Firms and International Trade: Evidence, Theory, and Policy Implications
Marc J. MelitzDepartment of Economics and Woodrow Wilson School
Princeton University
Extent of Firm/Plant Heterogeneity Within Sectors
Distribution of �rm size: overall and within sector� Standard deviation of log sales
Country # of producers Overall Within Sector (52 Manufacturing Sectors)France 76,456 1.82 1.70Italy 39,704 1.33 1.29Spain 31,446 1.26 1.18
U.S. (plants) 224,009 1.67
Large levels of heterogeneity in other key performance measures� Productivity� Standard deviation of log value added per worker for U.S. plants:Overall: .75 Within 4-digit sectors (450 sectors): .66
� Heterogeneity in other key measures: capital/labor ratio, investment, R&D, ...
�! Does static heterogeneity matter for aggregate cross-sector effects (i.e. for `GE' �elds)?Maybe �rms in a sector tend to change in the same way over time, so can think of an `average�rm' response for the sector?
2
Extent of Firm/Plant Heterogeneity Within Sectors: Dynamics
� Evidence on job �ows:� In the U.S., on average each year, 1 in every 10 jobs are created and destroyed byentering, exiting, expanding, contracting �rms
� Less than 10% of these job �reallocations� re�ect shifts across 4-digit sectors� This pattern holds for almost all countries for which such jobs data is available
� Other plant performance measures:
3
Extent of Firm/Plant Heterogeneity Within Sectors: Dynamics
� Do these within-sector �rm �uctuations represent �white noise� that is irrelevant for generalequilibrium analysis?
� Should one still focus on the �average� �rm response in a sector (even though this representsless than 10% of the �rm-level variance)?
� Evidence on U.S. plant-level heterogeneity and business cycles:� Davis & Haltiwanger (QJE 1992) show that U.S. business cycle �uctuations can not beexplained by �average� �rm response
� Next: Evidence for trade
4
Trade and Plant-Level Performance
� Very few plants/�rms export:� 21% of U.S. �rms plants� 14.6% of U.S. �rms (exporting plants are more likely to be owned by multi-plant �rm)� 17.4% of French �rms� Same pattern holds for both developed and developing countries
� Partitioning of �rms by export status occurs within narrowly de�ned sectors:
5
Trade and Plant-Level Performance (Cont.)
� Export intensity among exporting plants/�rms is typically very low:
6
Trade and Plant-Level Performance (Cont.)Are exporters different than non-exporters in the same sector?
All columns include 4-digit sector and state controls. Columns (d)-(f) add control for plant size (total employment)7
Trade and Plant-Level Performance (Cont.)Are exporters different than non-exporters in the same sector?
Ratio of productivity to mean within 4-digit sector 8
Trade and Plant-Level Performance (Cont.)
Causality of link between export status and productivity� Strong evidence for self-selection of more productive �rms into exporting� Colombia, Mexico, and Morocco: Clerides, Lach, and Tybout (1998, QJE)� U.S.: Bernard and Jensen (1999, JIE)� Taiwan: Aw, Chen, and Roberts (2001, JDE)
� Mixed evidence for �learning-by-exporting�� Some evidence in growing, developing countries (India, Sloevenia)
� However, exogenous causality from either export status or productivity is suspect:� New evidence shows that �rms make joint decisions concerning both export participationand technology use:� Bustos (2006 WP): Adoption of new technology & spending on new technology(Argentina)� Verhoogen (2006 WP): ISO 9000 certi�cation (Mexico)� Tre�er (2007 WP): One time productivity improvement (Canada)
9
Trade and Plant-Level Performance (Cont.)
Evidence on determinants of plant death (U.S. plants, over 5 year period)From Bernard & Jensen (ReStat, forthcoming):� Unconditionally, export status is associated with 12.6% reduction in probability of death� This is a very large number: Overall probability of death is 27%
� This is not driven only by higher productivity of exporters (low productivity is an importantsigni�cant predictor of exit):� Conditional on a full set of industry and �rm controls (productivity, size, capital-laborratio, ...), export status is still associated with a signi�cant 5-6% reduction in probabilityof death.
� Strong suggestive evidence for sunk costs of exporting
� Other effects of plant structure on probability of death:� Multi-product plants (lower conditional probability of death)� Ownership Structure also matters:
� Multi-plant �rm� Multinational �rm� Recent takeover� ... are associated with a higher conditional probability of death� (though lower overall death probability due to other observable characteristics)
10
Trade and Plant-Level Performance: Reallocations
Evidence for U.S. (Bernard & Jensen, 2002 WP):� Although exporter productivity does not grow faster than non-exporters
� Exporters grow faster than non-exporters (in terms of both shipments and employment)
� �! Explains 40% of U.S. manufacturing TFP growth� Half of this reallocation occurs within sectors
� Other suggestive evidence linking trade liberalization and productivity growth driven byreallocations:� Mexico: Tybout & Westbrook (1995 JIE)� Taiwan: Aw, Chen, & Roberts (2000 WBER)� Chile: Pavcnik (2002 ReStud)
� 19.3% productivity growth in manufacturing during 1979-1986� Contribution of 6.6% from increased productivity within plants� Contribution of 12.7% from reallocation towards more ef�cient producers
11
Direct Evidence for Effects of Trade Liberalization
Evidence for U.S. Manufacturing (Bernard, Jensen, & Schott, JME 2006)� Construct measure of U.S. trade costs by 4-digit sector over time (on import side)� Captures both �transport costs� (freight and insurance) and �policy barriers� (dutiesimposed)
� Enough variation over time and sectors to control for both 4-digit sector effects and yeareffects
� Test for effect of changing trade costs on industry productivity (within sector)
12
Trade Liberalization and Industry ProductivityEvidence for U.S. Manufacturing (Bernard, Jensen, & Schott, JME 2006):
13
Trade Liberalization and Reallocations: Market Entry and Exit
Plant-level estimations for high intra-industry trade sectors:� Effect on plant death:� Lower trade costs substantially increases the death probability for non-exporting plants:� A .9% trade cost decrease per year (a one standard deviation change) leads to a 3.7%increase in the probability of death. (Average probability of death in the sample is 26.8%)
� Probability of death for exporters increases by .2%
� Effect on export market entry:� Lower trade costs substantially increases the probability of export market entry fornon-exporters
� A .9% trade cost decrease per year (a one standard deviation change) leads to a 5.8%increase in the probability of exporting. (Average probability of becoming an exporter inthe sample is 27%)
�!These reallocations are driving an important part of the previously measured industryproductivity gains
14
Transitions Into and Out-of Exporting� Export market entry by existing non-exporters have important macroeconomic conse-quences� 40% (U.S.) to more than 50% (Colombia, Mexico) of export increases are driven by theentry of plants into the export market
� In an average year, 13.9% of U.S. plants begin exporting and 12.6% of exporters dropout of the export market
� Kehoe and Ruhl (2003 WP) �nd that newly traded goods account for a substantialamount of the increased trade following trade liberalization
1820
2224
2628
Log
Impo
rts
7 8 9 10 11 12Log Number of Imported Varieties
1.5
0.5
1Im
ports
Ann
ual G
row
th (d
etre
nded
)
.4 .2 0 .2 .4Imported Varieties Annual Growth (detrended)
15
Export Destinations� Exporting is not just a binary decision: A �rm decides where to export and how many exportdestinations to serve:
AFG ALB
ALG
ANG
ARG
AUL
AUT
BAN
BEL
BEN
BOL
BRA
BUL
BUK
BUR
CAM CAN
CEN
CHACHI CHNCOL
COS
COT
CUB
CZE
DEN
DOMECU
EGY
ELS
ETH
FIN
FRA
GEE
GER
GHA
GRE
GUAHON
HOK
HUNIND
INO
IRN
IRQ
IREISR
ITA
JAM
JAP
JOR
KEN
KORKUW
LIB
LIY
MAD
MAW
MAYMAL
MAUMAS MEX
MOR
MOZ
NEP
NET
NZE
NIC
NIGNIA
NOR
OMA PAKPAN
PAP
PAR
PERPHI
POR
ROMRWA
SAUSEN
SIE
SIN
SOM
SOU
SPA
SRI SUD
SWE
SWI
SYRTAI
TAN
THA
TOG
TRI
TUN
TUR
UGA
UNKUSA
URUUSR
VEN
VIE
YUGZAI
ZAMZIM
1010
010
0010
000
1000
00N
exp
orte
rs
1 10 100 1000 10000Market Size ($ billions)
Source: Eaton, Kortum, and Kramarz (AER 2004, P&P)
16
Export Destinations � with Control for Geography and Demand Conditions
AFGALB
ALG
ANG
ARG
AUL
AUT
BAN
BEL
BEN
BOL
BRA
BUL
BUK
BUR
CAM CAN
CEN
CHACHI CHNCOL
COS
COT
CUB
CZE
DEN
DOMECU
EGY
ELS
ETH
FIN
FRA
GEE
GER
GHA
GRE
GUAHON
HOK
HUNIND
INO
IRN
IRQ
IREISR
ITA
JAM
JAP
JOR
KEN
KORKUW
LIB
LIY
MAD
MAW
MAYMAL
MAUMAS MEX
MOR
MOZ
NEP
NET
NZE
NIC
NIGNIA
NOR
OMA PAKPAN
PAP
PAR
PERPHI
POR
ROMRWA
SAUSEN
SIE
SIN
SOM
SOU
SPA
SRISUD
SWE
SWI
SYRTAI
TAN
THA
TOG
TRI
TUN
TUR
UGA
UNKUSA
URUUSR
VEN
VIE
YUGZAI
ZAMZIM
1010
010
0010
000
1000
00N
exp
orte
rs
1 10 100 1000French Firms Market Size ($ billions)
Source: Eaton, Kortum, and Kramarz (AER 2004, P&P)17
Multiple Export Destinations: Firm Characteristics
Are �rms serving multiple export markets different?
19
Multiple Export Destinations: Firm Characteristics
Are �rms serving multiple export markets different?
20
Ricardian Models: Cross-Country Differences in Technology
Eaton-Kortum Model: Multi-Country Extension of DFS (AER 1977) with StochasticTechnology Draws (Ecma 2002; AER 2003 with Bernard & Jensen)� Fixed set of goods that can be produced in any country
� Production technology:� Aggregate differences across countries combined with idiosyncratic advantage/disadvantagefor particular goods (stochastic draw)
� Equilibrium:� Goods with bad technology draws are not produced and imported� Goods with good technology draws are produced and exported� Goods with intermediate draws are produced but not exported
� They are sheltered from import competition by trade costs� Model allows for a very �exible pattern of bilateral trade costs (requiring only atransitivity rule)
� Factor prices adjust to compensate for aggregate technology differences
� Model therefore explains why goods produced with better technologies are exportedwhereas goods with inferior technology may be produced but not exported
� Model emphasizes competition to produce a given good across countries and the effects oftrade costs (along with the key stochastic technology differences)
22
Monopolistic Competition Models with Firm-Level Fixed Costs and EndogenousEntry
� These models neglects competition to produce an individual variety
� Instead, models emphasize:� Firm/product level costs that determine �rm size� Key endogenous entry mechanism of monopolistic competition
� These models can be extended to incorporate additional �rm-level decisions:� Multinationals: horizontal and vertical FDI� Choice of technology and innovation� Expansion of product line
23
Firm Heterogeneity and Trade Induced Reallocations
The Impact of Trade on Intra-Industry Reallocations and Aggregate IndustryProductivity by Melitz (Ecma 2003)� Each �rm produces its own �variety� of a manufactured good.
� Developing this variety and one time production setup costs entail a sunk entry cost
� Following entry, �rms observe their productivity level� Prior to entry, only distribution of potential productivity levels is known (and commonacross �rms)
� Firms also face a �xed overhead production cost
� Exporting involves both a standard �per-unit� trade cost as well as a �xed export cost
� An entering �rm decides whether to produce (or exit) and then whether to export (or onlyserve its domestic market)
24
Trade Liberalization
� Forces least productive �rms to exit
� Re-allocates market shares towards more productive �rms� Resulting in higher average productivity
� Welfare gains� Combination of higher average productive and an ambiguous effect of product variety
28
Policy Implications
� Policies generate substantial inter-�rm reallocations � even within sectors� ... as �rms respond differently to same policy
Trade Liberalization� Induces substantial re-allocations that are not re�ected in sector-wide statistics� Short run transition costs may thus be under-estimated� Resistance to trade liberalization may be linked in part to greater uncertainty concerningjob tenure
� Challenge for trade policy associated with trade liberalization:� Want to cushion harshest shocks along transition� ... yet also need to allow re-allocations to proceed� �! This is what delivers long run gains to trade� ... and these gains can be more than just a more ef�cient allocation of market sharesacross �rms�
30
Policy Implications
� Firm-level trade barriers are substantial
� Need to understand how trade barriers affect �rm/product level export decision
� Sunk export market entry costs are substantial:� US $350K-$430K for Colombian manufacturing plants� Uncertainty greatly magni�es the effects of such sunk costs� Induces long transition dynamics
� Over 10 years, even for correctly perceived permanent changes in policy
� Information is an important component of trade barriers� Only way to explain large and persistent effect of distance on trade� Also explains large common language effect on trade
� Regulation costs substantially affect the extensive margin of trade� Time to process through customs� # of documents and signatures required� The effects of such costs are magni�ed when they occur at both ends of a tradingrelationship
31
Policy Implications: Regulation Costs of Trade
Relationship Between Unexplained Propensity to Trade (Import) and Regulation Costs
AFG
ALB
DZA
AGO
ARG
AUSAUT
BGD
BEL
BTN
BOL
BRA
SLB
BGR
BDI
KHM
CMR
CAN
CAF
LKA
TCD
CHLCHN COL
COG
CRI
BEN
DNK
DOM
ECU
SLV
FJI
FIN
FRA
GHA
KIR
GRC
GTM
GINGUY
HTIHND
HKG
HUN
ISL
IND
IDN
IRN IRQ
IRL
ISR
ITA
CIV
JAM
JPN
JOR KEN
KOR
KWT
LAO
LBN
MDG
MWI
MYS
MDVMLI
MRT
MUS
MEX
MNG
MAR
MOZ
OMN
NPL
NLD
NZL
NIC
NER
NGA
NORPAK
PNGPRY
PER
PHL
POL
PRT
ROU
RWA
SAU
SEN
SLE
SGP
VNM
ZAF
ZWE
ESP
SDN
SWECHE
SYR
THA
TGO
ARE TUNTUR
UGA
EGY
GBR
TZA
USA
BFA
URY VEN
YEM ZMB
20
24
6Im
porte
r Fix
ed E
ffect
50 100 150# days to import
Note: 1 unit of �xed effect represents an average 15% probability of trade with a given trade partner
32
Policy Implications: Regulation Costs of Trade
Relationship Between Unexplained Propensity to Trade (Import) and Regulation Costs:Controlling for real GDP and GDP per capita
DNK
SWE
DOM
SGP
SEN
FINNOR
CHN
MOZ
KOR
NLD
NPL
PHL
ESP
TZA
AUT
SLE
BEL
LKA
MRT
MUS
FJI
TGO
MWIJAM
ISR
THAIDNBEN
PAK
MLIIRL
EGY
USA
MYS
JPN
NGA
HTI
NZLUGAPNG
MDG
IND
CHLGHA
PRT
BFA
JOR
TUR
ZMB
CAN
MAR
PRY
BGD
NIC
URY
ISL
HND
KEN
GTM
HKG
PER
ROU
CIV
POL
TUN
GBR
BOL
AUSGUY
MEX
HUN
NER
GIN
ECU
AGOCOG
CRI
RWA
CMR
ARG
SLV
ZAF
BDI
TCDCOL
ZWE
FRA
IRN
SYR
CHE
VENBRA
GRC
DZA
CAF
ITA
21
01
2e(
imp_
fe |
X )
1 .5 0 .5 1e( limp_time | X )
coef = .61246452, se = .14787213, t = 4.14
Note: 1 unit of �xed effect represents an average 15% probability of trade with a given trade partner 33
Policy Implications
Interaction between trade and other policies with heterogeneous effects across �rms� Credit market imperfections� Affects smaller �rms � and generates signi�cant changes in trade patterns across sectorswith different levels of external �nance dependence
� This shows up in sectoral trade patterns (both extensive and intensive margins) and inproduct churning rates
� Labor market �exibility� Affects link between trade liberalization and unemployment in non-monotonic ways� Differences in labor market �exibility become a source of comparative advantage whensectors exhibit different intrinsic levels of �rm-level volatility
34
Policy Implications: Labor Market Flexibility and Comparative Advantage
ARG
AUS
AUT BEL
CAN
CHECZE
DEU
DNK
ESP
FINFRAGBR
GRC
HKG
IRL
ISR
ITA
JPNKOR
KWT
NLD
NOR
NZL
OMN
PRT
SAU
SGP
SVN
SWE
USA
0.0
1.0
2.0
3.0
4.0
5A
vera
ge P
urge
d C
ount
ry V
olat
ility
20 40 60 80 100Labor Market Flexibility
35