
Machinery
A. Sectoral FTA Estimates
All FTAs
0.451
0.713
0.006
-0.017
0.228
0.192
0.447
0.467
(0.076)∗∗
(0.120)∗∗
(0.072)
(0.057)
(0.041)∗∗
(0.065)∗∗
(0.062)∗∗
(0.118)∗∗
B. Individual FTA Estimates (NAFTA vs. All Other FTAs)
NAFTA
All Other FTAs
0.504
1.166
0.140
0.371
0.462
0.604
0.339
0.613
(0.067)∗∗
(0.063)∗∗
(0.095)
(0.032)∗∗
(0.030)∗∗
(0.051)∗∗
(0.136)∗
(0.144)∗∗
0.465
0.646
-0.038
-0.050
0.183
0.133
0.405
0.371
(0.082)∗∗
(0.118)∗∗
(0.052)
(0.032)∗∗
(0.056)∗
(0.060)∗∗
(0.082)∗∗
0.177
0.422
0.280
0.471
-0.067
0.242
(0.117)
(0.062)∗∗
(0.038)∗∗
(0.076)∗∗
(0.150)
(0.125)+
(0.068)
C. Significance Tests (NAFTA vs. the Average FTA)
NAFTA vs. Average
0.040
0.520
(0.103)
(0.107)∗∗
Robust standard errors, clustered by pair, are reported in parentheses. + p < 0.10 , * p < .05 , ** p < .01 .
All estimates are obtained with Santos-Silva and Tenreyro’s (2006) Poisson Pseudo-Maximum Likelihood estimator.
Following Baier and Bergstrand (2007), pair fixed effects are used to account for FTA endogeneity.
Time-varying exporter and importer fixed effects are used to control for the multilateral resistances.
3
Empirical Findings
I present my main evidence in Table 1. First, in Panel A, I estimate average sectoral FTA effects
using equation (5) in order to document how the effects of FTAs generally vary significantly across
industries.20 Notably there are two sectors, Wood and Paper, where it does not seem like FTAs
have impacted trade significantly. These initial results serve as a useful baseline for establishing
how much variation is missed out on by looking solely at average FTA effects.
Panels B and C then explore the implications of allowing for more specific FTA effects. In Panel
B, following equation (6), I separate out the individual effect of NAFTA from all other FTAs. In
both cases I control for the presence of all other FTAs. The estimates from Panel B reveal that the
effects of NAFTA are quite different from the average FTA estimate from Baier and Bergstrand
(β = 0.46), from the average sectoral FTA estimates from Panel A, and from the estimates of effects
of all other FTAs, excluding NAFTA (the “All Other FTAs” term in Panel B). For example, I find
that NAFTA has led to a significant increase in trade in Paper products among the three NAFTA
members, while the effect of all other FTAs is still negative, small, and marginally significant. Panel
C confirms the significance of these differences. These results lend support to my first hypothesis
for FTA-specific effects.
Then, in Table 2, I allow for country-specific and directional differences in the effects of NAFTA,
as in (7). Based on these estimates, I conclude that FTA effects are indeed directional. For example,
my estimates for Food reveal that NAFTA had strong positive effects on Canadian exports to Mexico (NAFTA Can-Mex= 1.614, std.err 0.165) and on US exports to Mexico (NAFTA U.S.-Mex=1.000,
std.err 0.163). However, I do not obtain statistically significant estimates of the effects of NAFTA on
20 Qualitatively,
the results are virtually identical to Anderson and Yotov (2014)’s results for this same specification,
and differ only because I include additional years in the data.
11
Table 2: Industry-Level Results: Directional NAFTA Effects
Food
Textile
Wood
Paper
Chemicals
Minerals
Metals
Machinery
Directional FTA Estimates (NAFTA)
NAFTA Can-Mex
NAFTA Mex.-Can.
NAFTA Can-U.S.
NAFTA U.S.-Can
NAFTA Mex.-U.S.
NAFTA U.S.-Mex
All Other FTAs
1.614
1.725
1.850
0.517
1.163
1.806
0.587
1.400
(0.165)∗∗
(0.125)∗∗
(0.211)∗∗
(0.116)∗∗
(0.091)∗∗
(0.163)∗∗
(0.114)∗∗
(0.133)∗∗
-0.249
1.544
0.754
1.769
1.931
1.472
1.674
1.291
(0.164)
(0.115)∗∗
(0.231)∗∗
(0.122)∗∗
(0.088)∗∗
(0.164)∗∗
(0.115)∗∗
(0.200)∗∗
0.595
1.348
0.209
0.081
0.412
0.643
0.330
0.441
(0.069)∗∗
(0.108)∗∗
(0.181)
(0.080)
(0.086)∗∗
(0.100)∗∗
(0.098)∗∗
(0.064)∗∗
0.510
0.832
-0.197
0.767
0.457
0.667
-0.028
0.481
(0.067)∗∗
(0.104)∗∗
(0.182)
(0.079)∗∗
(0.078)∗∗
(0.101)∗∗
(0.095)
(0.045)∗∗
-0.445
1.592
0.768
0.371
0.536
0.620
1.098
1.235
(0.162)∗∗
(0.128)∗∗
(0.176)∗∗
(0.104)∗∗
(0.064)∗∗
(0.153)∗∗
(0.104)∗∗
(0.207)∗∗
1.000
0.938
0.123
0.293
0.512
0.395
0.415
0.727
(0.163)∗∗
(0.132)∗∗
(0.148)
(0.099)∗∗
(0.060)∗∗
(0.152)∗∗
(0.102)∗∗
(0.119)∗∗
0.478
0.645
-0.041
-0.050
0.186
0.130
0.405
0.377
(0.081)∗∗
(0.118)∗∗
(0.068)
(0.052)
(0.032)∗∗
(0.056)∗
(0.060)∗∗
(0.079)∗∗
Robust standard errors, clustered by pair, are reported in parentheses. + p < 0.10 , * p < .05 , ** p < .01 .
All estimates are obtained with Santos-Silva and Tenreyro’s (2006) Poisson Pseudo-Maximum Likelihood estimator.
Following Baier and Bergstrand (2007), pair fixed effects are used to account for FTA endogeneity.
Time-varying exporter and importer fixed effects are used to control for the multilateral resistances.
Mexico’s exports to Canada and Mexico’s exports to the U.S. were actually negatively affected.21
The estimates for Food in Table 2 are also useful for pointing out a broader regularity in the
data that I wish to focus on. The NAFTA effects for U.S.-Mexico and Canada-Mexico trade are
much larger in magnitude than the effects for U.S. exports to Canada and vice versa. This pattern
suggests a loose correspondence between how much market access a country has gained as exporter vs. how much access to its own markets it offers in return. Here, the U.S. and Canada both
realize generally strong gains in market access on the export side, while we cannot say NAFTA has
done much for Mexican Food exporters.
Estimates from several other sectors suggest the same sort of regularity, with different orderings. Mexican Paper exports to Canada increase a prodigious amount, whereas other countries’
exports expand more modestly.22 We also see this same kind of pattern where one country seems
to generally gain the most most access to the other two markets in Metals and Machinery (again
Mexico in both cases).
I focus on illustrating these patterns to make a particular point that is important for the identification: one might be tempted to call the differences in how much trade increases I am identifying
as reflecting some measure of “comparative advantage”. Maybe these results are due to the U.S.
21 Trade barriers for Food are often dealt with seperately and given special exemptions in the negotiation of free trade
agreements and NAFTA is no exception. See Avery (1998) for an overview of the bargaining process that went on
between negotiators and agricultural special interests in solidifying the agreement.
22 I note however in the table below where I examine exporter-side effects versus importer-side effects that Mexico’s
large coefficient for Paper exports to Canada seems to be driven more by Canada’s opening its borders in general to
Paper products from both countries.
12
and Canada having comparative advantage in Food, Mexico having comparative advantage in
Machinery, and so on, the reasoning might go. But again I should call attention to the role of the
time-varying exporter and importer fixed effects in my specification. There is little question that
trade liberalization impacts relative prices across industries and that changes in relative prices in
turn cause factors of production to be re-allocated. But the structural model in (1) explicitly controls for the endogeneity of production and prices and, furthermore, the modularity of the model
allows me to separate the estimation of trade costs from these endogenous cross-sectoral linkages.
These estimates I am showing are indeed changes in market access as defined in the structural
model, controlling for all endogenous responses to trade within each country. This includes notions of comparative advantage.23
What can we say then about the possibility of “one way” effects in general? The industrylevel data is suggestive of certain patterns at a high level—Mexico seems to have had especially
strong gains as an exporter for instance (especially in the Textiles, Wood, Metals, and Machinery
sectors)—but to really make concrete statements one way or the other it is necessary to examine
results for aggregate trade. In Table 3, I show a number of specifications meant to highlight the
broad heterogeneity in FTA effects both between NAFTA and other FTAs and within NAFTA itself.
In column 1, I document that the average effect of FTAs on trade in my sample, across all industries combined is 0.401, which corresponds to an average increase in market access of e0.401 − 1 =
49.3%. Column 2 then replicates the specification from Table 1 Panel B. Unsurprisingly, NAFTA
appears to have promoted trade much more overall than the other FTAs in my sample—about
twice as much in fact.24 Furthermore, as I show explicitly in Column 3, the difference between
NAFTA and the average FTA is itself significant at the 5% level.
Columns 4 to 6 then isolate the directional impact of NAFTA on aggregate trade flows. The
patterns seen here for combined trade actually seem to disguise the amount of asymmetry we saw
in the industry-level results. While there is some definite variation across pairs (trade between
Mexico and Canada vs. trade between the U.S. and Canada) the asymmetry seen within pairs is
muted. In columns 5 and 6, we see for instance that each country generally tended to receive more
or less the same amount of access to its import markets that it offered to its partners in return.
One exception here—where we may be able to say there have been some asymmetric effects—
has been trade between the U.S. and Mexico. The results in column 4 show that Mexico overall
has received a e0.999 − 1 = 171% increase in access to U.S. markets, whereas the U.S. has received
an e0.615 − 1 = 85.0% increase in return.25 This last result is particularly interesting because it
goes directly against what would have been predicted to happen based on tariffs. As was well23 Note that this same argument also rules out any other type of general equilibrium response to trade cost reductions
(i.e. cross-sectoral re-optimization of input usages, as in Caliendo and Parro, 2014) as the source of these particular
variations. In such a model, a trade shock affects prices, which in turn feeds back into the quantity demanded for
products of a certain type. But again, all of this is captured in the structural model through the explicit representation
of production, expenditure, and prices, with all of these factors in turn being absorbed in the econometric model via the
use of time-varying exporter and importer fixed efffects.
24 e0.580 − 1 = 78.6%; e0.319 − 1 = 37.6%.
25 These increases in market access can also be described as a 63.2% decrease in US import barriers for Mexican products vs. a 45.9% decrease in Mexican import barriers for U.S. products, as stated in the opening remarks.
13
Table 3: NAFTA Effects, Aggregate Trade
(1)
Average FTA Effect
All Other FTAs
NAFTA
(2)
(3)
(4)
(5)
(6)
0.319
0.318
0.317
0.321
(0.042)∗∗
(0.042)∗∗
(0.042)∗∗
(0.042)∗∗
0.401
0.320
(0.065)∗∗
(0.042)∗∗
0.580
0.261
(0.095)∗∗
(0.100)∗∗
NAFTA Can-Mex
1.182
(0.073)∗∗
NAFTA Mex.-Can.
1.099
(0.074)∗∗
NAFTA U.S.-Can
0.470
(0.044)∗∗
NAFTA Can-U.S.
0.451
(0.044)∗∗
NAFTA Mex.-U.S.
0.999
(0.071)∗∗
NAFTA U.S.-Mex
0.615
(0.069)∗∗
NAFTA (Canadian exports)
0.443
(0.042)∗∗
NAFTA (Mexican exports)
1.076
(0.066)∗∗
NAFTA (U.S. exports)
0.507
(0.044)∗∗
NAFTA (Canadian imports)
0.417
(0.064)∗∗
NAFTA (Mexican imports)
0.845
(0.103)∗∗
NAFTA (U.S. imports)
0.586
(0.080)∗∗
Robust standard errors, clustered by pair, are reported in parentheses. + p < 0.10 , * p < .05 , ** p < .01 .
All estimates obtained with Santos-Silva and Tenreyro’s (2006) Poisson Pseudo-Maximum Likelihood estimator.
Following Baier and Bergstrand (2007), pair fixed effects are used to account for FTA endogeneity.
Time-varying exporter and importer fixed effects are used to control for the multilateral resistances.
14
publicized by the pro-NAFTA campaign at the time,26 Mexico’s tariffs on U.S. products were about
two-and-a-half times those on products shipped in the other direction. It would seem then that
focusing on tariff provisions does not necessarily do a good job of explaining how much market
access will change in response to the signing of a free trade agreement, a theme I will return to in
the ensuing discussion.
I continue with some further analysis of aggregate trade flows in order to demonstrate the
general robustness of the key messages of this paper. Table 4 presents a variety of alternate specifications and data samples meant to address potential concerns about the validity of my estimation
methods. In this table, Panel A shows how my findings for asymmetric effects within NAFTA
hold up across the different experiments and Panel B tests the finding from Table 3, column 3 that
NAFTA has promoted trade more effectively than other FTAs. For convenience I list again the
PPML results from Table 3 (specifically the results from column 6 and column 3, respectively),
such that any major deviations across other specifications can be easily verified.
I start by experimenting with using OLS instead of my preferred estimator, PPML. In this case
I do see some differences worth noting. As the results in column 2 seem to show, OLS generally
seems to underestimate the effects of FTAs compared with PPML. Still, within NAFTA there are
some clear asymmetries which are not inconsistent with the relative magnitudes shown in column
1. Indeed, the p value I calculate for the hypothesis that all directional NAFTA coefficients are
the same—not just for OLS, but for all specifications shown in Panel A actually—is approximately
0.000.27 In Panel B however, the test for whether NAFTA has promoted trade more than other
FTAs is narrowly outside the 10% significance range (p = 0.110). These deviations from the baseline results need not be considered cause for major concern however. As discussed in Santos Silva
and Tenreyro (2006), OLS is expected to perform poorly for estimating trade flows because the
process of log-linearizing equation (1) implicitly requires that the variance of the residual be proportional to the square of the conditional mean (otherwise, the estimates will not be consistent.)28
Furthermore, it is also necessary to drop all recorded “zeros” in the trade data—116 out of 11, 767
total observations in this case—in order to take logs. Nonetheless, comparing the OLS and PPML
results gives a useful basis for reconciling my results with those found in other work using OLS.
Column 4 then tests the implications of dropping the years 1992, 1996, and 2000, such that the
time-variation in the panel is reduced to every 4 years (specifically, the years 1990, 1994, 1998, and
2002). This experiment is meant to address concerns raised in Cheng and Wall (2002) about trade
being slow to adjust to changes in trade costs. This adjustment does not seem to diminish the
results, and if anything makes them slightly stronger.
A particularly important concern in this context is the possible presence of pre-existing trends
in the trade data that may have pre-dated NAFTA (or indeed, may have pre-dated the FTAs in
the data more generally). This may especially be a concern in the case of NAFTA, since Canada
26 See
for instance “NAFTA: Good for America”, Washington Post, July 4th, 1993.
test here is a standard Wald (F) test of whether all NAFTA variables in Panel A are equal.
28 PPML, which assumes the variance is directly proportional to the conditional mean, generally seems to be a better
fit to trade data.
27 The
15
and the U.S. already had an existing agreement in place (the “Canada-United States Free Trade
Agreement”, or simply “CUSFTA”) dating back to 1988. The baseline specifications in (6) and (7)
control for the effect of this earlier agreement on pre-NAFTA trade through the use of pair- and
direction- specific fixed effects—i.e. ηijk in (6) and ~ηijk in (7). However if the effects of CUSFTA were
still being “phased-in” gradually during the period 1990-1994, this may bias my main estimates. I
deal with these potential issues in two ways. First I adopt a set of pair-specific time trends, such
that trade costs between every pair of countries in the data are allowed to vary linearly over time.
While the inclusion of time trends overall seems to diminish the magnitudes of the coefficients I
obtain, this result is expected. After all, if FTA effects on trade are felt gradually over time, then the
time trend effects are likely picking up these phasing-in effects in addition to pre-existing trends.
Encouragingly, I still find significant differences within NAFTA (in this case driven mostly by
Mexico-Canada trade) and that NAFTA has promoted trade significantly more than the average
FTA.
A better strategy for untangling FTA phase-in effects from pre-trends in the data is to explicitly
allow for phasing-in (using lagged dummy variables) and also add controls for increases in trade
that may have occurred just before an FTA went into effect (using a lead dummy). Specifically,
I add 4- and 8-year lags of the “All Other FTAs” variable (FTA0 ), such that it FTAs become fully
operational over 8 years, in 3 stages.29 Importantly, adding lagged terms in this way not only
provides a reasonable means of controlling for trends in trade costs due to “outside” FTAs signed
by NAFTA countries—Mexico signed a whopping 35 FTAs during this period for example—it also
allows me to incorporate continuing effects from FTAs signed before 1990. CUSFTA for instance is
4 years old in 1992, and thus the coefficient on the 4-year lag for FTA0 can help explain potential
prior trends in U.S.-Canada trade. To then account for potential pre-trends more generally, I add 2year lead dummies for all FTA variables. These lead dummies allow me to distinguish, in the case
of NAFTA for example, large increases in trade that occurred starting with the implementation of
the agreement in 1994 from potential prior increases that may have occurred between 1990 and
1992.
Overall, the results for NAFTA variables in this final column are qualitatively very similar
to the baseline, which suggests that my original fixed-effects specifications in (6) and (7) do an
acceptable job of isolating the changes in trade costs that occur post-1994. The magnitudes on the
whole are slightly smaller, but we still see that Canada-Mexico trade has increased overall the most
of the 3 pairs, with Mexico-U.S. trade seeing the second largest increase, and Canada-U.S. trade
the least (but still with strong significance, despite the added controls for CUSFTA). We again not
see a significant asymmetry in U.S.-Mexico trade, but also now see more pronounced asymmetries
between the other pairs. I also note that in Panel B, the effect for NAFTA continues to significantly
stronger than that of the average FTA, where in this case I am comparing NAFTA to the combined
effect of an average FTA that has been fully phased-in over 8 years using a Wald test.30
29 This
approach to identifying phasing-in effects is the same as in Anderson and Yotov (2014) and is also fairly consistent with Baier and Bergstrand (2007), who use 5- and 10- year lags.
30 Specifically, the null hypothesis is that β N (the coefficient for NAFTA) is equal to the sum of β0 (“All Other FTAs”)
16
How concerned should we be though that the difference in magnitudes between column 1 and
column 5 indicates transient pre-trends in the data? In Panel B, I show that the estimated coefficient
for the lead dummy for NAFTA (“NAFTAt+1 ”) is indeed positive, if not significant (p = .141).
Interestingly however, the lead dummy for FTA0 (“All Other FTAst+1 ”) is not only positive, but
also strongly significant. What this result says is that it is actually a fairly general feature of the
trade data that agents within countries seem to be able to anticipate FTAs before they go into effect
and consistently begin investing in trade opportunities just before the actual signing occurs. It
follows then that my estimates, even though they are large to begin with, could conceivably be
underestimating the true trade impact of NAFTA.
Table 4: Robustness Checks, Aggregate Trade
PPML
OLS
Every 4 years
W/ time trends
W/ leads and lags
A. Directional FTA Effects (NAFTA)
NAFTA Can-Mex
1.182
0.716
1.280
0.535
1.134
(0.073)∗∗
(0.125)∗∗
(0.085)∗∗
(0.055)∗∗
(0.021)∗∗
NAFTA Mex.-Can.
1.099
0.523
1.416
0.073
0.816
(0.074)∗∗
(0.125)∗∗
(0.086)∗∗
(0.062)
(0.038)∗∗
NAFTA U.S.-Can
0.470
0.173
0.476
0.262
0.311
(0.044)∗∗
(0.070)∗
(0.048)∗∗
(0.035)∗∗
(0.077)∗∗
NAFTA Can-U.S.
0.451
0.156
0.575
0.221
0.484
(0.044)∗∗
(0.089)+
(0.049)∗∗
(0.034)∗∗
(0.080)∗∗
NAFTA Mex.-U.S.
0.999
0.415
1.268
0.230
0.874
(0.071)∗∗
(0.125)∗∗
(0.085)∗∗
(0.051)∗∗
(0.038)∗∗
NAFTA U.S.-Mex
0.615
0.255
0.748
0.240
0.418
(0.069)∗∗
(0.112)∗
(0.083)∗∗
(0.044)∗∗
(0.012)∗∗
All Other FTAs
0.318
0.154
0.323
0.107
(0.042)∗∗
(0.050)∗∗
(0.050)∗∗
(0.023)∗∗
All Other FTAs (Fully Phased-in)
0.187
(0.083)∗∗
B. Comparing NAFTA vs. the Average FTA
NAFTA vs. Average
0.261
0.220
0.341
0.133
(0.100)∗∗
(0.137)
(0.124)∗∗
(0.024)∗∗
vs. Fully Phased-in Avg FTA (χ2d f =1 )
8.78
(p: 0.003)∗∗
NAFTAt+1
0.286
(0.194)
All Other FTAst+1
0.246
(0.084)∗∗
Robust standard errors, clustered by pair, are reported in parentheses. + p < 0.10 , ∗ p < .05 , ∗∗ p < .01 . “All Other FTAs”
is “phased-in” over 8 years in the final column, using 4- and 8-year lags (including lags for pre-1990 FTAs). This specification
also includes one period leads for all FTA variables, including the directional NAFTA dummies in Panel A. Pair-specific time
trends are used in column 4. Time-varying exporter and importer fixed effects are used to control for the multilateral
resistances. Following Baier and Bergstrand (2007), pair fixed effects are used to account for FTA endogeneity.
To summarize the key takeaways from this analysis, I find strong evidence that individual
FTAs can have their own unique effect on trade and that, even within a given FTA, there are
widespread directional asymmetries (especially within individual industries). I do not however
find strong evidence for “one way trade” effects overall (although I do generally find that the
and its 4- and 8-year lags.
17
U.S. did indeed grant more market access to Mexico than it gained.) An examination of potential
concerns about the estimation provides support for the general robustness of my methods. We also
saw some evidence that what we would expect to happen ex ante based on tariffs is not always what
actually occurs when the FTA goes into effect, particularly in the case of U.S.-Mexico trade. The
following sections will discuss the possible causes of the asymmetries I identify as well as their
implications—both for further work on trade agreements and for how we estimate the welfare
impact of FTAs.
4
Discussion
I consider various intuitive explanations for why these asymmetries might be expected to arise in
the data. These results could be reflecting pre-existing patterns of comparative advantage for instance. Or they could be reflecting how much each country reduced its tariffs as a result of NAFTA
or, by the same token, the fact that the U.S. and Canada already had a free trade agreement in place
before NAFTA went into effect. In subsequent discussion, I show that none of these interpretations
adequately explains what I observe in the data. I then discuss possible alternative explanations as
well as the implications of my findings for further empirical work on FTAs.
4.1
Revealed Comparative Advantage
As noted, an attractive interpretation for the regularities I document might be that the differences
in directional estimates capture patterns of comparative advantage within the NAFTA trio and
thus can be thought of as indicators of “revealed comparative advantage”. It stands to reason for
instance that a country with comparative advantage in food production will experience a larger
increase in food exports and a smaller increase in food imports than its trading partner when the
two sides liberalize trade. I have already discussed how for each industry, the use of time-varying
country-level fixed effects in my estimation already controls for the endogenous response of prices
and overall production in each country to trade liberalization. Thus econometrically speaking,
I can safely say my results do not reflect comparative advantage in the classical sense.31 These
coefficients are instead capturing direct reductions in bilateral trade frictions that seem to favor
certain countries over others across different sectors.
It still might conceivably be the case though having comparative advantage in a particular sector somehow also directly benefits that sector on a bilateral basis relative to an FTA partner country. I examine this hypothesis by comparing my estimation results with two different empirical
measures of “revealed comparative advantage” (RCA). Panel A of Table 5 shows indices of RCA
across the three countries I calculate myself using the original Balassa (1965) formula with the 2002
trade flows from my data set. Panel B then shows alternative RCA indices taken from Leromain
31 See Costinot et al (2012) for a thorough illustration of how comparative advantage contributes to gains from trade
in a structural gravity setting.
18
Table 5: Measures of Revealed Comparative Advantage, NAFTA countries
Country
Food
Textiles
Wood
Paper
Chemicals
Minerals
Metals
Machinery
3.09
0.68
0.52
1.70
0.94
A. Balassa (1965) RCA index, 2002 trade
Canada
1.10
0.32
3.00
Mexico
0.51
0.97
1.51
0.24
0.29
1.00
0.42
1.32
U.S.
0.89
0.36
0.46
1.25
1.19
0.67
0.59
1.10
B. Leromain and Orefice (2013) RCA index, 2010 trade
Canada
1.07
0.88
1.21
1.07
0.85
1.14
1.03
1
Mexico
1
0.86
0.78
1.07
0.98
0.82
0.98
0.9
U.S.
1.02
0.86
1.13
1.07
1.06
1.02
0.98
1.07
For the Balassa RCA index, a value greater than 1 indicates comparative advantage with respect to the
rest of the world. The index values for the Leromain and Orefice (2013) RCA measure are based on the
Costinot et al (2012) structural estimation procedure. A value greater than 1 indicates better than
average productivity in a given sector relative to the rest of the world.
and Orefice (2013), who use the Costinot et al (2012) structural Ricardian estimation procedure to
reveal relative productivity differences across sectors for 2010 trade.32
The evidence for this case is mixed. To illustrate, let us focus on the cases I noted in reviewing
the directional results in Table 2. For the Food sector, we saw the U.S. and Canada realizing large
market access gains with Mexico not seeming to make any. Interestingly, the U.S. and Canada do
indeed seem to have genuine comparative advantage over Mexico in Food.33 My results for Paper
(where Mexico realized the largest gains) are contrary however: the Balassa measure says Mexico
is dominated by the U.S. and Canada in terms of comparative advantage, whereas the Leromain
and Orefice indices suggests the countries are at par with each other. The RCA numbers also fail
to predict Mexico’s large gains in Metals and Machinery, with the exception of the Balassa figure
for Machinery, which does agree.
If I focus solely on the Leromain and Orefice statistics, my preferred RCA measure, I fail to
find robust correspondence between RCA and my directional estimates in Table 2 in seven out of
the eight sectors, with Food as the only convincing exception. The Balassa measure finds Mexico
to be strong in machinery, consistent with my findings for the Mexican machinery sector in Table
2—and is also reasonably consistent with the Textiles results—but is otherwise just as weak a fit
for the other sectors.
I conclude that my findings of discrepancies in bilateral trade gains from NAFTA do not merely
reflect patterns of comparative advantage, though comparative advantage could still have some
meaningful correlation. I do show later in my welfare analysis that the asymmetries in NAFTA’s
32 The
main advantage of the Costinot et al method is its consistency with Ricardian theory. As Leromain and Orefice
explain, comparative advantage is an ex ante characteristic of a country, whereas the Balassa formula is based on ex
post realizations of trade flows, which may depend on other factors besides comparative advantage. The Costinot et al
procedure corrects this issue by recovering revealed measures of Ricardian productivity from a structural model.
33 It is important to note that neither of these measures take into account the influence of agricultural policy, which
most likely plays a role in these figures (as well as in my own results for Food). Rather they are inferred from trade data.
19
Table 6: Tariffs, before and after NAFTA
Exporter
Importer
Food
Textiles
Wood
Paper
Chemicals
Minerals
Metals
Machinery
A. 1989-1991 Import Tariffs (ad valorem, Effectively Applied)
Mexico∗
10.02
Canada
U.S.
2.40
Mexico
Canada
5.32
Mexico
U.S.
6.20
U.S.
Canada
5.24
U.S.
Mexico∗
10.81
Canada
15.29
14.95
5.17
10.09
13.72
7.14
14.08
7.52
1.11
0.28
3.79
1.56
1.93
0.38
13.67
13.18
3.14
6.39
3.32
0.71
4.14
13.16
4.05
3.57
5.07
5.83
3.38
3.81
16.5
6.98
3.19
7.50
6.44
4.31
5.40
16.98
14.91
6.46
10.13
14.44
10.03
13.99
B. 1999 Import Tariffs (ad valorem, Effectively Applied)
Canada
Mexico
27.47
8.05
4.80
4.43
5.06
3.44
3.97
3.54
Canada
U.S.
2.76
0.00
0.00
0.00
0.00
0.00
0.00
0.00
Mexico
Canada
11.78
5.82
0.10
0.27
0.48
0.23
1.37
0.33
Mexico
U.S.
3.41
1.08
0.01
0.00
0.31
1.43
1.01
0.14
U.S.
Canada
0.01
0.00
0.00
0.00
0.00
0.00
0.00
0.00
U.S.
Mexico
20.19
4.22
3.98
5.03
4.77
3.76
4.41
2.93
Weighted average tariffs computed by World Bank WITS software using UNCTAD TRAINS tariff database.
∗ Mexican
tariffs only available for 1991 in Panel A. All other Panel A tariffs from 1989.
effects seem to have done a reasonably good job of promoting gains from specialization, and it
stands to reason this may have been partly by design.
4.2
Tariff liberalization
Another standard interpretation of my results might be that I am capturing patterns of tariff liberalization associated with NAFTA. In Table 6, I show representative ad valorem “effectively applied”
tariff levels for each country pair from before and after NAFTA, using data from the UNCTAD
TRAINS tariff database, accessed using the World Bank WITS interface. For pre-NAFTA tariffs
(Panel A), all values shown are from the year 1989, with the exception of Canadian exports to
Mexico, which use 1991 tariffs because of missing data for 1989. The “after NAFTA” tariffs (Panel
B) are all from 1999. While it is not appropriate to use raw tariff changes to compare implications
for trade across sectors—because different sectors have different trade elasticities—within a given
sector, one might expect larger tariff changes to correspond with larger trade increases.
Again however, I do not find a satisfactory explanation for the directional effects I observe.
In the Food sector for instance, the incidence of Mexican and Canadian import tariffs actually
increased in the post-NAFTA period,34 yet these increases have not deterred the flow of imports
overall, rather only for Mexican exporters. That is not to say the tariff story should be dismissed
outright, since tariffs are a clear, known impediment to trade, and where there are high tariffs
originally we should see large gains in trade flows ex post. The Wood sector seems like a good
34 As discussed in Carlberg and Rude (2004), Mexico actually raised tariffs on meat products significantly just before
the signing of NAFTA. If anything this should have downward-biased the estimates for Mexican Food imports in Table
2. Effective increases in tariff incidence can also occur when newly traded products appear in the data post-NAFTA
along the highest tariff lines.
20
example. The U.S. only had a 1.11% tariff on Canadian Wood products before NAFTA—whereas
Mexico’s was 14.95%—and indeed Canada’s Wood exports to Mexico increased by more than its
exports to the U.S. did. Likewise, it may be worth noting that Mexico is the only country that
had high tariffs to begin with in Metals and they did seem to allow the largest increase in market
access in Table 2. We cannot however rationalize their large gains in market access as an exporter
N,Metals
of Metals (β exp
= MEX = 1.273) based on tariffs alone.
Indeed, as I noted earlier when I reviewed the directional results for aggregate trade, in general
there is major variation in the effects of NAFTA that cannot be explained based on tariffs. As it
was with Metals, Mexico also had the highest tariffs in the Machinery sector, yet realized the largest
gains in market access as an exporter in that industry as well. There is no clear reason to be gleaned
from Table 6 why the magnitudes for NAFTA’s effect on Textile trade would be as large as they are,
or why Canada’s import markets for Minerals seemed to open up more than Mexico’s did. Clearly,
free trade agreements can influence trade along other channels besides explicit trade cost reduction
and the majority of the variation in directional NAFTA effects I observe must be occurring along
these other channels.
A related subject I should address here is the the fact that the U.S. and Canada already had a
pre-existing FTA in place, the “Canada-United States Free Trade Agreement”, before the signing of
NAFTA. In this case, this does seem to be a compelling source of some of the variation. NAFTA is
still associated with a significant increase in Canada-U.S. trade, but if we look across all industries,
there have generally been larger decreases in trade barriers for U.S.-Mexico and Canada-Mexico
trade than there have been for U.S.-Canada trade. Again though, this observation does not contradict the finding that NAFTA affected trade in ways other than through tariffs.
4.3
Broader Implications
An interesting puzzle remains: what are the sources of this directional variation in NAFTA effects
within sectors we are seeing in the data? Are there industry-specific political pressures forcing
countries to trade off trade concessions across sectors as part of the agreement? Could it instead
be that the answer lies at the at the micro level of the firm (e.g. FDI, fixed costs of exporting,
matching between exporting and importer firms, etc.)? Or is there perhaps some pair-specific
aspect to comparative advantage yet to be identified formally in the literature? Generally when
we find impediments to trade other than tariffs that could still be influenced by policy, we call
these “non-tariff barriers”, but what is the nature of these barriers?
Valid interpretations of these findings likely could potentially need to incorporate some combination of these factors discussed above. Freund and Pierola (2012) for instance show that a
country’s revealed comparative advantage in trade is often wholly determined by a small handful
of firms and can change abruptly with the entry or exit of a large firm. Sudden entry of large firms
is hypothesized to depend on FDI, which could conceivably be geared towards pair-specific trade
flows. The extent to which FTAs may be designed to directly promote FDI in certain sectors over
others, e.g. by selective relaxation of “rules of origin” restrictions, has not been thoroughly studied
21
theoretically or empirically.
What this paper clearly establishes however is that non-tariff barriers can have a substantial role in driving the effects of free trade agreements. While I do not claim this is necessarily
surprising—it goes without saying that both tariff and non-tariff barriers play a role in restricting
trade—both the magnitudes and asymmetries in the FTA effects I identify are far different than
what would have been reasonably predicted based solely on what we can measure. If we really
want to explain why it is some FTAs promote more trade than others, which member countries
will benefit more in which industries, or which channels will drive the promotion of trade, we
first need to accept that what we consider “trade costs” in the quantitative trade literature is actually a very expansive category with a very large conceptual dimension. As a second recent paper
by Head and Mayer (2013) has documented, trade costs in general depend majorly on abstract
concepts such as information asymmetry, cultural biases, and lack of common institutions.
It stands to reason then that such conceptual barriers to trade could also concievably be considered “non-tariff barriers”—just as observable barriers such as product regulations are—and they
may interact with observable barriers in ways that are hard to predict, both for economists and
for policymakers. Furthermore, it also stands to reason a free trade agreement—as it is designed
holistically to bring two countries closer together in trade via many institutional shifts, not by any
one particular instrument—can also have an effect on this conceptual aspect of trade costs, the part
that is hard to precisely measure.
But that of course is not a very satisfying answer if you ultimately care about being able to
predict what impact a given FTA will have ex ante. To really get deeper into this question, we want
to thoroughly explore what it is we can say about the heterogeneity of FTAs based on what can be
observed. This paper is meant to mark the beginning of a long-term project that will try to get at
exactly that. In new work pursuant to the empirical findings discussed above, Yotov and Zylkin
(2014), my co-author and I expand the method used above to show asymmetries in the effects of
NAFTA to a larger scale examination of the direction-specific effects of FTAs in general. The empirical approach implied from the NAFTA results is straight-forward: using directional FTA effects as
the dependent variable, we estimate a regression using exporter and importer fixed effects among
the explanatory variables alongside other bilateral variables that may be deemed theoretically important, including observable changes in tariffs and data on other observable commonalities in
FTA provisions. Interestingly, we also find that standard gravity variables such as distance, colonial relationships, and common borders play a significant role. In general terms, we establish the
interesting new fact that FTA effects—like the trade flows they are derived from—can themselves
be modeled surprisingly successfully via a simple gravity equation.
The ultimate contribution to the literature from this line of research will be to devise better,
more informed ways to address the question of what it is FTAs actually do, not just for trade
but for welfare as well. We have reached a point in the trade literature where the models we
use to simulate the general equilibrium effects of trade policy shocks are sophisticated enough to
give us very good answers regarding what effect a given shock will have throughout the broader
22
economy. Caliendo and Parro (2014) for instance combine notions of comparative advantage, interfirm trade, input utilization differences across sectors, and non-traded goods within a structural
general equilibrium gravity model to construct a very comprehensive, yet very coherent, picture of
the different channels by which trade liberalization can affect welfare. Taking things a step further,
computational general equilibrium models, such as the one used by CEPII to forecast the effects
of TTIP (Fontagn´e et al., 2013), are widely known for their deep complexity and for their ability to
analyze many different kinds of shocks and outcome vectors.
But as I will try to make as clear in my welfare analysis that follows, the results these excellent models can produce will tend to be significantly constrained by how they model what happens at the partial equilibrium level, i.e. what happens to trade costs to set off the endogenous
responses implied by the model. Caliendo and Parro (2014) for instance use tariff changes to simulate NAFTA’s effects. The CEPII study on TTIP by contrast acknowledges that US-EU tariffs are all
but non-existent at present, but is only able to deal with non-tariff barriers by assuming an across
the board 25% reduction. Naturally then, a desireable research agenda will be to move towards
more detailed analysis that can generate reasonable sector- and directional-specific counterfactuals
for any two countries or regions that have yet to form an FTA. While there is still much more to
be written regarding how to generate these predictions, better accounting for these forces should
make for a significant refinement of how we project the welfare benefits from FTA formation, as I
show in the following section.
5
Welfare Analysis
In this section I illustrate that these large directional differences in direct FTA effects in turn can also
have important implications for general equilibrium welfare analysis of FTA formation. The initial
simulation structure I use here is the same as in Anderson and Yotov (2014), where welfare effects
occur strictly through changes in buyer and producer prices and accordingly may be thought of as
“terms of trade effects”. This particular simulation method is especially suitable for demonstration
purposes here because the first-order price effects it identifies hold important implications for the
additional welfare channels addressed in other quantitative models. I then introduce a novel way
of measuring gains from specialization as a way of illustrating these implications.
I start by specifying a multi-sector monopolistic competition model where there are Nik identical single product firms in each industry in each country, with each industry k firm in country
i supplying a differentiated variety qik at a factory gate price pi∗k . Consumers have common CES
preferences over all varieties within each industry, with elasticity parameter σk . It follows that
nominal exports from i to j of goods of type k can be written as
Xijk
=
Nik
pi∗k tijk
Pjk
23
!1− σ k
· Ekj ,
(8)
1− σ k
with tijk the standard “iceberg cost” of transporting goods of type k from i to j and Pjk ≡ ∑i Nik pi∗k tijk
the CES price index for buyers in j.
It is straightforward to re-write the CES Demand function in (8) in the form of a structural
1− σ k
serve as the parameter
gravity equation as in (1) by adjusting the notation. Let φijk ≡ tijk
1− σ k
. It follows then from an accounting
measuring of bilateral market access and let Φkj ≡ Pjk
k
1− σ
= Yik /Ωik , as in (1), with Ωi still defined as in (3) as a measure of access
identity that Nik pi∗k
to world markets for sellers in i.35 The structural interpretation for Φkj in (2) likewise follows from
the same accounting identity.
On the production-side, I assume the following cost function for an individual firm in industry
k in country i:
Cik =
wik
Aik
!
qik + f ik .
All firms use only one factor of production, labor, which may or may not be mobile across sectors.
Wages for workers in sector k are given by wik . Aik is a measure of worker productivity in each
sector, such that one unit of labor can be used to produce Aik units of qik . I also require that labor be
used both for production and for fixed cost expenditures.
The presence of the fixed entry cost f ik implies that the number of firms in each industry is
determined endogenously. Specifically, one can write the free entry condition for industry k as
pi∗k qik =
wik
Aik
!
qik + f ik .
Profit maximization under CES Demand dictates each k firm charges a constant markup in equi
librium of µk = σk / σk − 1 . Plugging pi∗k = µk · wik /Aik into the free entry condition yields the
following expression for firm quantity qik :
qik = σk − 1 f ik .
(9)
That is to say, the total quantity produced by each individual firm in equilibrium is a constant
pinned down by parameters: all variation in real output at the industry-level occurs at the exten
sive margin of the firm.36 Furthermore, using Nik pi∗k qik = wik Lik implies that Nik = Aik Lik / µk qik ;
thus the number of firms will only vary when labor is allowed to move across industries.
1− σ k
Finally, using the fact Nik pi∗k
= Yik /Ωik —with Yik = N pi∗k qik —in combination with (9)
if Yi = ∑ Xij = Si ∑ j φij Ej /Φ j , then Si = Yi /Ωi follows directly.
36 While this feature of the model obviously represents a simplification, it nonetheless is frequently observed that
most variation in real world trade flows occurs at the firm extensive margin, as opposed to the intensive margin. See
the discussion in Head and Mayer (2014), section 5.3 for a synopsis of these findings.
35 Specifically,
24
then enables me to write:
Nik pi∗k
ek = σk
where A
i
=⇒
pi∗k
=⇒
wi
−1
σk − 1
(σk −1)/σk
1− σ k
Nik pi∗k qik
Ωik
− 1 1
k
k
=
qik σ Ωik σ
1
ek Ωk σk ,
=
A
i
i
Aik / f ik
=
1/σk
(10)
is a combined parameter governing what one
might consider the “quasi-Ricardian” forces in the model. That is to say, in the case where labor is
ek ’s across countries are what would determine the tendency
fully mobile, relative differences in A
i
towards specialization, absent the influence of trade frictions.37
Ultimately the world we are talking about is not frictionless however, and thus we have two
potentially important sources for gains from trade to talk about in this setting: the initial “terms
of trade” effects—as changes in Ωik and Φkj dictate changes in buyer and seller prices in each
economy—and the ensuing “gains from specialization”, as labor adjusts to a new efficient allocation. I distinguish between these two channels by comparing the case where labor is held fixed
with what occurs when labor is fully mobile. This decomposition is useful for my purposes because the impact of Ωik and Φkj on domestic prices plays an important structural role across a very
general range of quantitative trade models. Examining the additional specialization gains then
effectively illustrates how differences in the initial terms of trade effects can translate into even
larger measurement differences as more welfare channels are introduced.
5.1
Terms of Trade Effects
Because firm entry is pinned down by the allocation of labor, the case where labor is held fixed is
structurally equivalent to an “endowment economy” setting of the type explored in Anderson and
Yotov (2014), where all welfare effects occur through prices. Note that, by the duality between Φkj
and the buyer price index, and by the dependence of wages on Ωik in (10), we have the following
intuitive relationships between price changes and access to world markets:
∆ ln pi∗k = ∆ ln wik =
1
∆ ln Ωik
σk
∆ ln Pjk =
1
∆ ln Φkj
1 − σk
That is, sellers face higher aggregate demand for their products when they are more able to reach
world markets (higher Ωik ) and, by σk > 1, buyers likewise enjoy lower prices when there is more
competition in their import markets (higher Φkj ). Regarded separately, these two effects allow for
a simple decomposition of how trade integration affects buyers vs. producers in each economy.
Regarded together, the combined effect provides a useful notion of the “terms of trade” effect from
37 I
say “quasi-Ricardian” because specialization in this context is not only influenced by relative differences in unit
labor efficiencies (Aik ’s) as in a classical Ricardian model, but also by the incidence of fixed costs.
25
liberalization.
Adapting Anderson and Yotov’s method, I generate a solvable R-by-K system of equations for
worker wages wik by summing (8) over j and dividing by world output. The resulting expression
is
Yik
=
Yk
∑
Nik µk wik /Aik
Φkj
j
1− σ k
φijk
·
Ekj
Yk
,
(11)
where Yik , Ekj , Y k , and Φkj are each themselves functions of wages. To close the model, I assume that
each country’s expenditure on manufacturing goods is a constant share δj of its total manufacturing income ∑k Yik and that buyers in each country allocate expenditure across sectors according to
a Cobb-Douglas function with common share parameters αk .
To simulate general equilibrium FTA effects within this structure, I impose the normalization
that all initial wages wik are equal to one, such that all labor allocations Lik are given by initial
k
1990 output levels, i.e. Lik = Yi,1990
. By then introducing a new system of φijk ’s reflecting NAFTA, I
can recover changes in buyer and producer prices, terms of trade, and welfare using the linkages
implied by the structural model. Specifically, I solve (11) for wik subject to the following:
Yik
wik Lik
=
Yk
∑k wik Lik
Ekj
Yk
(12)
δj ∑k wik Lik
=
∑ j δj ∑k wik Lkj
1− σ k
Φkj = ∑ Nik µk wik /Aik
φijk .
(13)
(14)
j
k
In practice, I solve
(11)-(14) twice:
once with 1990 trade costs and all wi = 1 in order to solve
1− σ k
for the combined Nik µik /Aik
terms, and then again for new wages using new post-NAFTA
φijk ’s. In addition, I take values for elasticity parameters σk from data compiled by Broda et al
2006. I also construct expenditure share parameters δj and αk from 1990 trade and output
αk /1−σk
data.38 Finally, I construct the buyer price index in each country as Pi = ∏k Φik
, and in
turn calculate welfare as real income Wi =
∑k wik Lik
.
Pi
Intuitively, national welfare increases with
seller prices (via wage income) and decreases with buyer prices. To compute aggregate changes in
national supplier prices (“pi ”), I simply compare ∑k Lik with ∑k wik Lik .39
The first three panels of Table 7 then show the simulation results for the effects of NAFTA
on price indices and welfare for each of the three specifications shown in Tables 1 and 2 for each
38 Note that all αk terms cancel out of (13). However these terms will still be needed for constructing aggregate buyer
prices and welfare.
39 Note that one last normalization is needed when we endogenize Y and E because the system in (11) is homogeneous
k
of degree zero. So I impose ∑i wik Lik = Y1990
both pre- and post-NAFTA, such that total nominal production in each
sector stays the same over time.
26
NAFTA country and for an aggregate non-NAFTA group. All changes shown are relative to the
1990 baseline year. Changes in producer prices and welfare for the non-NAFTA group are aggregated by total output shares and buyer prices are aggregated by expenditure shares. To establish
a baseline, I first note the main features of Panel A, which assumes NAFTA has been no different
than any other FTA. Notably, Mexico and Canada show larger net positive benefits from NAFTA
than the U.S. This result is intuitive because Canada and Mexico are both relatively small countries
compared with the U.S., with Mexico seeing the largest gains because it is the smallest of the three.
The Rest of the World, which as also very large, only suffers very mildly from trade diversion.
When I allow for NAFTA to have its own distinct effect on trade costs however (Panel B), we see
favorable differences for both buyers and sellers in all three NAFTA countries, with each country
realizing roughly 50% larger implied gains than in the baseline case. These results not only confirm
the message of Panel B in Table 1—that NAFTA has been significantly more effective than the
average FTA in promoting trade—but also quantify that message in terms of real income effects.
What this approach reveals is that using the simplifying assumption of a single FTA effect across
all agreements can greatly distort the true welfare gains from trade liberalization for individual
countries, even large ones. NAFTA of course is just one example—focusing on other agreements
instead would likely reveal further large differences, both positive and negative.
Panel C generates price and welfare results from trade costs generated using my main specification with asymmetric NAFTA effects. Noting that Mexico had the largest gains in both exports
and imports in Table 3, it is unsurprising that the differences in welfare changes between Panels
B and C heavily favor Mexico, and that these gains are large (> 2%) for both buyers and sellers.
Clearly, Mexico has benefited the most from NAFTA out of the three countries. It is also reasonable that these gains have come largely at the expense of Canada, who has had the smallest overall
increase in market integration from NAFTA, again going by columns 5 and 6 of Table 3. Both the
U.S. and the Rest of the World, largely since they are both very big regions, are less sensitive to
these asymmetries.
One last experiment I consider is comparing what we might have expected to happen based on
tariff changes with what actually happened. The last panel of Table 7 shows the counterfactual for
prices and welfare if changes in trade costs were driven solely by the tariff changes shown in Table
6. As we can see, estimating welfare effects based solely on tariff changes not only mischaracterizes
the composition of each country’s gains from NAFTA—it shows U.S. producers benefiting at the
expense of producers in the other two countries (as well as U.S. buyers) for instance—but also
greatly underestimates the magnitudes of the overall welfare gains in general. It is clear then that
tariffs may do a poor job of predicting the true effects of FTAs and that more work needs to be
done understanding other means by which FTAs can affect trade.
5.2
Gains from specialization
What then do these first-order effects on prices imply in turn for more nuanced welfare calculations? A recent Handbook of International Economics chapter by Costinot and Rodriguez-Clare (2014)
27
Table 7: Initial Terms of Trade Effects (NAFTA only)
A. Average FTA effect
B. Average NAFTA effect
%∆Supplier Prices
%∆Buyer Prices
%∆Welfare
%∆Supplier Prices
%∆Buyer Prices
%∆Welfare
Canada
1.73
-0.43
2.17
2.43
-1.09
3.52
Mexico
2.52
-1.11
3.63
3.18
-2.57
5.75
U.S.
0.06
-0.16
0.22
0.15
-0.19
0.34
ROW
-0.10
-0.09
-0.01
-0.16
-0.14
-0.02
C. Directional NAFTA effects
D. Using Tariffs
%∆Supplier Prices
%∆Buyer Prices
%∆Welfare
%∆Supplier Prices
%∆Buyer Prices
%∆Welfare
Canada
2.34
-0.89
3.23
-0.38
-1.27
0.89
Mexico
5.35
-3.53
8.87
0.84
-1.52
2.37
U.S.
0.14
-0.20
0.34
0.30
0.20
0.10
ROW
-0.18
-0.16
-0.02
-0.10
-0.09
-0.01
Supplier prices and welfare for the ROW region aggregated by output shares.
Buyer prices for ROW aggregated by expenditure shares.
All changes relative to 1990 price and welfare levels, assuming fixed supply quantities.
surveys the many additional modeling considerations that can be built on top of this basic structure, including firm heterogeneity, trade in intermediate goods, and having additional factors of
production. However one important channel they do not explicitly isolate—which plays a role in
any model with multiple sectors—is the welfare impact from the specialization of factors of production in response to trade. In my setting, this impact is easily obtained by comparing the gains
from trade with and without free movement of labor.
This experiment is distinct from the one in Costinot et al. (2012), because I allow the shifting
incidence of trade costs to play an important role in shaping the gains from specialization rather
than focusing strictly on Ricardian forces. Examining (10), all else equal, workers will gravitate
1/σk
towards the sector that enjoys the largest increase in Ωik
, which may not necessarily be the
sector they would move to in a world with global free trade. Redding (2014) also examines the
implications of labor mobility for welfare, but his analysis is concerned with the movement of
labor across geographic regions rather than across industries.
For the simulation, I essentially repeat the procedure described above as written, only this time
I require that new equilibrium wages be the same across all sectors (that is, wik = wi ) and that
the labor allocation adjusts such that labor markets clear. It is also important to keep in mind
that ∆Nik = ∆Lik , such that as labor moves across sectors, new varieties will appear in expanding
sectors and some varieties will disappear in receding sectors. The overall increase in the number
of varieties in each industry as countries realize efficiency gains from specialization is then a major
source of the additional global welfare benefits. Specialization also holds important implications
for the aggregate incidence of trade costs, as discussed in Anderson and Yotov (2011).
Table 8 then shows the additional specialization gains that obtain under the three most interesting cases discussed above: the case where NAFTA’s effects are assumed to be symmetric, the
case where NAFTA’s effects are allowed to be asymmetric, and the case where all asymmetries are
28
assumed to be due to tariffs. In each case, the first column shows the “Initial Terms of Trade Gains”
taken from Table 7 and the third column shows the new welfare change when labor is allowed to
be move freely. The second column, “Gains from Specialization”, is simply the difference between
the other two columns.
In all cases, gains from specialization are non-negative. However there are important differences both across the three cases and across the different regions. In the “Average NAFTA effects”
case for example (Panel A), it is interesting to note that Mexico not only enjoys the largest additional gains from specialization, but also derives the largest share of its overall welfare gains from
specialization: roughly 25% of its gains are from specialization, vs. 17% for Canada, and only
10.5% for the U.S. I emphasize these differences because they show that the implications for how
we measure the specialization channel do not neatly carry over from initial biases identified in the
endowment economy case.
Panel B, which uses Directional NAFTA effects, then helps clarify this point. Here the share of
gains that is attributed to specialization increases for all three NAFTA countries. This occurs even
though the initial terms of trade gains in this case are smaller for Canada and unchanged for the
U.S. Since all I have done here compared with Panel A is allow for asymmetries within NAFTA,
this experiment goes to show that how we measure the incidence of trade costs matters a great
deal for how we attribute gains to different welfare channels. Mexico’s gains in this case jump all
the way to 14.4%, with almost 40% of those gains coming from specialization. One might even
argue these results suggest that NAFTA was intentionally implemented in such a way that each
could could realize large specialization gains, backing up some of the observations made based on
Table 5.
The results based on tariffs (Panel C) then make it clear that it is not just the presence of asymmetries that generate large gains from specialization, but rather the specific nature of the heterogeneity. We saw in Table 7 that using tariffs under-predicts the welfare gains for all three countries.
We see now that gains from specialization are also much smaller in this case, both in terms of magnitude and as a share of overall gains. Most strikingly, the implied specialization gains for Canada
become negligible.
Clearly, the mere fact of trade liberalization is not sufficient to generate large specialization
gains. A particular liberalization will affect countries with different geographic and productive
configurations in different ways. Another striking result from Table 8 for example is that the additional gains for the Rest of the World are of a similar magnitude as those for the U.S. Why is this?
It turns out that, relatively speaking, the pre-existing configuration of U.S. production was already
relatively well aligned with the opportunities presented by NAFTA. I find in the “Average NAFTA
effects” case for instance that only 0.40% of the US labor force is induced to move across sectors
(perhaps again this feature of NAFTA was in part by design.) For comparison, labor movement
in the ROW region is of a very similar magnitude, 0.44%, even though ROW is not involved in
any liberalization in this experiment and even though the ROW aggregate being considered here
is almost four times as large as the U.S.
29
Table 8: Gains from Specialization (NAFTA only)
Initial Terms of Trade Gains (%∆)
Gains from Specialization (%∆)
Total Welfare Impact (%∆)
A. Average NAFTA effects
Canada
3.52
0.72
4.24
Mexico
5.75
1.91
7.66
U.S.
0.34
0.04
0.38
ROW
-0.02
0.06
0.04
B. Directional NAFTA effects
Canada
3.23
1.38
4.61
Mexico
8.87
5.57
14.44
U.S.
0.34
0.08
0.42
ROW
-0.02
0.08
0.06
Canada
0.89
0.01
0.90
Mexico
2.37
0.28
2.65
C. Using Tariffs
U.S.
0.10
0.02
0.12
ROW
-0.01
0.00
-0.01
“Initial Terms of Trade Gains” are the total welfare effects from the previous table.
Welfare changes for ROW aggregated by output shares.
All changes relative to 1990 welfare levels, assuming market clearing in labor markets.
Overall, my welfare analyis illustrates that NAFTA’s effectiveness at promoting trade relative
to other FTAs translates into larger welfare gains for all three countries that would otherwise be
measured. Furthermore, when we allow NAFTA to affect trade costs asymmetrically, Mexico benefits by far the most (and Mexican producers in particular). Comparisons with predictions based
on tariffs reveal that tariff changes do not necessarily determine either the magnitude or withincountry distribution of an FTA’s impact. These differences are all especially pronounced when we
allow for gains from specialization, which themselves are heavily dependent on the underlying
patterns of liberalization.
6
Closing Remarks
The question of how much free trade agreements actually increase trade between countries remains an important, actively studied topic in the trade literature. While most empirical work has
focused on identifying the average effect of FTAs on trade, my refined approach makes it plain that
some FTAs increase trade more than others and furthermore that the effects of FTAs are not necessarily symmetric between partners within a particular sector. Accounting for NAFTA’s particular
effects on trade in a general equilibrium simulation setting reveals quantifiably large implications
for welfare: ignoring these variations both understates the overall gains from trade and also misses
important differences in the incidence of price effects for buyers vs. sellers in each country, as well
as the degree to which these price effects translate into additional gains from specialization. I show
30
that US sellers in particular did not see the benefits that would otherwise be predicted using tariffs.
In general, the gains and losses for buyers and sellers in each country correspond intuitively to the
differences in market access gains I identify for NAFTA in my empirical analysis. It is not surprising then that basing predictions on tariffs does an especially poor job of predicting NAFTA’s
effects, since tariffs in general do not correspond to the actual patterns of market liberalization I
demonstrate.
I envision there being significant potential for future work both expanding on the approach I
have pursued in this paper and examining the reasons for the differences I obtain, and there is
already a substantial research pipeline underway. My findings for trade within NAFTA motivate a
larger-scale project geared towards characterizing how each individual FTA currently in existence
has promoted trade across sectors and how the effects of those FTAs have differed depending on
the direction of flows, which I am following up on in further work. What is clear from this paper
is that non-tariff barriers can play a substantial role in driving the effects of free trade agreements
and there is much more to be said about how we can potentially clarify and anticipate how this
channel may function for future FTAs.
Another important dimension to think about is what the changing nature of Free Trade Agreements means for theoretical research on trade policy. To say the theoretical literature on trade
liberalization is “enormous” would be understating its true size. Yet by and large, this literature
has focused on strategic games involving tariffs, or analogous instruments. What the empirical
findings outlined here suggest is that to truly come up to speed with modern evolutions in trade
liberalization, economists must take a more expansive view of how trade liberalization now occurs
in a modern world context. Trade integration these days for instance may not necessarily involve
the removal of trade-negating institutions—such as the authority to tax at the border—so much
as they create new institutions that promote trade—such as a supra-national commercial court intended to give foreign producers and investors more legal certainty over their competitive rights
to export and invest. It also has not been explored how harmonizing product regulations and relaxing labeling requirements, which allow goods from foreign sellers to be more easily perceived
as similar to domestic products, may work along the information friction aspect of trade costs.
Does trade liberalization still tend to be optimal when it involves consumers receiving muddier
signals about the processes used to produce the goods they choose from? Or is the real channel
that they gain “trust” for products produced under the aegis of an agreement requiring common
production standards? These are important questions for new theories to explore.
Combined with more nuanced theory and more thorough empirical unpacking of the heterogeneity in the effects of FTAs, the tools and ideas described in this paper should mark an important first step towards my ultimate research agenda: making informed, pair- and direction-specific
predictions about how hypothetical FTAs not yet formed will affect the welfare of member and
non-member countries in general equilibrium.
31
References
[1] Anderson, James. E. 2010. “The gravity model,” NBER Working Paper No. 16576.
[2] Anderson, James E. and Eric van Wincoop. 2002. “Borders, Trade, and Welfare,” in D. Rodrick
and S. Collins (eds), Brookings Trade Forum 2001, Brookings Institution.
[3] Anderson, James E. and Eric van Wincoop. 2003. “Gravity with Gravitas,” American Economic
Review, 93(1): 170-92.
[4] Anderson, James E., and Yoto V. Yotov. 2011. “Specialization: Pro-and Anti-Globalizing, 19902002.” NBER Working Paper No. 16301.
[5] Anderson, James E. and Yoto V. Yotov. 2014. “Terms of Trade and Global Efficiency Effects of
Free Trade Agreements,” NBER Working Paper No. 17003.
[6] Arkolakis, Costas, Costinot, Arnaud, Donaldson, Dave, and Rodr´ıguez-Clare, Andr´es. 2012.
“The Elusive Pro-Competitive Effects of Trade,” Unpublished, Massachusetts Institute of
Technology.
[7] Arkolakis, Costas, Costinot, Arnaud, and Rodr´ıguez-Clare, Andr´es. 2012. “New trade models,
same old gains?” American Economic Review, 102(1): 94-130.
[8] Armington, Paul S. 1969. “A Theory of Demand for Products Distinguished by Place of Production,” Staff Papers - International Monetary Fund, 16(1): 159-178.
[9] Avery, William P. 1998. Domestic interests in NAFTA bargaining. Political Science Quarterly,
113(2), 281-305.
[10] Baier, Scott L. and Jeffrey H. Bergstrand. 2002. “On the endogeneity of international trade
flows and free trade agreements.” Manuscript.
[11] Baier, Scott L. and Jeffrey H. Bergstrand. 2004. “Economic determinants of free trade agreements,” Journal of International Economics, 64: 29-63.
[12] Baier, Scott L. and Jeffrey H. Bergstrand. 2007. “Do free trade agreements actually increase
members’ international trade?” Journal of International Economics, 71(1): 72-95.
[13] Baier, Scott L. and Jeffrey H. Bergstrand. 2009. “Estimating the effects of free trade agreements
on international trade flows using matching econometrics” Journal of International Economics,
77(1): 63-76.
[14] Baier, Scott L., Jeffrey H. Bergstrand., and Michael Feng. 2013. “Economic Integration Agreements and the Margins of International Trade.” Manuscript.
[15] Bhagwati, Jagdish N., Pravin Krishna, and Arvind Panagariya, eds. 1999. Trading blocs: alternative approaches to analyzing preferential trade agreements. Mit Press, 1999.
32
[16] Behrens, Kristian, Giordano Mion, Yasasada Murata, and Jens Suedekem. 2014. “Trade,
Wages, and Productivity,” International Economic Review, forthcoming.
[17] Broda, Christian, Joshua Greenfield, and David Weinstein. 2006. “From Groundnuts to Globalization: A Structural Estimate of Trade and Growth,” NBER Working Paper No. 12512.
[18] Caliendo, Lorenzo and Fernando Parro. 2014. “Estimates of the Trade and Welfare Effects of
NAFTA,” Review of Economic Studies. Forthcoming.
[19] Carlberg, Jared and James Rude. 2004. “Intercambio de comercio agroalimentario entre
Canad´a y M´exico bajo el Tratado de Libre Comercio.” Revista Mexicana de Agronegocios, 8(15),
282-293.
[20] Cheng, I.-Hui and Howard J. Wall. 2002. “Controlling for heterogeneity in gravity models of
trade,” Federal Reserve Bank of St. Louis Working Paper vol. 1999-010C.
[21] Cipollina, Maria and Salvatici, Luca, 2010. “Reciprocal Trade Agreements in Gravity Models:
A Meta-Analysis,” Review of International Economics, 18(1): 63-80.
[22] Costinot, Arnaud, Dave Donaldson, and Ivana Komunjer. 2012. “What Goods Do Countries
Trade? A Quantitative Exploration of Ricardo’s Ideas,” Review of Economic Studies, 79(2): 581605.
[23] Costinot, Arnaud and Andr´es Rodr´ıguez-Clare. 2014. “Trade Theory with Numbers: Quantifying the Consequences of Globalization,” Handbook of International Economics, 4: 197-262.
[24] Eaton, Jonathan and Samuel Kortum. 2002. “Technology, Geography, and Trade,” Econometrica, 70(5): 1741-1779.
[25] Fontagn´e, Lionel, Julien Gourdon, and Sebastien Jean. 2013. “Transatlantic Trade: Whither
Partnership, Which Economic Consequences?” CEPII Policy Brief 2013-1.
[26] Frankel, Jeffery, Ernesto Stein, and Shang-jin Wei. 1995. “Trading blocs and the Americas: The
natural, the unnatural, and the super-natural,” Journal of Development Economics, 47(1): 61-95.
[27] Frankel, Jeffery. 1997. “Regional Trading Blocs,” Institute for International Economics, Washington, DC.
[28] Fruend, Caroline and Martha D. Pierola. 2012. “Export Superstars,” World Bank Policy Research Working Paper No. 6222.
[29] Ghosh, Sucharita and Steven Yamarik. 2004. “Are Preferentrial Trade Agreements Trade Creating? An Application of Extreme Bound Analysis,” Journal of International Economics, 63(2):
369-395.
[30] Head, Keith and Thierry Mayer. 2013. “What separates us? Sources of resistance to globalization,” Canadian Journal of Economics/Revue canadienne d’´economique, 46(4), 1196-1231.
33
[31] Head, Keith and Thierry Mayer. 2014. “Gravity equations: Workhorse, Toolkit, and Cookbook,” Handbook of International Economics, 4: 131-196.
[32] Helpman, Elhanan, Marc Melitz, and Yona Rubinstein. 2008. “Estimating Trade Flows: Trading Partners and Trading Volumes.” Quarterly Journal of Economics, 123(2): 441-487.
[33] Krugman, Paul. 1980. “Scale Economies, Product Differentiation, and the Pattern of Trade,”
American Economic Review, 70(5): 950-959.
[34] Krueger, Anne O. 1999. “Trade Creation and Trade Diversion under NAFTA,” NBER Working
Paper No. 7429.
[35] Leromain, Elsa and Gianluca Orefice. 2013. “New Revealed Comparative Advantage Index:
Dataset and Empirical Distribution,” CEPII Working Paper No. 2013-20.
[36] Magee, Chris 2003. “Endogenous preferential trade agreements: an empirical analysis,” Contributions to Economic Analysis and Policy, 2. Berkeley Electronic Press.
[37] Melitz, Marc J. 2003. “The impact of trade on intra-industry reallocations and aggregate industry productivity,” Econometrica, 71(6): 1695-1725.
[38] Melitz, Marc J., and Gianmarco I. P. Ottaviano. 2008. “Market size, trade, and productivity,”
Review of Economic Studies, 75(1): 295-316.
[39] Nicita, Alessandro and Marcelo Olarreaga. “Trade, Production, and Protection 1976-2004,”
World Bank Economic Review, 21(1), 165-171.
[40] Redding, Stephen J. 2014. “Goods Trade, Factor Mobility and Welfare,” NBER Working Paper
No. 18008.
[41] Romalis, John. 2007. “NAFTA’s and CUSFTA’s Impact on International Trade”, Review of Economics and Statistics, 89(3): 416-435.
[42] Santos Silva, Jorge and Sylvana Tenreyro. 2006. “The Log of Gravity,” Review of Economics and
Statistics, 88(4): 641-658.
[43] Shikher, Serge. 2012. “Predicting the effects of NAFTA: Now we can do it better!” Journal of
International and Global Economic Studies, 5(2): 32-59.
[44] Tinbergen, Jan. 1962. “An Analysis of World Trade Flows” in Shaping the World Economy, Ed.
Jan Tinbergen. New York, NY: Twentieth Century Fund.
[45] Trefler, Daniel. 1993. “Trade Liberalization and the Theory of Endogenous Protection: An
Econometric Study of U.S. Import Policy,” Journal of Political Economy, 101: 138-60.
34
[46] UNCTAD. 2005. “Methodologies, Classification, Quantification and Development Impacts of
Non-Tariff Barriers (NTBs),” Expert Meeting on Methodologies, Classifications, Quantification and
Development Impacts of Non-Tariff Barriers.
[47] Villarreal, M. Angeles and Ian F. Ferguson. 2014. “NAFTA at 20: Overview and Trade Effects,”
Washington, DC: Congressional Research Service.
[48] Viner, Jacob. 1950. The Customs Union Issue. Carnegie Endowment for Peace.
[49] Wei, Shang-Jin, 1996. “Intra-National versus International Trade: How Stubborn are Nations
in Global Integration?” NBER Working Paper 5531.
[50] Wooldridge, Jeffrey M. 2002. Econometric Analysis of Cross-Section and Panel Data. Cambridge,
MA: MIT Press.
[51] Yotov, Yoto V. and Thomas Zylkin. “How do Free Trade Agreements actually affect Trade?”
Unpublished, Drexel University.
35