ClassNotes - To Be Checked - Maybe PHD
Uploaded by
EnkhmaaClassNotes - To Be Checked - Maybe PHD
Uploaded by
EnkhmaaElements of Advanced International Trade1
Treb Allen2 and Costas Arkolakis3
February 2016 [New version: preliminary]
1 This set of notes and the problem sets accomodating them is a collection of material de-
signed for an international trade course at the graduate level. We are grateful to Cristina Arel-
lano, Jonathan Eaton, Timothy Kehoe and Samuel Kortum. We are also grateful to Steve Redding
and Peter Schott for sharing their teaching material and to Federico Esposito, Monica Morlacco,
Ananth Ramananarayan, Olga Timoshenko, Philipp Stiel, Alex Torgovitsky, and Yipei Zhang for
valuable input and for various comments and suggestions. All remaining errors are purely ours.
2 Northwestern University and NBER
3 Yale University and NBER
Abstract
These notes are prepared for a Ph.D. level course in international trade.
i
Contents
1 An introduction into deductive reasoning 1
1.1 Inductive and deductive reasoning . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Employing and testing a model . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 International Trade: The Macro Facts . . . . . . . . . . . . . . . . . . . . . . . 3
2 An introduction to modeling 4
2.1 The Heckscher-Ohlin model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Autarky equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Free Trade Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.3 No specialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.4 Specialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.5 The 4 big theorems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 Models with Constant Elasticity Demand 16
3.1 Constant Elasticity Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 The gravity setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Armington model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.2 Gravity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
ii
3.3.3 Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 Monopolistic Competition with Homogeneous Firms and CES demand . . . 25
3.4.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4.2 Consumers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.3 Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.4 Gravity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4.5 Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.5 Ricardian model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5.1 The two goods case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5.2 Autarky . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5.3 Free trade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.6 Problem Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 Models with CES demand and production heterogeneity 36
4.1 Introduction to heterogeneity: The Ricardian model with a continuum of
goods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.1.1 Where DFS stop and EK start . . . . . . . . . . . . . . . . . . . . . . . 38
4.2 A theory of technology starting from first principles . . . . . . . . . . . . . . 41
4.2.1 Order Statistics and Various Moments . . . . . . . . . . . . . . . . . . 42
4.3 Application I: Perfect competition (Eaton-Kortum) . . . . . . . . . . . . . . . 47
4.3.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3.2 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.3 Gravity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3.4 Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3.5 Discussion of EK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Application II: Bertrand competition (Bernand, Eaton, Jensen, Kortum) . . . 60
iii
4.4.1 Supply side . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.4.2 Gravity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.4.3 Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.4.4 Discussion of BEJK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.5 Application III: Monopolistic competition with CES (Chaney-Melitz) . . . . 66
4.5.1 Model Set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.5.2 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.5.3 The Pareto Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.5.4 Trade with firm heterogeneity . . . . . . . . . . . . . . . . . . . . . . . 75
4.5.5 Next steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.7 Problem Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5 Closing the model 83
5.1 General Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2 Solving for the Equilibrium when the Profit share is constant . . . . . . . . . 86
5.3 Endogenous Entry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.4 Labor Mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.5 Problem Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6 Model Characterization 92
6.1 The Concept of a Model Isomorphism . . . . . . . . . . . . . . . . . . . . . . 92
6.1.1 An Exact Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1.2 A Partial Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.2 General Equilibrium: Existence and Uniqueness . . . . . . . . . . . . . . . . 96
6.3 Analytical Characterization of the Gravity Model . . . . . . . . . . . . . . . 100
iv
6.4 Computing the Equilibrium (Alvarez-Lucas) . . . . . . . . . . . . . . . . . . 105
6.5 Conducting Counterfactuals: The Dekle-Eaton-Kortum Procedure . . . . . . 106
6.6 Problem Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7 Gains from Trade 110
7.1 Trade Liberalization and Firm Heterogeneity . . . . . . . . . . . . . . . . . . 110
7.1.1 Trade Liberalization and Welfare gains (Arkolakis-Costinot-Rodriguez-
Clare) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.1.2 Ex-ante Gains (Atkeson Burstein 2010) . . . . . . . . . . . . . . . . . . 115
7.1.3 Discussion: gains from trade and the trade heterogeneity . . . . . . . 117
7.2 Global Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.3 Problem Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8 Extensions: Modeling the Demand Side 121
8.1 Extension I: The Nested CES demand structure . . . . . . . . . . . . . . . . . 121
8.2 Extension II: Market penetration costs . . . . . . . . . . . . . . . . . . . . . . 122
8.3 Extension III: Multiproduct firms . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.3.1 Gravity and Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
8.4 Extension IV: General Symmetric Separable Utility Function . . . . . . . . . 132
8.4.1 Monopolistic Competition with Homogeneous Firms (Krugman 79) 132
8.4.1.1 Consumer’s problem . . . . . . . . . . . . . . . . . . . . . . 132
8.4.1.2 Firm’s Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.4.1.3 Monopolistic Competition with Heterogeneous Firms (Arkolakis-
Costinot-Donaldson-Rodriguez Clare) . . . . . . . . . . . . 137
8.4.1.4 Firm Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.4.1.5 Gravity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
v
8.4.1.6 Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.5 Problem Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
9 Modeling Vertical Production Linkages 145
9.1 Each good is both final and intermediate . . . . . . . . . . . . . . . . . . . . . 145
9.2 Each good has a single specialized intermediate input . . . . . . . . . . . . . 146
9.3 Each good uses a continuum of inputs . . . . . . . . . . . . . . . . . . . . . . 147
10 The gravity estimator and trade cost estimates 150
10.1 The Head and Ries procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
10.2 Gravity estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
10.2.1 The traditional gravity estimator . . . . . . . . . . . . . . . . . . . . . 152
10.2.2 The fixed effects gravity estimator . . . . . . . . . . . . . . . . . . . . 154
10.2.3 The ratio gravity estimator . . . . . . . . . . . . . . . . . . . . . . . . 156
10.2.4 The Anderson and van Wincoop procedure . . . . . . . . . . . . . . . 157
10.3 The no arbitrage condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
10.3.1 The Eaton and Kortum (2002) approach . . . . . . . . . . . . . . . . . 160
10.3.2 The Donaldson (2014) approach . . . . . . . . . . . . . . . . . . . . . 162
10.3.3 The Allen (2012) approach . . . . . . . . . . . . . . . . . . . . . . . . . 163
11 Parameter Choices for Gravity Models 166
11.1 Recovering the trade elasticity from gravity estimates . . . . . . . . . . . . . 166
11.1.1 The Eaton and Kortum procedure(s) . . . . . . . . . . . . . . . . . . . 168
11.1.2 Gravity trade models where β 6= 0 . . . . . . . . . . . . . . . . . . . . 169
11.2 Calibration of a firm-level model of trade . . . . . . . . . . . . . . . . . . . . 173
11.3 Estimation of a firm-level model . . . . . . . . . . . . . . . . . . . . . . . . . 175
11.3.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
vi
11.3.2 Estimation, simulated method of moments . . . . . . . . . . . . . . . 179
12 Some facts on disaggregated trade flows 181
12.1 Firm heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
12.2 Trade liberalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
12.3 Trade dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
13 Appendix 194
13.1 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
13.1.1 The Fréchet Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 194
14 Figures and Tables 198
vii
List of Figures
14.1 Sales in France from firms grouped in terms of the minimum number of
destinations they sell to. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
14.2 French entrants and market size. Source: Eaton, Kortum, and Kramarz (2011).200
14.3 Distribution of sales for Portugal and means of other destinations group in
terciles depending on total sales of French firms there. . . . . . . . . . . . . . 200
viii
Chapter 1
An introduction into deductive
reasoning
1.1 Inductive and deductive reasoning
Inductive or empirical reasoning is the type of reasoning that moves from specific ob-
servations to broader generalizations and theories. Deductive reasoning is the type of
reasoning that moves from axioms to theorems and then applies the predictions of the
theory to the specific observations.
Inductive reasoning has failed in several occasions in economics. According to Prescott
(see Prescott (1998)) “The reason that these inductive attempts have failed ... is that the
existence of policy invariant laws governing the evolution of an economic system is in-
consistent with dynamic economic theory. This point is made forcefully in Lucas’ famous
critique of econometric policy evaluation.”
Theories developed using deductive reasoning must give assertions that can be fal-
sified by an observation or a physical experiment. The consensus is that if one cannot
1
potentially find an observation that can falsify a theory then that theory is not scientific
(Popper).
A general methodology of approaching a question using deductive reasoning is the
following:
1) Observe a set of empirical (stylized) facts that your theory has to address and/or
are relevant to the questions that you want to tackle,
2) Build a theory,
3) Test the theory with the data and then use your theory to answer the relevant ques-
tions,
4) Refine the theory, going through step 1
1.2 Employing and testing a model
A vague definition of two methodologies: calibration and estimation
Calibration is the process of picking the parameters of the model to obtain a match
between the observed distributions of independent variables of the model and some key
dimensions of the data. More formally, calibration is the process of establishing the re-
lationship between a measuring device and the units of measure. In other words, if you
think about the model as a “measuring device” calibrating it means to parameterize it to
deliver sensible quantitative predictions.
Estimation is the process of picking the parameters of the model to minimize a func-
tion of the errors of the predictions of the model compared to some pre-specified targets.
It is the approximate determination of the parameters of the model according to some
pre-specified metric of differences between the model and the data to be explained.
It is generally considered a good practice to stick to the following principles (see
Prescott (1998) and the discussion in Kydland and Prescott (1994)) when constructing
2
quantitative models:
1. When modifying a standard model to address a question, the modification contin-
ues to display the key facts that the standard model was capturing.
2. The introduction of additional features in the model is supported by other evidence
for these particular additional features.
3. The model is essentially a measurement instrument. Thus, simply estimating the
magnitude of that instrument rather than calibrating the model can influence the ability
of the model to be used as a measuring instrument. In addition the model’s selection
(or in particular, parametric specification) has to depend on the specific question to be
addressed, rather than the answer we would like to derive. For example, “if the question
is of the type, how much of fact X can be accounted for by Y, then choosing the parameter
values in such a way as to make the amount accounted for as large as possible according
to some metric makes no sense.”
4. Researchers can challenge existing results by introducing new quantitatively rele-
vant features in the model, that alter the predictions of the model in key dimensions.
1.3 International Trade: The Macro Facts
Chapter 2 of Eaton and Kortum (2011) manuscript
3
Chapter 2
An introduction to modeling
2.1 The Heckscher-Ohlin model
The Heckscher-Ohlin (H-O) model of international trade is a general equilibrium model
that predicts that patterns of trade and production are based on the relative factor endow-
ments of trading partners. It is a perfect competition model. In its benchmark version it
assumes two countries with identical homothetic preferences and constant return to scale
technologies (identical across countries) for two goods but different endowments for the
two factors of production. The model’s main prediction is that countries will export the
good that uses intensively their relatively abundant factor and import the good that does
not. We will present a very simple version of this model. Country i’s representative con-
sumer’s problem is
max a1 log c1i + a2 log c2i
s.t. p1 c1i + p2 c2i ≤ ri k̄i + wi l¯i
4
The production technologies of good ω in the two countries are identical and given by
bω 1 − bω
yiω = zω kiω i
lω , i, ω = 1, 2
and where 0 < b2 < b1 < 1. This implies that good 1 is more capital intensive than
good 2. Assume for simplicity that k̄1 /l¯1 > k̄2 /l¯2 . This implies that country 1 is capital
abundant relative to country 2. Finally, goods, labor, and capital markets clear. One of the
common assumptions for the H-O model is that there is no factor intensity reversal which
in our example is always the case given the Cobb-Douglas production function (one good
is always more capital intensive than the other, with the capital intensity given by bω ).
2.1.1 Autarky equilibrium
We first solve for the autarky equilibrium for country i. This is easy especially if we
consider the social planner problem, but we will compute the competitive equilibrium
instead. The Inada conditions for the consumer’s utility function imply that both goods
will be produced in equilibrium. Thus, we just have to take FOC for the consumer and
look at cost minimization for the firm. For the consumer we have
max a1 log c1i + a2 log c2i
s.t. p1i c1i + p2i c2i ≤ ri k̄i + wi l¯i
5
which implies
a1 = λi p1i c1i (2.1)
a2 = λi p2i c2i (2.2)
p1i c1i + p2i c2i = ri k̄i + wi l¯i (2.3)
This gives
a2 i i
p2i c2i = p c . (2.4)
a1 1 1
The firm’s cost minimization problem
min ri kiω + wi lω
i
bω 1 − bω
s.t. yiω ≤ zω kiω i
lω
implies the following equation, under the assumption that both countries produce both
goods,
bω
l i wi = ri kiω . (2.5)
( 1 − bω ) ω
We can also use the goods market clearing to obtain
bω 1 − bω
ciω = zω kiω i
lω =⇒
b
bω w i ω
i i
cω = zω lω . (2.6)
1 − bω r i
bω 1 − bω
Zero profits in equilibrium, piω zω kiω i
lω = ri kiω + wi lωi , combined with (2.5),
6
give us
ri kiω
piω = bω
i i (1− b )
bω zω rwkiω bω ω (1−bωbω ) wri
i
1 − bω i bω
wi r
=
z ω ( 1 − b ω ) 1 − bω ( b ω ) bω
We can also derive the labor used in each sector. From the consumer’s FOCs, together
with the expressions for piω and ciω derived above, we obtain:
aω = λi piω ciω =⇒
aω
( 1 − bω ) = wi lωi
λi
this implies that
(1 − b2 ) a2
l1i = l2i
(1 − b1 ) a1
We can use the labor market clearing condition and get
l2i + l1i = l¯i =⇒
(1 − b2 ) a2
l1i + l1i = l¯i =⇒
(1 − b1 ) a1
(1 − b1 ) a1
l1i = l¯i . (2.7)
(1 − b2 ) a2 + (1 − b1 ) a1
The results are similar for capital and thus,
b1 a1
ki1 = k̄i ,
b1 a1 + b2 a2
7
This implies
l¯i r i ∑ ω ( 1 − bω ) a ω
= (2.8)
k̄i wi ∑ ω bω a ω
Thus, in a labor abundant country capital is relatively more expensive as we would ex-
pect. We can finally use the goods’ market clearing conditions combined with the optimal
i , ki to get the values for ci ’s as a function solely of parameters and endow-
choices for lω ω ω
ments,
bω 1 − bω
(1 − bω ) aω ¯i
bω a ω
ciω = zω k̄i l . (2.9)
∑ω ω ω
0 b 0 a 0 ∑ ω 0 ( 1 − bω 0 ) a ω 0
2.1.2 Free Trade Equilibrium
In the two country example, free trade implies that the price of each good is the same in
both countries. Therefore, we will denote free trade prices without a country superscript.
In the two country case it is important to distinguish among three conceptually different
cases: in the first case both countries produce both goods, in the second case one country
produces both goods and the other produces only one good, and in the last case each
country produces only one good.
We first define the free trade equilibrium. A free trade equilibrium is a vector of allo-
cations for consumers ĉiω , i, ω = 1, 2 , allocations for the firm k̂iω , lˆω
i , i, ω = 1, 2 , and
prices ŵiω , r̂ω , p̂ω , i, ω = 1, 2 such that
1. Given prices consumer’s allocation maximizes her utility for i = 1, 2
2. Given prices the allocations of the firms solve the cost minimization problem in
8
i = 1, 2,
bω − 1 1 − bω
bω pω zω kiω i
lω ≤ ri , with equality if yiω > 0
bω − bω
(1 − bω ) pω zω kiω lωi
≤ wi , with equality if yiω > 0
3. Markets clear
∑ ĉiω = ∑ ŷiω , ω = 1, 2
i i
∑ k̂iω = k̄i for each i = 1, 2
ω
∑ lˆωi = l¯i for each i = 1, 2 .
ω
2.1.3 No specialization
We analyze the three cases separately. First, let’s think of the case in which both countries
produce both goods.
max a1 log c1i + a2 log c2i
s.t. p1 c1i + p2 c2i ≤ ri k̄i + wi l¯i
a1 = λi p1 c1i (2.10)
a2 = λi p2 c2i (2.11)
p1 c1i + p2 c2i = ri k̄i + wi l¯i (2.12)
9
This implies again that
a2
p2 c2i = p1 c1i (2.13)
a1
When both countries produce both goods the firms cost minimization problem implies
the following two equalities,
bω − 1 1 − bω
bω pω zω kiω i
lω = ri ,
bω − bω
(1 − bω ) pω zω kiω lωi
= wi ,
which in turn imply
i wi
bω l ω
= kiω . (2.14)
( 1 − bω ) r i
Additionally, from zero profits,
bω 1 − bω
ri wi
pω = (2.15)
z ω ( b ω ) bω ( 1 − b ω ) 1 − bω
and, of course, technologies (by assumption) and prices (due to free trade) are the same
in the two countries. Notice that the equality (2.15) is true for i = 1, 2 this implies that
bω 1 − bω b 1 − bω
r1 w1 = r 2 ω w2 ω = 1, 2
1 bω 2 1 − bω
r w
= ω = 1, 2
r2 w1
Noticing that the above expression holds for ω = 1, 2 and replacing these two equa-
10
tions in one another we have
(1−bb2 )b1 −1+b1
w2
2
= 1 =⇒
w1
w2 = w1
and of course
r2 = r1 .
This shows that we have factor price equalization (FPE) in the free trade equilibrium.
From the cost minimization of the firm we have
bω 1 − bω
bω pω zω kiω i
lω = ri kiω =⇒
bω pω yiω = ri kiω =⇒
ri kiω
pω yiω = .
bω
Summing up over i and using FPE we have
! !
r
pω ∑ yiω =
bω ∑ kiω . (2.16)
i i
The equations (2.10) and (2.11) imply
aω aω
1 2
+ = p ω c + c (2.17)
λ1 λ2 ω ω
11
Using goods market clearing, ∑i ciω = ∑i yiω , we have
aω r
∑ λi = pω c1ω + c2ω = pω ∑ yiω =
bω ∑ kiω =⇒
i i i
1
bω a ω ∑ = r ∑ kiω =⇒
i
λi i
!
1
∑ λi ∑ bω aω = r ∑ ∑ kiω =⇒
i ω ω i
r k̄1 + k̄2
1
∑ λi
=
∑ ω bω a ω
(2.18)
i
and in a similar manner
w l¯1 + l¯2
1
∑ λi
=
∑ ω ( 1 − bω ) a ω
. (2.19)
i
Using (2.18) and (2.19) we can determine the w/r ratio
l¯1 + l¯2 r ∑ ω ( 1 − bω ) a ω
= . (2.20)
1
k̄ + k̄ 2 w ∑ ω bω a ω
Assuming that one country is more capital abundant than the other (say k̄1 /l¯1 > k̄2 /l¯2 ),
the equilibrium factor price ratio r/w under free trade lies in between the autarky factor
prices of the two countries (determined in equation 2.8).
Using the relationships for the capital labor ratio (2.14) together with the above ex-
pression and factor market clearing conditions we can derive the equilibrium labor used
from each country in each sector. Using the capital labor ratios for both goods and for
12
both countries we get:
w ¯i i
b1
i b2
l − l2 + l2 = k̄i
r (1 − b1 ) (1 − b2 )
l2i (1 − b2 ) (1 − b1 ) r k̄i
b1
= −
l¯i b2 − b1 w l¯i (1 − b1 )
i
(1 − b2 ) (1 − b1 ) ∑ ω bω a ω l¯1 + l¯2 k̄i
l2 b1
= −
l¯i b1 − b2 (1 − b1 ) ∑ω (1 − bω ) aω k̄1 + k̄2 l¯i
You may notice two things in this expression. First, if initial endowments of the two
countries are inside a relative range, there is diversification since l ji > 0. If the endow-
ments of a country for a given good are not in this range, then a country specializes in the
other good (this range of endowments that implies diversification in production is com-
monly referred to as the cone of diversification). Second, conditional on diversification
labor abundant countries use relatively more labor in the labor intensive sector.
What is the share of consumption for each country? We can use the FOC from the
consumer’s problem to obtain
a2
p1 c1i 1+ = r k̄i + wl¯i =⇒
a1
r k̄i + wl¯i
c1i = =⇒
p1 1 + aa21
1 wl¯i
c1i = (2.21)
( 1 − bω ) p 1 1 + a2
a1
where in the last equivalence we used equation (2.14). Obtaining the rest of the alloca-
tions and prices is straightforward. In fact, you can show that if the production function
exhibits CRS and the capital-labor ratio for both countries is fixed (in a given sector), total
13
production can be represented by1
! bω ! 1 − bω
yω = zω ∑ kiω ∑ lωi .
i i
i by combining expression (2.18) with (2.16), (2.17) and using
We can determine ∑i kiω , ∑i lω
the market clearing condition. This gives
bω a ω ∑ ki
= i ωi , (2.22)
∑ ω bω a ω ∑i k̄
and similarly for labor.
2.1.4 Specialization
[HW]
1 Assume that k1 /l 1 = k2 /l 2 . We only have to prove that given this assumption
b 1− b b 1− b b 1− b
A k1 + k2 l1 + l2 = A k1 l1 + A k2 l2 .
Using repeatedly the condition we have that
!b !b 2 b
k1 + k2 k1 l1 k l2
A =A + A ⇐⇒
l1 + l2 l1 l1 + l2 l2 l1 + l2
!b !b
k1 + k2 k1
= ⇐⇒
l1 + l2 l1
! !
k2 l 1 /l 2 + k2 k1
= ⇐⇒
l1 + l2 l1
k2 k1
2
= 1 ,
l l
which holds by assumption completing the proof.
14
2.1.5 The 4 big theorems.
In this final section for the H-O model we will state main theorems that hold in the bench-
mark model with two countries and two goods. Variants of these theorems hold under
less or more restrictive assumptions. Our approach will still be as parsimonious as possi-
ble.2
Theorem 1. Assume countries engage in free trade, there is no specialization (thus there is diver-
sification) in equilibrium and there is no factor intensity-reversal, then factor prices equalize across
countries.
Proof. See main text
Theorem 2. (Rybczynski) Assume that the economies remain always incompletely specialized.
An increase in the relative endowment of a factor will increase the ratio of production of the good
that uses the factor intensively.3
Proof. TBD
Theorem 3. (Stolper-Samuelson) Assume that the economies remain always incompletely spe-
cialized. An increase in the relative price of a good increases the real return to the factor used
intensively in the production of that good and reduces the real return to the other factor.
Proof. TBD
Theorem 4. (Hekchser-Ohlin) Each country will produce the good which uses its abundant factor
of production more intensively.
Proof. TBD
2 For a detailed treament you can look at the books of Feenstra (2003) and Bhagwati, Panagariya, and
Srinivasan (1998).
3 If prices were fixed a stronger version of the theorem can be proved.
15
Chapter 3
Models with Constant Elasticity
Demand
Suppose there is a compact set S of countries. For now, we assume that S is discrete,
although having a continuum of countries does not change things much. Whenever it is
possible, we refer to an origin country as i and a destination country as j and order the
subindices such that Xij is the bilateral trade from i to j. We define as X j the total spending
of country j. We further denote by L j the population of country j and let each consumer
have a single labor unit that is inelastically supplied.
There are three common assumptions made about the market structure by trade theo-
rists. The first is that markets in every country are perfectly competitive, so the price of a
good is simply equal to its marginal cost. The second is that there is Bertrand competition
so that the price of a good depends on the marginal cost of the least cost producer as well,
potentially, on the cost of the second cheapest producer. The third is that production is
monopolistically competitive so that the firm does not perceive any immediate competitor
but it is affected by the overall level of competition. We will consider each below.
16
We also assume throughout that labor is the only factor of production (we will add
intermediate inputs later on). We also assume that there are iceberg trade costs {τij }i,j∈S .
This means that in order from one unit of a good to arrive in destination j, destination i
must ship τij units. Iceberg trade costs are so called because a fraction τij − 1 “melts” on
its way from i to j, much as if you were towing an iceberg. We almost always assume that
τij ≥ 1 and usually assume that τii = 1 for all i ∈ S, i.e. trade with oneself is costless.
Furthermore, we sometimes assume that the following triangle inequality holds: for all
i, j, k ∈ S: τij τjk ≥ τik . The triangle inequality says that it is never cheaper to ship a good
via an intermediate location rather than sell directly to a destination.
3.1 Constant Elasticity Demand
We first introduce one of the most remarkably simple as well as versatile demand func-
tions that will be the basis of our analysis for the next two chapters, the Constant Elas-
ticity of Substitution (CES) demand function. Why do we do so? CES preferences have
a number of attractive properties: (1) they are homothetic; (2) they nest a number of spe-
cial demand systems (e.g. Cobb-Douglas); and (3) they are extremely tractable. Trade
economists often do not believe that CES preferences are a good representation of actual
preferences but are seduced into making frequent use of them due to their analytical con-
venience.
In particular, assume that the representative consumer in country j derives utility Uj
from a set of varieties Ω the consumption of goods shipped from all countries i ∈ S:
! σ−σ 1
1 σ −1
Uj = ∑ aij (ω ) qij (ω )
σ σ , (3.1)
ω ∈Ω
where σ ≥ 0 is the elasticity of substitution and aij (ω ) is an exogenous preference shifter.
17
A couple of things to note: first, qij (ω ) is the quantity of a good shipped from i that ar-
rives in j (the amount shipped is τij qij (ω )); second, the fact that there is a representative
consumer is not particularly important: we can always assume that workers (with identi-
cal preferences) are the ones consuming the goods and Uj can be interpreted as the total
welfare of country j in this case (homotheticity is, of course, crucial for this property to be
true).
We now solve the representative consumer’s utility maximization problem. Given the
importance of CES in the class we will proceed to do the full derivation for any given
good ω ∈ Ω. Let the spending of country j be denoted X j and let the price of a good (net
of trade costs) from country i in country j be pij . Then the utility maximization problem
is: ! σ−σ 1
1 σ −1
max ∑ aij (ω ) σ qij (ω ) σ s.t. ∑ qij (ω ) pij (ω ) ≤ X j ,
{qij (ω )}ω∈Ω i ∈Ω ω ∈Ω
where I ignore the constraint that qij (ω ) > 0 (why is this okay?).
The Lagrangian is:
! σ−σ 1 !
1 σ −1
L: ∑ aij (ω ) qij (ω )
σ σ −λ ∑ qij (ω ) pij (ω ) − X j
ω ∈Ω ω ∈Ω
First order conditions (FOCs) imply:
! σ−1 1
∂L 1 σ −1 1 1
∂qij (ω )
= 0 ⇐⇒ ∑ aij (ω ) qij (ω )
σ σ aij (ω ) σ qij (ω )− σ = λpij (ω )
ω ∈Ω
∂L
= 0 ⇐⇒ X j = ∑ qij (ω ) pij (ω )
∂λ ω ∈Ω
18
Using the FOCs for two different goods, ω and ω 0 :
aij (ω ) pijσ (ω ) qij (ω )
= σ 0
aij (ω 0 ) pij (ω ) qij (ω 0 )
Rearranging and multiplying both sides by pij (ω ) yields:
1
qij ω 0 pij ω 0 = qij (ω ) pij (ω )σ (ω ) aij ω 0 p1ij−σ ω 0
aij (ω )
Summing over all ω 0 ∈ Ω yields:
1 1− σ
∑ qij ω 0 pij ω 0 = qij (ω ) pij (ω )σ ∑ aij ω 0 pij ω 0
⇐⇒
ω ∈Ω
0 aij (ω ) ω 0 ∈Ω
1
Xj = qij (ω ) pij (ω )σ Pj1−σ
aij (ω )
1
1 − σ 1− σ
where the last line used the second FOC and Pj ≡ ∑ω 0 ∈Ω aij (ω 0 ) pij (ω 0 ) is known
Xj
as the Dixit-Stiglitz price index. It is easy to show that Uj = Pj , i.e. dividing income by
the price index gives the total welfare of country j. Rearranging the last line yields the
CES demand function:
qij (ω ) = aij (ω ) pij−σ (ω ) X j Pjσ−1 , (3.2)
Equation (3.2) implies that the quantity consumed in j of a good produced in i will be
increasing with j0 s preference for the good (aij ), decreasing with the price of the good
( pij ), increasing with j0 s spending ( X j ), and increasing with j0 s price index.
Note that the value of total trade is simply equal to the price times quantity. In what
follows, let us denote the value of trade of good ω from country i to country j as Xij (ω ) ≡
19
pij (ω ) qij (ω ). Then we have:
Xij (ω ) = aij (ω ) p1ij−σ (ω ) X j Pjσ−1 . (3.3)
The only thing left to construct bilateral trade is to solve for the optimal price and aggre-
gate across varieties, which we will do after a brief discussion of the gravity equation.
3.2 The gravity setup
It is helpful to provide a brief motivation of why we are interested in writing down a flex-
ible model in the first place. “Classical” trade theories (Ricardo, Heckscher-Ohlin), while
extremely useful in highlighting the economic forces behind trade, are very difficult to
generalize to a set-up with many trading partners and bilateral trade costs. Because the
real world clearly has both of these, the classical theories do not provide much guidance
in doing empirical work. Because of this difficulty, those doing empirical work in trade
began using a statistical (i.e. a-theoretic) model known as the gravity equation due to its
similarity the Newton’s law of gravitation. The gravity equation states that total trade
flows from country i to country j, Xij , are proportional to the product of the origin coun-
try’s GDP Yi and destination country’s GDP Yj and inversely proportional to the distance
between the two countries, Dij :1
Yi × Yj
Xij = α . (3.4)
Dij
For a variety of reasons (which we will go into later on in the course), this gravity equation
is often estimated in a more general form, which we refer to as the generalized gravity
1 This is actually in contrast to Newton’s law of gravitation, where the force of gravity is inversely propor-
tional to the square of the distance.
20
equation:
Xij = Kij γi δj , (3.5)
where Kij is a measure of the resistance of trade between i and j, γi measures the origin
“size” and δj measures the destination “size” (note that each country has two different
measures of size)..
The gravity equation (3.4) and its generalization (3.5) have proven to be enormously
successful at explaining a large fraction of the variation in observed bilateral trade flows;
indeed, it is probably not too much of an exaggeration to say that the gravity equation
is one of the most successful empirical relationships in all of economics. Because it was
originally proposed as a statistical relationship, however, the absence of a theory justifying
the relationship made it very difficult to ask any meaningful counterfactual questions; e.g.
“what would happen to trade between i and j if the tariff was lowered between i and k?”
3.3 Armington model
The Armington model (Armington, 1969) is based on the premise that each country pro-
duces a different good and consumers would like to consume at least some of each coun-
try’s goods. This assumption is of course ad hoc, and it completely ignores the “classical”
trade forces such as increased specialization due to comparative advantage. However,
as we will see, the model (when combined with Constant Elasticity of Substitution (CES)
preferences as in (Anderson, 1979)) provides a nice characterization of trade flows be-
tween many countries.2
The Armington model (as formulated by (Anderson, 1979)) was important because
2 Actually, in the main text, Anderson (1979) considers Cobb-Douglass preferences and writes that “there
is little point in the exercise” of generalizing to CES preferences, doing so only in an appendix. Despite his
reluctance to do so, the paper has been cited thousands of times as the example of an Armington model with
CES preferences.
21
it provided the first theoretical foundation for the gravity relationship. It is also a great
place to start our course, as one of the great surprises of the international trade literature
over the past fifteen years has been how robust the results first present in the Armington
model are across different quantitative trade models. By now, as we will discuss in this
chapter, models that yield the gravity relationship (3.5) are ubiquitous and much of the
rest of what follows will focus on analyzing their common properties.
3.3.1 The model
We now turn to the details of the Armington models and in particular to the supply side
of this model, given CES demand.
The Armington assumption is that each country i ∈ S produces a distinct variety of
a good. Because countries map one-to-one to varieties, we index the varieties by their
country names (this will not be true for Bertrand and monopolistic competition when we
have to keep track both of varieties and countries).
Suppose that the market for each country/good is perfectly competitive, so that the
price of a good is simply the marginal cost. Suppose each worker can produce Ai units
of her country’s good and let wi be the wage of a worker. Then the marginal cost of
wi
production is simply Ai . This implies that the price at the factory door (i.e. without
wi
shipping costs) is pi = Ai . What about with trade costs? Recall that with the iceberg
formulation, τij ≥ 1 units have to be shipped in order for one unit to arrive. This means
that τij ≥ 1 units have to be produced in country i in order for one unit to be consumed in
country j. Hence the price in country j of consuming one unit from country i is:
wi
pij = τij . (3.6)
Ai
22
Note that this implies that:
pij
= τij , (3.7)
pi
i.e. the ratio of the price in any destination relative to the price at the factory door is
simply equal to the iceberg trade cost. Equation (3.7) is called a no-arbitrage equation, as
it means that there is no way for an individual to profit by buying a good in country i and
sell in country j (or vice versa). Note, however, that there may still be profitable trading
opportunities between triplets of countries even if equation (3.7) holds when the triangle
inequality is not satisfied.
3.3.2 Gravity
Assuming that each country produces a different good ω, and substituting equation (3.6)
into equation (3.3) yields a gravity equation for bilateral trade flows:
1− σ
wi
Xij = aij τij1−σ X j Pjσ−1 . (3.8)
Ai
To the extent that trade costs are increasing in distance, the value of bilateral trade flows
will decline as long as σ > 1.
We can actually use equation to get a little close to the true gravity equation. The total
income in a country is equal to its total sales:
1− σ
wi
Yi = ∑ Xij = ∑ aij τij1−σ
Ai
X j Pjσ−1 ⇐⇒
j j
1− σ
wi
= Yi / ∑ aij τij1−σ X j Pjσ−1
Ai j
23
Replacing this expression in the equation (3.8) yields:
! !
Yi Xj
Xij = aij τij1−σ , (3.9)
Π1i −σ Pj1−σ
where Π1i −σ ≡ ∑ j aij τij1−σ X j Pjσ−1 bears a striking resemblance to the price index [insert
foreshadowing here]. Equation (3.9) which shows that the bilateral trade spending is
related to the product of the GDPs of the two countries (gravity!!), the distance/tradecost
and a GE component.
Equation (3.9) is actually about as close as we will ever get to the original gravity
equation. This is because all of our theories say that bilateral trade flows depend on more
than just the bilateral trade costs and the incomes of the exporter and importer; what also
matters is so-called “bilateral resistance”: intuitively, the greater the cost of exporting in
general, the smaller the Π1i −σ ; conversely, the greater the cost of importing in general, the
smaller the Pi1−σ . This means that trade between any two countries depends not only on
the incomes of those two countries but also the “cost” of trading between those countries
relative to trading with all other countries. This point was made in the enormously famous
and influential paper “Gravity with Gravitas: A Solution of the Border Puzzle” (Anderson
and Van Wincoop, 2003).
3.3.3 Welfare
We will now show that welfare in relationship to trade is given by a simple equation
involving the trade to GDP ratio and parameters of the model (but no other equilibrium
variables). We will be revisiting this relationship multiple times in these notes. To begin
24
define λij as the fraction of expenditure in j spent on goods arriving from location i:
Xij
λij ≡ .
∑k Xkj
From equation (3.8) we have:
1− σ
wi
aij τij1−σ Ai
λij = 1−σ ⇐⇒
wk
∑k akj τkj1−σ Ak
1− σ
wi
λij = aij τij1−σ Aiσ−1 (3.10)
Pj
1− σ
wk
since Pj1−σ ≡ ∑k akj τkj1−σ Ak . Remember from the CES derivations above that the
Xj
utility of the representative agent is the real wage, i.e. Wj = Pj . Assume that τjj = 1. Then
by choosing i = j, equation (3.10) implies that welfare can be written as:
X j 1−1 σ σ−1 1
Uj = λ a Aj, (3.11)
w j jj jj
i.e. welfare depends only on changes in the trade to GDP ratio, λ jj , with an elasticity of
−1/ (σ − 1) which is the inverse of the trade elasticity.
3.4 Monopolistic Competition with Homogeneous Firms and CES
demand
With the Armington model, we saw how we could justify the gravity relationship in trade
using the ad-hoc assumption that every country produces a unique good as well as the
assumption that consumers have a “love of variety” (i.e. they want to consume at least a
little bit of every one of the goods). In this section, we will dispense of the first assumption
25
by introducing firms into the model. However, we will continue to rely heavily on the
second assumption by assuming that each firm produces a unique variety and consumers
would like to consume at least a little bit of every variety.
The model considered today was introduced by Krugman (1980) and was an impor-
tant part of the reason he won a Nobel prize. A key feature of the Krugman (1980) model
is that there are increasing returns to scale, i.e. the average cost of production is lower the
more that is produced. All else equal, this will lead to gains from trade, since by taking
advantage of demand from multiple countries, firms can lower their average costs. To
succinctly model the increasing returns from scale, we suppose that a firm has to incur
a fixed entry cost f ie in order to produce. (The e might seem like unnecessary notation;
however, we keep it here because in future models there will be both an entry cost and
a fixed cost of serving a particular destination). We assume the fixed cost of entry (like
the marginal cost) is paid to domestic workers so that f ie is the number of workers em-
ployed in the entry sector (think of them as the workers who build the firm). Some of the
results toward the end of this section are based on the subsequent analysis of Arkolakis,
Demidova, Klenow, and Rodrı́guez-Clare (2008).
3.4.1 Setup
The main departure from the perfect competition paradigm is that in monopolistic com-
petition each differentiated variety is produced (potentially) by a different firm, where
there is a measure Ni of firms in country i. This number of firms is determined in equilib-
rium by allowing firms to enter after incurring a fixed cost of entry in terms of domestic
labor, f ie .
26
3.4.2 Consumers
As in the Armington model, we assume that consumers have CES preferences over vari-
eties. Hence a representative consumer in country j ∈ S gets utility Uj from the consump-
tion of goods shipped by all other firms in all other countries, where:
ˆ ! σ−σ 1
σ −1
Uj = ∑ Ωi
qij (ω ) σ dω , (3.12)
i ∈S
where q j (ω ) is the quantity consumed in country j of variety ω. Note that for simplicity, I
no longer include a preference shifter (although one could easily be incorporated) so that
consumers treat all firms in all countries equally.
The consumer’s utility maximization problem is very similar to the Armington model
(which shouldn’t be particularly surprising, given preferences are virtually the same). In
particular, a consumer in country j ∈ S optimal quantity demanded of good ω ∈ Ω is:
qij (ω ) = pij (ω )−σ X j Pjσ−1 ,
where:
ˆ ! 1−1 σ
Pj ≡ ∑ Ωi
pij (ω ) 1− σ
dω (3.13)
i ∈S
is the Dixit-Stiglitz price index.
The amount spent on variety ω is simply the product of the quantity and the price:
xij (ω ) = pij (ω )1−σ X j Pjσ−1 . (3.14)
Note that we derived a very similar expression in the Armington model, from which the
gravity equation followed almost immediately. In this model however, this is the amount
spent on the goods from a particular firm, so we now need to aggregate across all firms in
27
country i to determine bilateral trade flows between i and j, i.e.:
ˆ ˆ
Xij ≡ xij (ω ) dω = X j Pjσ−1 pij (ω )1−σ dω. (3.15)
Ωi Ωi
3.4.3 Firms
All firms in country i have a common productivity, zi , and produce one unit of the good
1
using zi units of labor. The optimization problem faced by a firm ω from country i is:
τij
max ∑ p j (ω ) q j (ω ) − wi q j (ω ) − wi f ie s.t. q j (ω ) = p j (ω )−σ X j Pjσ−1
{ p j ( ω ) } j∈S j ∈ S zi
We can substitute the constraint into the maximand and write the equivalent uncon-
strained problem of choosing the price to sell to each location as:
τij σ
max ∑ p1j −σ (ω ) X j Pjσ−1 − wi p −
zi j
(ω ) X j Pjσ−1 − wi f ie
{ p j ( ω ) } j∈S j ∈ S
Note that the constant marginal cost assumption implies that the country can treat each
destination as a separate optimization problem (this will come in helpful in models we
will see later on).3 Profit maximization implies that optimal pricing for a firm selling from
country i to country j is
σ τij wi
pij (zi ) = , (3.16)
σ − 1 zi
3 Notice that here we haven’t introduced fixed costs of exporting. Introducing these costs will change the
analysis in that we may have countries for which all the firms chose not to export depending on values of
the fixed costs and other variables. More extreme predictions can be delivered if the production cost f i is
only a cost to produce domestically and independent of the exporting cost. However, in order to create a true
extensive margin of firms (i.e. more firms exporting when trade costs decrease) requires heterogeneity either
in the productivities of firms (as we will do later on in the notes) or in the fixed costs of selling to a market
(see Romer (1994)).
28
and since all firm decision will depend on parameter’s and firm productivity we drop the
ω notation from here on. We will make the notation a bit cumbersome by carrying around
the zi ’s in order to allow for direct comparison of our results with the heterogeneous firms
example that will be studied later on.
3.4.4 Gravity
Because every firm is charging the same price, we can substitute the price equation (3.16)
into the gravity equation (3.15) to yield:
ˆ 1− σ
σ wi
Xij = X j Pjσ−1 τij dω ⇐⇒
Ωi σ − 1 zi
1− σ 1− σ
σ wi
Xij = τij1−σ Ni X j Pjσ−1 (3.17)
σ−1 zi
´
where Ni ≡ Ωi dω is the measure of firms producing in country i. Comparing this equa-
tion to the one derived for the Armington model with monopolistic, we see that the two
expressions are nearly identical - the only difference here is that we have to keep track of
the mass of firms Mi and all trade flows are smaller (if σ > 1) as a result of the markups.
3.4.5 Welfare
It turns out welfare can be written similarly to the Armington model. First, note that
substituting equation (3.16) for the equilibrium price charged into the price index equation
(3.13) yields:
1− σ 1− σ
σ wk
Pj1−σ ≡
σ−1 ∑ τkj1−σ zk
Nk .
k
Xij
As above, define λij ≡ ∑k Xkj
to be the fraction of expenditure of country j on goods
sent from country i. Then using equation (3.17), we can write λij as a function of the price
29
index in j:
1− σ
σ 1− σ 1− σ wi
Ni X j Pjσ−1
σ −1 τij zi
λij = 1− σ
σ 1− σ 1− σ wk
∑k Nk X j Pjσ−1
σ −1 τkj zk
1− σ
wi
τij1−σ zi Ni
= 1− σ
w
∑k τkj1−σ zkk Nk
1− σ τ 1− σ wi 1− σ N
σ ij zi i
= . (3.18)
σ−1 Pj1−σ
This equation can be rewritten as
σ wi 1 1
Pj = τij Ni1−σ λijσ−1 . (3.19)
σ−1 zi
Since equation (3.19) holds for any i and j, we can focus on the particular case where i = j.
Then assuming τjj = 1, we can write equation (3.19) as:
wj
1 1
σ
Pj = Nj1−σ λ jjσ−1 ⇐⇒
σ−1 zj
wj σ−1
1 1
= z j Njσ−1 λ jj1−σ , (3.20)
Pj σ
i.e. the real wage is declining in λ jj or equivalently, increasing in trade openness. Note,
however, that unlike the Armington model, firms are making positive profits, so that
the real wage no longer captures the welfare of a location. To deal with this issue, we
introduce a free entry condition.
30
3.5 Ricardian model
The Ricardian model is a model of perfect competition where countries produce the same
goods using different technologies. The Ricardian model predicts that countries may spe-
cialize in the production of certain ranges of goods.
3.5.1 The two goods case
We consider the simple version of the model with two countries and two goods. In order
to get as much intuition as possible we will first consider the case where both countries
specialize in the production of one good.
The production technologies in the two countries i = 1, 2 are different for the two
goods ω = 1, 2 and given by
yi (ω ) = zi (ω ) l i (ω ) , i, ω = 1, 2 .
Assume that country 1 has absolute advantage in the production of both goods
z2 (1) < z1 (1) ,
z2 (2) < z1 (2) .
Assume that country 1 has comparative advantage in the production of good 1 and coun-
try 2 in good 2
z1 (2) z1 (1)
2
< 2 . (3.21)
z (2) z (1)
31
Assume Cobb-Douglas preferences. The consumer’s problem is
max a (1) log ci (1) + a (2) log ci (2)
s.t. p (1) ci (1) + p (2) ci (2) ≤ wi l¯i .
Consumer optimization implies that
a (2)
p (2) c i (2) = p (1) c i (1) (3.22)
a (1)
p (1) ci (1) + p (2) ci (2) = wi l¯i (3.23)
3.5.2 Autarky
Using firms cost minimization and the Inada conditions (that ensure that the consumer
actually wants to consume both goods) from the consumer problem we directly obtain
that
p i (1) z i (1) = w i = p i (2) z i (2) .
Using the goods market clearing
ci (ω ) = yi (ω ) for ω = 1, 2 ,
together with labor market clearing
l i (ω ) = a (ω ) l¯i ,
we get labor allocated to each good. Using the production function and goods market
32
clearing we can obtain the rest of the allocations.
3.5.3 Free trade
Under free trade international prices equalize. Relative productivity patterns will deter-
mine specialization. There can be three possible specialization patterns, two where one
country specializes and the other diversifies and one where both countries specialize.
Proposition 1. [Specialization]Under the assumptions stated, at least one country specializes in
the free trade equilibrium.
Proof. If not then the firm’s cost minimization together with the consumer FOCs would
imply
z1 (2) z1 (1)
= ,
z2 (2) z2 (1)
a contradiction.
In the three different equilibria that can emerge the countries export what they have
comparative advantage on (specialization into exporting). Under free trade this relative
price has to be in the range (given the Inada conditions in consumption):
z1 (2) p (2) z1 (1)
≤ ≤
z2 (2) p (1) z2 (1)
To consider an example of how the wages are determined notice that for the country that
is under incomplete specialization equations cost minimization implies
p (1) z i (1) wi p (1) z i (2)
= =⇒ = ,
p (2) z i (2) wi p (2) z i (1)
i.e. this country sets the relative price of the two goods. Now assume that country 1 is
33
incompletely specialized which means that country 2 specializes in good 2 and normalize
w1 = 1. Because of free trade and perfect competition it must be the case that the cost of
producing good 2 in both countries is the same, i.e.
w1 w2 2 z2 (2)
= =⇒ w = < 1 = w1 .
z1 (2) z2 (2) z1 (2)
Notice that using the wages and the zero profit conditions for country 1 we now get
p (1) z1 (1) = 1 and p (2) z1 (2) = 1
z1 (1) p (2)
1
= .
z (2) p (1)
Finally using the budget constraints of the individual we can determine the levels of con-
sumption and verify that the equilibrium is consistent with our initial assumption for the
patterns of specialization (i.e. indeed country 2 exports good 2 and country 1 exports
good 1)
3.6 Problem Sets
1. Dixit-Stiglitz Preferences. Suppose that a consumer has wealth W, consumes from a
set of differentiated varieties ω ∈ Ω, and solves the following CES maximization
problem:
ˆ σ−σ 1 ˆ
σ −1
max U = q (ω ) σ dω s.t. p (ω ) q (ω ) dω ≤ X, (3.24)
{q(ω )} Ω Ω
where σ > 0, q (ω ) is the quantity consumed of variety ω and p (ω ) the price of
variety ω.
34
W
(a) Find a price index P such that in equilibrium U = P.
(b) Derive the optimal q (ω ) as a function of W, P and p (ω ).
(c) Show
that σ is the elasticity of substitution, i.e. for any ω, ω 0 ∈ Ω, σ =
q(ω )
d ln
q(ω 0 )
.
∂U/∂q(ω 0 )
d ln ∂U/∂q(ω )
(d) What happens as σ → ∞? σ → 1? σ → 0?
35
Chapter 4
Models with CES demand and
production heterogeneity
The purpose of this chapter is to develop a general model for production heterogene-
ity in which different assumptions on technology and competition will give us different
workhorse frameworks important for the quantitative analysis of trade. Our analysis of
the general framework is based on the exposition of (Eaton and Kortum (2011)) and ear-
lier results of (Kortum (1997)) and (Eaton and Kortum (2002)). We start with a simple
extension of the Ricardian model with intra-sector heterogeneity.
4.1 Introduction to heterogeneity: The Ricardian model with a
continuum of goods
The model of Dornbusch, Fischer, and Samuelson (1977) is based on the Ricardian model
where trade and specialization patterns are determined by different productivities.1 There
1 The notes in this chapter are partially based on Eaton and Kortum (2011).
36
is absolute advantage due to higher productivity in producing certain goods, but also
comparative advantage due to lower opportunity cost of producing some goods. The
main drawback of the simple Ricardian model, similar to that of the Heckscher-Ohlin
model, is in the complexity of solving for the patterns of specialization for a large number
of industries.
Breakthrough: Dornbusch, Fischer, and Samuelson (1977) used a continuum of sectors.
The characterization of the equilibrium ended up being very easy.
• Perfect competition
• 2 countries (H, F)
• Continuum of goods ω ∈ [0, 1]
• CRS technology (labor only)
• Cobb-Douglas Preferences with equal share in each good
• Iceberg trade costs τHF , τFH
We normalize the domestic wage to 1. We want to characterize the set of goods pro-
duced and exported from each country. Without loss of generality we will characterize
production and exporting for country F. We first need to compare the price of a good
ω potentially offered by country H to country F to the corresponding price of the good
produced by F in order to determine the set of goods produced by country F in equi-
librium. For this purpose, we will order the goods in a decreasing order of domestic to
foreign productivity and define ω as the good with the lowest productivity produced in
the foreign country. Thus, the foreign country produces goods [ω, 1] while the domestic
[0, ω̄ ]. When trade costs exist then the two sets will overlap, ω̄ > ω, but if τHF = τFH then
ω̄ = ω. A simple condition that determines which are the goods that will be produced by
37
country F dictates that the price of these goods in country F has to be lower than the price
of imported goods, i.e.
wF τHF z (ω ) τ
< =⇒ A (ω ) ≡ H > HF .
z F (ω ) z H (ω ) z F (ω ) wF
Therefore, we can define
wF
A (ω ) = . (4.1)
τHF
which determines that only products that will be produced by country F.
To find the set of goods that F will be exporting we need to determine set of goods
produced by country H. Using a similar logic this simply entails finding the ω that satis-
fies
A (ω̄ ) = τFH w F (4.2)
and [0, ω̄ ] is the set of goods produced by the home country. Thus, F produces goods
[ω, 1] and exports [ω̄, 1] since the domestic does not produce any of the goods in that last
set.
z F (ω )
In order to get sensible relationships from the model, DFS parametrize z H (ω )
by using
a monotonic function. In this last case we can invert A and get the exact range of goods
produced by each country, i.e. effectively determine ω and ω̄ as a function of parameters
and w F . Subsequently, we can solve for the equilibrium wage, using the labor market
clearing
L H = ω (w F ) w F L F + ω̄ (w F ) L H .
4.1.1 Where DFS stop and EK start
Eaton and Kortum (2002) (henceforth EK) treat productivities zi (ω ) as an independent
realization of a random variable Zi independently distributed according to the same dis-
38
tribution Fi for each good ω in country i. Given the continuum of goods (using a LLN
argument) we can determine with certainty the fraction of goods produced by each coun-
try. This way EK are able to overcome the complications faced by the standard Ricardian
framework and go much further in developing an analytical quantitative trade frame-
work.
Assume that the random variable Zi follows the Frechet distribution2 :
h i
Pr ( Zi ≤ z) = exp − Ai z−θ .
The parameter Ai > 0 governs country’s overall level of efficiency (absolute advantage)
(with more productive countries having higher Ai ’s). The parameter θ > 1 governs vari-
ation in productivity across different goods (comparative advantage) (higher θ less dis-
persed).
Now we will split the [0, 1] interval by thinking of ω̄ as the probability that the relative
productivity of F to H is less than Ã, where à can either be defined by (4.2). Therefore,
in order to determine ω̄ which is defined as the share of goods that the domestic country
produces we simply compute the probability that the domestic country is the cheapest
provider of the good across all the range of productivities. For example using (4.2) for the
2 See the appendix for the properties of the Frechet distribution and the next chapter for a deriva-
tion from first principles.
39
definition of à we can derive
ω̄ = λ HH
zF
= Pr ≤ Ã
zH
= Pr z F ≤ Ãz H
ˆ +∞ h −θ i
= exp − A F Ãz dFH (z)
0 | {z } | {z }
density of z H (ω )
Pr(z(ω )≤ ÃzH (ω ))
ˆ +∞ h −θ i h i
= exp − A F Ãz θ A H (z)−θ −1 exp − A H (z)−θ dz
0
AH
=
A H + A F Ã−θ
Country H is spending (1 − ω̄ ) w H L H on imports (given Cobb-Douglas) which implies
A F (w F τHF )−θ
XFH = wH LH
A H + A F (w F τHF )−θ
Notice that this relationship is similar to the one derived with the assumption of the
Armington aggregator, equation (3.10), but with an exponent −θ for the elasticity of trade
with respect to trade costs. A lower value of θ generates more heterogeneity. This means
that the comparative advantage exerts a stronger force for trade against resistance impose
by the geographic barrier τin . In other words with low θ there are many outliers that
overcome differences in geographic barriers (and prices overall) so that changes in w’s
and τ’s are not so important for determining trade.
40
4.2 A theory of technology starting from first principles
We start with a very general technological framework under the following assumptions.
Time is continuous and there is a continuum of goods with measure µ (Ω). Ideas for good
ω (ways to produce the same good with different efficiency) arrive at location i at date t
at a Poisson rate with intensity
āRi (ω, t)
where we think of ā as research productivity and R as research effort. The quality of ideas
is a realization from a random variable Q drawn independently from a Pareto distribution
with θ > 1, so that
−θ
Pr [ Q > q] = q/q ,q≥q
¯ ¯
where q is a lower bound of productivities. Note that the probability of an idea being big-
¯
ger than q conditional on ideas being bigger than a threshold, is also Pareto (see appendix
for the properties of the Pareto distribution).
The above assumptions together imply that the arrival rate of an idea of efficiency
Q ≥ q is
−θ
āRi (ω, t) q/q .
¯
(normalize this with q→ 0, ā → +∞ such that āq−θ → 1 in order to consider all the ideas
¯ ¯
in (0, +∞)). We also assume that there is no forgetting of ideas. Thus, we can summarize
the history of ideas for good ω by
ˆ t
Ai (ω, t) = Ri (ω, τ ) dτ.
−∞
The number of ideas with efficiency Q > q0 is therefore distributed Poisson with a param-
−θ
eter Ai (ω, t) (q0 ) (using the previous normalization).
41
The unit cost for a location i of producing good ω with an efficiency of q is c = wi /q.
Given all the above, the expected number of techniques providing unit cost less than c is
distributed Poisson with parameter
Φi (ω, t) cθ
where
Φi (ω, t) = Ai (ω, t) wi−θ .
But notice that this delivers back unit costs that are conditionally Pareto distributed
h w w i
Pr C ≤ c0 |C ≤ c = Pr Q ≥ 0 = q0 | Q ≥
=q
c c
0 −θ θ 0 θ
A (ω, t) (q ) q c
= −θ
= 0
= .
A (ω, t) (q) q c
In what follows set
Φ = Φi (ω, t) .
4.2.1 Order Statistics and Various Moments
The generality of this approach still allow for a number of order statistics and key mo-
ments to be computed. We start by computing the distribution of the order statistics in
the model.
Definition 1. C (k) is the k’th lowest unit cost technology for producing a particular good.
Given this definition we have the main theorem for the joint distribution of the order
statistics C (k)
42
Theorem 5. The joint density C (k) , C (k+1) is
g C (k) = ck , C (k+1) = ck+1 ≡ gk,k+1 (ck , ck+1 )
θ2 −1 θ −1
= Φk+1 cθk
k c k +1 exp − Φc θ
k +1
( k − 1) !
for 0 < ck ≤ ck+1 < ∞ while the marginal density of C (k) is:
θ −1
gk ( c k ) = Φk cθk
k exp − Φc θ
k
( k − 1) !
for 0 < ck < +∞
Proof. We start by looking at costs C ≤ c̄. The distribution of a cost C conditional on C ≤ c̄
is:
c θ
F (c|c̄) = c ≤ c̄
c̄
F (c|c̄) = 1 c > c̄
The probability that a cost is less than ck is F (ck |c̄). Thus, if we have n techniques with
unit cost less than c̄, where ck ≤ ck+1 ≤ c̄, the probability that the k’th lowest cost is less
than ck while the remaining are greater than ck+1 is given by the multinomial:
n
h i
k n−k
Pr C (k) ≤ ck , C (k+1) ≥ ck+1 |n = F (ck |c̄) (1 − F (ck+1 |c̄))
k
Taking the negative of the cross derivative of this expression with respect to ck , ck+1 gives
k −1
n!F (ck |c̄) [1 − F (ck+1 |c̄)]n−k−1 F 0 (ck |c̄) F 0 (ck+1 |c̄)
gk,k+1 (ck , ck+1 |c̄, n) =
( k − 1) ! ( n − k − 1) !
43
for ck+1 ≥ ck and n ≥ k + 1. For n < k + 1 we can define gk,k+1 (ck , ck+1 |c̄, n) = 0. We also
know that n is drawn from a Poisson distribution with parameter Φc̄θ , the expectation of
this joint distribution unconditional on n is:
∞
n
exp −Φc̄θ Φc̄θ
gk,k+1 (ck , ck+1 |c̄) = ∑ gk,k+1 (ck , ck+1 |c̄, n) =
n =0 | n!
{z } | {z }
(k)
prob n ideas arrived for conditional on n prob C =ck ,
a particular good C ( k +1) = c k +1
∞
n
exp −Φc̄θ Φc̄θ n!F (ck |c̄)k−1 [1 − F (ck+1 |c̄)]n−k−1 F 0 (ck |c̄) F 0 (ck+1 |c̄)
= ∑
n = k +1
n! ( k − 1) ! ( n − k − 1) !
k +1
Φc̄θ exp −Φc̄θ F (ck+1 |c̄) F (ck |c̄)k−1 F 0 (ck |c̄) F 0 (ck+1 |c̄)
=
( k − 1) !
∞ m exp −Φc̄θ F (c |c̄) [1 − F (c |c̄)]m
∑ exp −Φc̄θ Φc̄θ
k +1 k+
m =0 m!
k + 1 k −1 0
Φc̄θ exp −Φc̄θ F (ck+1 |c̄) F (ck |c̄) F (ck |c̄) F 0 (ck+1 |c̄)
= 1
( k − 1) !
Substituting using the expression F (c|c̄) we have that
θ2 −1 θ −1
gk,k+1 (ck , ck+1 |c̄) = Φk+1 cθk
k c k +1 exp − Φc θ
k +1
( k − 1) !
Now by letting c̄ → ∞ we can integrate for the entire range of c ≥ ck and derive the
marginal density by making use of the above expression. We have that
ˆ ∞
gk ( c k ) = gk,k+1 (ck , ck+1 )dck+1
ck
ˆ ∞
θ2 −1 −1 −Φck+1
Φk+1 cθk
θ
= k cθk+ 1e dck+1 .
( k − 1) ! ck
44
Now by making the substitution u = cθk+1 ,
ˆ ∞ ˆ ∞
−1 −Φcθk+1
cθk+ 1e dck+1 =θ −1
e−Φu du
ck cθk
= θ −1 Φ−1 e−Φck .
θ
Therefore
θ2 −1
−1 −1 −Φcθk
gk ( c k ) = Φk+1 cθk
k θ Φ e
( k − 1) !
θ −1 −Φcθk
= Φk cθk
k e , (4.3)
( k − 1) !
as asserted.
This result will be the base for a series of lemmas to be discussed later on. First, by
h i
noticing that Fk0 (ck ) = gk (ck ) we can directly compute the probability Pr C (k) ≤ c̃k :
Lemma 1. The distribution of the k0 th lowest cost C (k) is:
υ
h i k −1 Φc̃θk
∑ e−Φc̃υ
(k) θ
Pr C ≤ c̃k = Fk (c̃k ) = 1 − (4.4)
υ =0 υ!
This Lemma implies that the distribution of the lowest cost (k = 1) is the Frechet dis-
tribution
F1 (c̃1 ) = 1 − exp −Φc̃1θ
Now in this context we will assume that ideas are randomly assigned to goods across
the continuum. Given that there is a large number of goods (say of measure µ (Ω)) in the
continuum we can drop the ω notation by simply denoting Ai (ω, t) = Ai (t) /µ (Ω) to be
the average number of ideas available for a good, in location i at time t. Given the above,
the measure of goods with cost less than c is Φi (t) /µ (Ω) cθ and the distribution of the
45
lowest cost C (1) (the frontier idea) is
F1 (c1 ) = 1 − exp − (Φi (t) /µ (Ω)) c̃1θ
Thus a set of µ (Ω) F1 (c1 ) ideas can be produced at a cost less than c1 . We will proceed
under this convention in the rest of this chapter.
Using the following proposition and the assumption of the CES demand we can di-
rectly derive the price index
Proposition 2. For each order k, the b’th moment (b > −θk ) is
b b Γ [(θk + b) /θ ]
E C (k)
= Φ−1/θ ,
( k − 1) !
´ +∞
where Γ (α) = 0 yα−1 e−y dy.
Proof. First consider k = 1, where suppressing notation we denote by the marginal den-
sity of C (k) ,
θ k θk−1
h i
gk ( c ) = Φ ck exp −Φck
θ
( k − 1) !
b ˆ +∞
E C (1)
= cb g1 (c) dc
0
ˆ +∞ h i
= Φθcθ +b−1 exp −Φcθ dc
0
changing the variable of integration to υ = Φcθ and applying the definition of the gamma
46
function, we get
b ˆ +∞
E C (1)
= (υ/Φ)b/θ exp [−υ] dυ
0
θ+b
= (Φ)−b/θ Γ (4.5)
θ
well defined for θ + b > 0. For general k we have
b ˆ +∞
(k)
E C = cb gk (c) dc
0
ˆ +∞
θ h i
= cb Φk cθk−1 exp −Φcθ dc
0 ( k − 1) !
ˆ +∞
Φ k − 1 h i
= cb+θk−θ θΦcθ −1 exp −Φcθ dc
( k − 1) ! 0
Φ k −1
b + θ ( k −1)
(1)
= E C (4.6)
( k − 1) !
Using the general technology framework we developed above and different assump-
tions on the competition structure we will be able to derive main quantitative models that
are widely used in the recent international trade literature. We now provide a number of
illustrations of this general framework that arise from explicilty specifying the competi-
tion and market structure.
4.3 Application I: Perfect competition (Eaton-Kortum)
A main factor inhibiting the use of “classical trade theory” in empirical gravity models
was the perceived intractability of such a problem. The traditional Ricardian comparative
advantage framework relied on two countries and two goods; many attempts to gener-
47
alize the framework quickly led to a nightmare of corner solutions. While Dornbusch,
Fischer, and Samuelson (1977) provided a tractable framework for a continuum of va-
rieties with two countries, it was thought to be impossible to extend the framework to
many countries and arbitrary trade costs (both which are necessary to deliver a gravity-
like equation). (The closest generalization to many countries was the local comparative
static analysis of Wilson (1980)).
This is where Eaton and Kortum (2002) enters. Using a model that bears a resem-
blance to a discrete-choice framework (a la McFADDEN (1974)), they show how to derive
gravity expressions for trade flows in a world with many countries, arbitrary trade costs
(i.e. arbitrary geography), where trade is only driven by technological differences across
countries (i.e. comparative advantage). The Eaton and Kortum (2002) framework not only
shattered the age-old belief that there couldn’t be a Ricardian gravity model, the model
developed turns out to be remarkably elegant. It is also surprising that despite looking
very different from the models we have considered thus far, the Eaton and Kortum (2002)
trade expression remains formally isomorphic to those models.
4.3.1 Model Setup
Let us now turn to the set-up of the model.
The World
In this model, there a finite number of countries i ∈ S ≡ {1, ..., N }; (unlike previous mod-
els, there are technical difficulties in extending the model to a continuum of countries).
There are a continuum of goods Ω. However, unlike in the Krugman (1980) and Melitz
(2003) models which follow, every country is able to produce every good. Countries, how-
ever, vary (exogenously) in their productivity of each good; in particular, let zi (ω ) denote
48
country i’s efficiency at producing good ω ∈ Ω.
The Eaton and Kortum (2002) has no concept of a firm. Instead, it is assumed that all
goods ω ∈ Ω are produced using the same bundle of inputs with a constant returns to
scale technology. Let the cost of a bundle of inputs in country i ∈ S be ci so that the cost
ci
of producing one unit of ω ∈ Ω in country i ∈ S is zi ( ω )
.
Finally, like the previous models we considered, suppose there is an iceberg trade cost
τij ≥ 1 of trading a good from i ∈ S to j ∈ S.
Supply
Each good is assumed to be sold in perfectly competitive markets, so that the price a
consumer in country j ∈ S would pay if she were to purchase good ω ∈ Ω from country
i ∈ S is:
ci
pij (ω ) = τij . (4.7)
zi ( ω )
However, consumers in country j ∈ S are assumed to only purchase good ω ∈ Ω from the
country who can provide it at the lowest price, so the price a consumer in j ∈ N actually
pays for good ω ∈ Ω is:
ci
p j (ω ) ≡ min pij (ω ) = min τij . (4.8)
i ∈S i ∈S zi ( ω )
The basic idea behind the Eaton and Kortum (2002) is already present in equation (4.8): a
country j ∈ S will be more likely to purchase good ω ∈ Ω from country i ∈ S if (1) it has
a lower unit cost ci ; (2) it has a higher good productivity zi (ω ); and/or (3) it has a lower
trade cost τij .
One of the major innovations of the Eaton and Kortum (2002) model is that the pro-
49
ductivity zi (ω ) is treated as a random variable drawn independently and identically for
each ω ∈ Ω. Define Fi to be the cumulative distribution function of the productivity in
country i ∈ S. That is, for each i ∈ S, for all ω ∈ Ω:
Fi (z) ≡ Pr {zi (ω ) ≤ z}
Eaton and Kortum (2002) assume that Fi (z) is Fréchet distributed so that for all z ≥ 0:
n o
Fi (z) = exp − Ti z−θ , (4.9)
where Ti > 0 is a measure of the aggregate productivity of country i (note that a larger
value of Ti decreases Fi (z) for any z ≥ 0, i.e. it increases the probability of larger values
of z and θ > 1 (which is assumed to be constant across countries) governs the distribu-
tion of productivities across goods within countries (as θ increases, the heterogeneity of
productivity across goods declines).
Why make this particular distributional assumption for productivities? In a micro-
foundation that is related to the general technological framework we have discussed
above Kortum (1997) showed that if the technology of producing goods is determined by
the best “idea” of how to produce, then the limiting distribution is indeed Fréchet, where
Ti reflects the country’s stock of ideas. More generally, consider the random variable:
Mn = max { X1 , ...., Xn } ,
where Xi are i.i.d. The Fisher–Tippett–Gnedenko theorem states that the only (normal-
ized) distribution of Mn as n → ∞ is an extreme value distribution, of which Fréchet
is one of three types (Type II). Note that a conditional logit model assumes that the er-
50
ror term is Gumbel (Type I) extreme value distributed. If random variable x is Gumbel
distributed, ln x is distributed Fréchet; hence, loosely speaking, the Fréchet distribution
works better for models that are log linear (like the gravity equation), whereas the Gumbel
distribution works better for models that are linear.
Demand
As in previous models, consumers have CES preferences so that the representative agent
in country j has utility:
ˆ σ−σ 1
σ −1
Uj = q j (ω ) σ dω ,
Ω
where q j (ω ) is the quantity that country j consumes of good ω. Note that unlike the
Krugman (1980) model, not every good produced in every country will be sold to country
j. Indeed, good ω ∈ Ω will be produced by all countries but county j will only purchase
it from one country. However, like the previous models considered, the CES preferences
will yield a Dixit-Stiglitz price index:
ˆ 1−1 σ
1− σ
Pj ≡ p j (ω ) dω (4.10)
Ω
4.3.2 Equilibrium
We now consider the equilibrium of the model. Instead of relying on the CES demand
equation as in the previous models, we use a probabilistic formulation in order to solve
the model.
Prices
In perfect competition only the lowest cost producer of a good will supply that particular
good. Thus, we want to derive the distribution of the minimum price over a set of prices
51
offered by producers in different countries
p j = min p1j , ..., p Nj
In order to find this distribution we take advantage of the properties of extreme value
distributions for productivity z. Everything turns out to work beautifully!
First, let us consider the probability that country i ∈ S is able to offer country j ∈ S
good ω ∈ Ω for a price less than p. Because the technology is i.i.d across goods, this
probability will be the same for all goods ω ∈ Ω. Define:
Gij ( p) ≡ Pr pij (ω ) ≤ p
Using the perfect competition price equation (4.7) and the functional form of the Fréchet
distribution (4.9), we have:
Gij ( p) ≡ Pr pij (ω ) ≤ p ⇐⇒
ci
Gij ( p) = Pr τij ≤ p ⇐⇒
z (ω )
i
ci
Gij ( p) = Pr τij ≤ zi (ω ) ⇐⇒
p
ci
Gij ( p) = 1 − Pr zi (ω ) ≤ τij ⇐⇒
p
ci
Gij ( p) = 1 − Fi τij ⇐⇒
p
( −θ )
ci
Gij ( p) = 1 − exp − Ti τij (4.11)
p
Consider now the probability that country j ∈ S pays a price less than p for good ω ∈ Ω.
Again, because the technology is i.i.d across goods, this probability will be the same for
52
all goods ω ∈ Ω. Define:
Gj ( p) ≡ Pr p j (ω ) ≤ p
Because country j ∈ S buys from the least cost provider, using equation (4.8) and some
basic tools of probability, we can write:
Gj ( p) = Pr min pij (ω ) ≤ p ⇐⇒
i ∈S
= 1 − Pr min pij (ω ) ≥ p ⇐⇒
i ∈S
= 1 − Pr ∩i∈S pij (ω ) ≥ p ⇐⇒
= 1 − ∏ 1 − Gij ( p)
(4.12)
i ∈S
Substituting equation (4.11) into equation (4.12) yields:
Gj ( p) = 1 − ∏ 1 − Gij ( p)
⇐⇒
i ∈S
( −θ )
ci
= 1 − ∏ exp − Ti τij ⇐⇒
i ∈S
p
( )
−θ
= 1 − exp − p θ
∑ Ti ci τij ⇐⇒
i ∈S
n o
= 1 − exp − pθ Φ j , (4.13)
−θ
where Φ j ≡ ∑i∈S Ti ci τij . Equation (4.13) tells us what the distribution of prices will
be across goods for country j. This, in turn, will allow us to calculate the price index in
53
country j, Pj . Starting with the definition of the price index from equation (4.11), we have:
ˆ 1−1 σ
1− σ
Pj ≡ p j (ω ) dω ⇐⇒
Ω
ˆ ∞
Pj1−σ = p1−σ dGj ( p) ⇐⇒
ˆ0 ∞ o
d n
Pj1−σ = p 1− σ
1 − exp − p Φ j
θ
dp ⇐⇒
0 dp
ˆ ∞ n o
Pj1−σ = θΦ j pθ −σ exp − pθ Φ j dp.
0
Define x ≡ pθ Φ j so that with a change of variables we have:
ˆ ∞ 1−θ σ
x
Pj1−σ = exp {− x } dx ⇐⇒
0 Φj
ˆ ∞
− 1−θ σ 1− σ
Pj1−σ = Φj x θ exp {− x } dx ⇐⇒
0
θ+1−σ
− 1−θ σ
Pj1−σ = Φj Γ ⇐⇒
θ
1
θ + 1 − σ 1− σ
− 1θ
Pj = Φj Γ ,
θ
´∞
where Γ (t) ≡ 0 x t−1 e− x dx is the Gamma function.
Hence, the equilibrium price index in country j ∈ N can be written as:
!− 1θ
−θ
Pj = C ∑ Ti ci τij , (4.14)
i ∈S
θ +1− σ
1−1 σ
where C ≡ Γ θ . [Class questions: What does this mean if trade is costless? When
trade is infinitely costly?]
54
4.3.3 Gravity
Now suppose we are interested in determining the probability that i ∈ S is the least cost
provider of good ω ∈ Ω to destination j ∈ S. Because all goods receive i.i.d. draws and
there are a continuum of varieties, by the law of large numbers, this probability will be
equal to the fraction of goods i sells to j. Define:
πij ≡ Pr pij (ω ) ≤ min pkj (ω ) ⇐⇒
k∈S\ j
ˆ ∞
= Pr min pkj (ω ) ≥ p dGij ( p) ⇐⇒
0 k∈S\ j
ˆ ∞ n
o
= Pr ∩k∈S\ j pkj (ω ) ≥ p dGij ( p) ⇐⇒
ˆ0 ∞
∏ 1 − Gkj ( p) dGij ( p)
= (4.15)
0 k ∈ S \i
Substituting the distribution of price offers from equation (4.11) into equation (4.15) yields:
ˆ ∞
∏
πij = 1 − Gkj ( p) dGij ( p) ⇐⇒
0 k∈S\ j
ˆ ∞
( −θ )!
( −θ )!!
ck d ci
= ∏ exp −Tk p τkj
0 k∈S\ j dp
1 − exp − Ti
p
τij dp ⇐⇒
ˆ
− θ ∞ θ −1 n o
= Ti ci τij θp exp − pθ Φ j dp ⇐⇒
0
−θ
Ti ci τij n o
= − exp − pθ Φ j |0∞
Φj
−θ
Ti ci τij
= (4.16)
Φj
Hence, the fraction of goods exported from i to j just depends on i0 s share in j0 s Φ j . Note
that more productive countries, countries with lower unit costs, and countries with lower
bilateral trade costs (all relative to other countries) will comprise a larger fraction of the
55
goods sold to j.
Note that πij is the fraction of goods that j ∈ S purchases from i ∈ S; it may not be
the fraction of j0 s income that is spent on goods from country i. However, it turns out
that with the Fréchet distribution, the distribution of prices of goods that country j actually
purchases from any country i ∈ S will be the same. To see this, note that the probability
country i ∈ S is able to offer good ω ∈ Ω for a price lower than p̃ conditional on i having
the lowest price is simply the product of inverse of the probability that i has the lowest
cost good and the probability that j receives a price offer lower that p̃:
ˆ p̃
1
Pr pij (ω ) ≤ p̃| pij (ω ) ≤ min pkj (ω ) = Pr min pkj (ω ) ≥ p dGij ( p) ⇐⇒
k ∈ S \i πij 0 k ∈ S \i
ˆ p̃
1
∏
= 1 − Gkj ( p) dGij ( p) ⇐⇒
πij 0 k ∈ S \i
−θ
1 Ti ci τij n o
p̃
= − exp − pθ Φ j |0
πij Φj
−θ
1 Ti ci τij n o
= 1 − exp − p̃θ Φ j
πij Φj
= Gj ( p̃) .
Intuitively, what is happening is that origins with better comparative advantage (lower
trade costs, better productivity, etc.) in selling to j will exploit its advantage by selling
a greater number of goods to j exactly up to the point where the distribution of prices it
offers to j is the same as j’s overall price distribution.
While this result depends heavily on the Fréchet distribution, it greatly simplifies the
process of determining trade flows. Since the distribution of prices offered to an importing
country j ∈ S is independent of the origin, country j0 s average expenditure per good does
not depend on the source of the good. As a result, the fraction of goods purchased from a
56
particular origin (πij ) is equal to the fraction of j0 s income spent on goods from country i,
Xij
λij ≡ Yj i.e.
−θ
Ti ci τij
λij = . (4.17)
Φj
This implies that the total expenditure of j on goods from country i is:
Xij = πij Ej ,
where from equation (4.16) we have:
−θ
Ti ci τij
Xij = Ej (4.18)
Φj
Supposing that ci = wi and substituting in equation (4.14) for the price index yields:
Xij = C −θ τij−θ wi−θ Ti Ej Pjθ . (4.19)
Hence, the Eaton and Kortum (2002) model yields a nearly identical gravity equation to
the Armington model of Anderson (1979), except that the relevant elasticity is θ instead
of σ − 1.
As in previous models, we can also push the gravity equation a little bit further. Note
that in general equilibrium, the total income of a country will equal the amount it sells to
all other countries:
Yi = ∑ Xij (4.20)
j∈S
57
Substituting gravity equation (4.19) into equation (4.20) yields:
Yi = ∑ C−θ τij−θ wi−θ Ti Ej Pjθ ⇐⇒
j∈S
Yi = C −θ wi−θ Ti ∑ τij−θ Ej Pjθ ⇐⇒
j∈S
Yi
C −θ wi−θ Ti = (4.21)
∑ j∈S τij−θ Ej Pjθ
Now substituting equation (4.21) back into the gravity equation (4.19) yields:
Xij = C −θ τij−θ wi−θ Ti Ej Pjθ ⇐⇒
Yi Ej
Xij = τij−θ × −θ
× −θ , (4.22)
Πi Pj
where
!− 1θ
Ek
Πi ≡ ∑ τik−θ
Pk−θ
.
k∈S
We will see soon that if trade costs are symmetric, Pi = Πi .
Why is the trade elasticity different in this model? Recall that in the Armington and
Krugman (1980) models, how responsive trade flows were to trade costs depended on
how demand for a good was affected by the good’s price, which was determined by con-
sumer’s elasticity of substitution. In this model, however, changes in trade costs affect the
extensive margin, i.e. which goods an origin country trades with a destination country. As
the bilateral trade costs rise, the origin country is the least cost provider in fewer goods;
the greater the θ, the less heterogeneity in a country’s productivity across different goods,
so there are a greater number of goods for which it is no longer the least cost provider.
Hence, the Eaton and Kortum (2002) model is similar to the Melitz (2003) model in that
the elasticity of trade to trade costs ultimately depends on the density of producers/firms
58
that are indifferent between exporting and not exporting: the greater the heterogeneity in
productivity, the lower the density of these marginal producers.
4.3.4 Welfare
From the CES preferences, the welfare of a worker in country i ∈ S can be written as:
wi
Wi ≡ . (4.23)
Pi
Xij
Recall from above that λij ≡ Ej is the fraction of j0 s expenditure spent on i. From gravity
equation (4.19) we then have that:
λij = C −θ τij−θ wi−θ Ti Pjθ ,
which, given τii = 1, implies:
λii = C −θ Wi−θ Ti ⇐⇒
−1 1
Wi = C −1 λii θ Tiθ . (4.24)
Hence, as in the Krugman (1980) model, welfare can be expressed as a function of tech-
nology and the openness of a country. Indeed, as far as I am aware, Eaton and Kortum
(2002) were the first to derive this expression, although the expression is usually known as
the “ACR” equation after Arkolakis, Costinot, and Rodrı́guez-Clare (2012), who derived
the conditions under which it holds more generally (we will see more of this in several
weeks).
59
4.3.5 Discussion of EK
The Eaton and Kortum model is a key defining point in the new trade literature not only
because it has provided a simple framework to model comparative advantage but also
because it has provided a set of probabilistic tools to think about the determination of
trade and allocation of resources without having to worry about mathematical intricacies
(e.g. corner solutions in the standard Ricardian setup etc.). These tools have been use
henceforth for a variety of applications that we will discuss in many of the remaining
chapters. Still, the structure we developed above can be useful with different forms of
competition as is illustrated by Bernard, Eaton, Jensen, and Kortum (2003) and Melitz
(2003) that model other forms of competition (Bertrand and monopolistic, respectively)
and also work that brings models of trade closer to the trade data in many dimensions.
4.4 Application II: Bertrand competition (Bernand, Eaton, Jensen,
Kortum)
• Consider the case where different producers have access to different technologies. If
we assume Bertrand competition, the cost distribution will be given by the frontier
producer (k = 1 in the expression 4.4) but prices are related to the distribution of
the second lowest cost (k = 2). Since the lowest cost supplier is the one that will sell
the good, the probability that a good is supplied from i to j is the same as in perfect
competition, equation (4.17).
60
4.4.1 Supply side
The price of a good ω in market j is:
p j (ω ) = min C2j (ω ) , m̄C1j (ω )
where we will define Cij (ω ) to be the cost of the i’th minimum cost producer of good ω in
country j, and m̄ = σ/ (σ − 1) is the optimal markup that a monopolist firm would charge
(assuming CES preferences with a demand elasticity σ). Given heterogeneity among tech-
nological costs for firms we will derive the distribution of costs and markups in each given
country.
Efficiency, markups, and measured productivity
Define again C (k) as the k’th lowest unit cost technology for producing a particular
good. We have the following Lemma
Lemma 2. The distribution of C (k+1) conditional on C (k) = ck is:
h i h i
Pr C (k+1) ≤ ck+1 |C (k) = ck = 1 − exp −Φ cθk+1 − cθk , cθk+1 ≥ cθk ≥ 0 (4.25)
Proof. Using Bayes’ rule we have
ˆ c k +1
h
( k +1) (k)
i gk,k+1 (ck , c)
Pr C ≤ c k +1 | C = ck = dc
ck gk ( c k )
ˆ c k +1 h i
= θΦcθ −1 exp −Φcθ + Φcθk dc
ck
h i
= 1 − exp −Φ cθk+1 − cθk .
61
The above relationship also implies that (define m0 = ck+1 /ck )
" # " #
C ( k +1) c k +1 ( k ) C ( k +1) 0 (k)
Pr ≤ |C = ck = Pr ≤ m |C = ck
ck ck C (k)
h i
= 1 − exp −Φcθk m0 − 1
θ
The distribution of the ratio M0 = C (2) /C (1) given C (1) = c1 is:
h i h i
Pr M0 ≤ m0 |C (1) = c1 = 1 − exp −Φc1θ m0 − 1
θ
.
We have that the lower c1 , the more likely a high markup. Thus, in this context low-
cost producers are more likely to charge a high markup and their measured (revenue)
productivity is more likely to appear as higher. In this model, revenue productivity is
associated one-to-one with the markup that the firm charges, since it equals
M (ω ) wi τij q (ω ) /z (ω ) / l (ω ) = M (ω ) wi τij q (ω ) /z (ω ) / [q (ω ) /z (ω )] = M (ω ) wi τij .
| {z } | {z }
revenue labor used
Notice that from this expression is straightforward that market structures/demand func-
tions which imply costant markup imply no revenue productivity variation across firms.
Notice that the markup with Bertrand competition that BEJK consider is
( )
C (2) ( ω )
M (ω ) = min , m̄ (4.26)
C (1) ( ω )
We start by characterizing the distribution of the ratio M0 = C (2) /C (1) . Conditional on the
62
(2)
second lowest cost in market j being Ci = c2 , we have
h i h i
(2) (1) (2)
Pr M0 ≤ m0 |Cj = c2 = Pr c2 /m0 ≤ Cj ≤ c2 |Cj = c2
´ c2
0 g j ( c1 , c2 ) dc1
= c´2 /m
c2
0 g j ( c1 , c2 ) dc1
c2θ − (c2 /m0 )
θ
=
c2θ
−θ
= 1 − m0 . (4.27)
This derivation implies that the distribution of this M0 does not depend on c2 and is also
Pareto.3 Thus, the unconditional distribution is also Pareto.4 Given the markup func-
tion for the case of Betrand competition, equation (4.26), we have proved the following
proposition:
3 Using again the results of the theorem we can derive the distribution of m for each k. Notice that
h i ˆ ck
Pr C (k) ≤ ck |C (k+1) = ck+1 = gk,k+1 (c, ck+1 ) /gk+1 (ck+1 ) dc
0
ˆ ck θ 2 Φk+1 cθk−1 cθ −1 exp −Φcθ
( k −1) ! k + 1 k + 1
= θ (k+1)−1 −Φcθ
dc
0 θ
(k)!
Φ k +1 c k +1 e k +1
ck θk
= (4.28)
c k +1
and simply replacing for ck = ckm+1 in expression (4.28) (and given that C (k+1) = ck+1 ) we can get
" #
C ( k +1) h c i
Pr ( k )
≤ m|C ( k +1)
= ck+1 = 1 − Pr C (k) ≤ k+1 |C (k+1) = ck+1
C m
= 1 − m−θk
4 An alternative derivation of the distribution of the markups can be obtained for m < m̄. To compute the
unconditional distribution of productivities for m < m̄ we have that
ˆ +∞ h i
Pr [ M ≤ m] = 1 − exp −Φc1θ mθ − 1 Φθc1θ −1 exp −Φc1θ dc1
0
+∞ +∞
= − exp −Φc1θ + exp −Φc1θ = 1 − 1/mθ
0 0
63
Proposition 3. Under Bertrand competition the distribution of the markup M is:
1 − m−θ if m < m̄
Pr [ M ≤ m] = FM (m) = .
1 if m ≥ m̄
With probability m̄−θ the markup is m̄. The distribution of the markup is independent of C (2) .
4.4.2 Gravity
Lengthy derivations can be used (see the online appendix of BEJK) to show that the joint
distribution of the lowest and the second lowest cost of supplying a country, conditional
on a certain country being the supplier, is independent of the country of origin, and is
given by equation (4.28) for k = 1. Therefore, the market share of country i in j equals to
the probability that country i is the supplier and thus
−θ
Ai wi τij Φij
λij = −θ = (4.29)
∑kN=1 Ak wk τkj Φj
It is worth noting that the profits of firms depend on both their cost but also their competi-
tor’s cost. An interesting implication of the model is that the share of profit to aggregate
revenue is constant and equals to 1/ (θ + 1) . The proof of this can be found in the online
appendix of BEJK.
64
4.4.3 Welfare
Using the derivations for the distribution of markups, we can also derive a price index for
this case. We have
ˆ h ∞ i
Pj1−σ = E p1j −σ | M = m θm−θ −1 dm
1
ˆ m̄ ˆ +∞ 1− σ
(2) 1− σ (2)
= E Cj θm−(θ +1) dm + E m̄Cj /m θm−(θ +1) dm
|1 {z } | m̄ {z }
marginal cost pricing Dixit-Stiglitz pricing
1− σ
(2)
θ
=E Cj 1 − m̄−θ + m̄−θ
1+θ−σ
where in the second equality we used the fact that the distribution of markups is indepen-
(2) 1− σ
dent of the second lowest cost, equation (4.27). We have already calculated E Ci
in equation (4.6). Thus, the price index under Bertrand competition is
Pi = γ BC Φi−1/θ
1/(1−σ)
2θ + 1 − σ
−θ
−θ θ 1/(1−σ )
γ BC
= 1 − m̄ + m̄ Γ
1+θ−σ θ
and given the gravity expression the welfare as a function of trade is the same as in the
case of Perfect competition
4.4.4 Discussion of BEJK
• Develop a firm level model and explicitly tests its predictions with firm-level data.
• Model a framework where firms markups are variable and depending on compe-
tition. Alternative models of variable markups can be developed in monopolistic
competition by allowing for a preference structure that departs from the CES aggre-
65
gator (see section 8.4).
• Develop a methodology of simulating an artificial economy with heterogeneous
firms and finding the parameters of this economy that brings the predictions of the
model closer to the data.
• Acknowledges the fact that “measured productivity” when measured as nominal
output over employment is constant in models with constant markups. It developes
a model that can deliver variable markups, thus measured productivity differentials.
4.5 Application III: Monopolistic competition with CES (Chaney-
Melitz)
The Krugman (1980) model, while micro-founding the gravity equation based on a story
of equilibrium firm entry, made the simplifying assumption that all firms were ex-ante
identical. With the advent of digitized data on firm-level trading partners, however, it
became clear that there existed an enormous heterogeneity in firm’s exporting behavior
Bernard, Jensen, Redding, and Schott (2007) provides an excellent overview of the em-
pirical patterns concerning firms in international trade, of which we mention just a few.
First, the vast majority of firms do not export; in the U.S. in 2000, only 4% of firms were
exporters. Second, amongst those 4% of firms that did export, 96% of the value of exports
came from just 10% of exporters. Third, comparing the firms that export to those that
do not, the exporting firms tend to be larger, more productive, more skill- and capital-
intensive, and to pay higher wages. These differences are apparent even before exporting
begins, suggesting that more productive firms choose to export (rather than the act of
exporting increasing the productivity of firms).
In response to these new empirical findings, Melitz (2003) developed an extension
66
of the Krugman (1980) where firms varied (exogenously) in their productivities and self
selected into exporting. This model has proven enormously successful for a number of
reasons: first, it is able to capture many (but not all) of the empirical facts mentioned
above, most notably that larger firms will be more likely to export; second, the model
has proven incredibly flexible, generating a huge number of “extensions” to capture ad-
ditional empirical patterns; and third, the model generates a new (potential) source for
gains from trade: if falling trade costs leads higher productivity firms to grow and lower
productivity firms to shrink, this reallocation of factors of production will increase the av-
erage productivity of a country. While there is some debate about whether this is actually
an additional gain from trade (as we will see in a few weeks), the idea that greater trade
can make a country more productive by increasing competition has made the (rare) leap
from academic to political discourse; for example the U.S. trade representative web page
lists as one of its major “benefits of trade” the fact that “trade expansion benefits families
and businesses by supporting more productive, higher paying jobs in our export sectors.”
4.5.1 Model Set-up
Let us now turn to the set-up of the model.
The world
As in the previous models, there is a compact set S of countries, where I will keep the
notation that i is an origin country and a j is a destination country. Each country i ∈ S
is be populated by an exogenous measure Li of workers/consumers where each worker
supplies her unit of labor inelastically. Suppose that labor is the only factor of production.
67
Supply
As in the Krugman (1980) model, suppose that there is a continuum Ω of possible varieties
that the world can produce, and suppose that every firm in the world produces a distinct
variety ω ∈ Ω. Let the set of varieties produced by firms located in country i ∈ S be
denoted by Ωi ⊂ Ω. (Note that Ωi is an equilibrium object, as it will depend on the
number of firms that are actively producing).
Instead of the fixed entry cost in Krugman (1980) model, suppose that there is a mass
Mi of firms from country i ∈ S and that firms must incur a fixed cost f ij > 0 to export to
each destination j ∈ S.5
The major innovation of the Melitz (2003) model is that firms are heterogeneous. To
model this, we suppose that each firm in i ∈ S has a productivity ϕ drawn from some
cumulative distribution function Gi ( ϕ), i.e. it costs a firm with productivity ϕ exactly
1
ϕ units of labor to produce a single unit of its differentiated variety. In what follows,
we will sometimes identify each firm by its productivity (since all firms with the same
productivity within a particular country will act the same way) and sometimes identify
each firm by its variety ω (since every firm produces a unique variety).
Finally, as in previous models, we suppose that all firms within a country are subject
to iceberg trade costs {τij }i,j∈S .
Demand As in the Krugman (1980) model, we assume that consumers have CES pref-
erences over varieties. Hence a representative consumer in country j ∈ S gets utility Uj
from the consumption of goods shipped by all other firms in all other countries, where:
Melitz (2003), it was assumed that there was an additional entry cost f ie that determined the equilib-
5 In
rium mass of firms Mi . In the Chaney (2008) version of the model, Mi was assumed (for simplicity) to be
proportional to the income in the origin. The Chaney (2008) version of the model has become more widely
used because it allows for arbitrary bilateral trade costs (the original Melitz (2003) model imposed symmetry).
68
ˆ ! σ−σ 1
σ−σ 1
Uj = ∑ Ωij
qij (ω ) dω , (4.30)
i ∈S
where qij (ω ) is the quantity consumed in country j of variety ω.
4.5.2 Equilibrium
We now consider the equilibrium of the model.
Optimal demand
The consumer’s utility maximization problem is identical to that of Krugman (1980): A
consumer in country j ∈ S optimal quantity demanded of good ω ∈ Ω is:
qij (ω ) = pij (ω )−σ Yj Pjσ−1 , (4.31)
where:
ˆ ! 1−1 σ
Pj ≡ ∑ Ωi
pij (ω ) 1− σ
dω (4.32)
i ∈S
is the Dixit-Stiglitz price index.
The amount spent on variety ω is simply the product of the quantity and the price:
xij (ω ) = pij (ω )1−σ Yj Pjσ−1 . (4.33)
To determine total trade flows, we need to aggregate across all firms in country i:
ˆ ˆ
Xij ≡ xij (ω ) dω = Yj Pjσ−1 pij (ω )1−σ dω. (4.34)
Ωi Ωi
Unlike the Krugman (1980) model, firms with different productivities will charge different
69
prices, so the integral in equation (4.34) becomes more complicated.
Optimal supply
We now determine the equilibrium prices that a firm with productivity ϕ sets (where we
now identify firms by their productivity). The optimization problem is:
wi
max ∑ pij ( ϕ) qij ( ϕ) − τij qij ( ϕ) − f ij s.t. qij ( ϕ) = pij ( ϕ)−σ Yj Pjσ−1 .
{qij ( ϕ)} j∈S j∈S ϕ
Substituting the constraint into the maximand yields:
w
max ∑ pij ( ϕ) 1− σ
Yj Pjσ−1 − i τij pij ( ϕ)−σ Yj Pjσ−1 − f ij
{qij ( ϕ)} j∈S j∈S ϕ
The first order condition implies that a firm from i ∈ S with productivity ϕ, conditional
on selling to destination j, will charge a price:
σ wi
pij ( ϕ) = τij (4.35)
σ−1 ϕ
Combining the optimal price (equation (4.35)) and the optimal demand (equation (4.31))
gives the total revenue of a firm (conditional on exporting) to be:
1− σ
σ wi
xij ( ϕ) ≡ pij ( ϕ) qij ( ϕ) = τij Yj Pjσ−1 (4.36)
σ−1 ϕ
70
and variable profits conditional on entering (note that we now define πij ( ϕ) as profits
without the fixed costs):
wi
πij ( ϕ) ≡ pij ( ϕ) −
τij qij ( ϕ)
ϕ
−σ
σ wi wi σ wi
= τij − τij τij Yj Pjσ−1
σ−1 ϕ ϕ σ−1 ϕ
1− σ 1− σ
1 σ wi
= τij Yj Pjσ−1
σ σ−1 ϕ
1
= xij ( ϕ) (4.37)
σ
Note that both revenue and profits are increasing in a firm’s productivity. [Class question:
Why is this?]
Aggregation
We now discuss how to use the optimal behavior on the part of each firm to construct the
aggregate variables necessary to generate a gravity equation. Let µij ( ϕ) be the (equilib-
rium) probability density function of the productivities of firms from country i that sell to
country j and let Mij be the (equilibrium) measure of firms exporting from i to j.
Then we can write the average prices charged by all firms in i ∈ S selling to j ∈ S as:
ˆ ˆ ∞ 1− σ
1− σ σ wi
pij (ω ) dω = Mij τij µij ( ϕ) dϕ ⇐⇒
Ωi 0 σ−1 ϕ
1− σ ˆ ∞
σ
= wi τij Mij ϕσ−1 µij ( ϕ) dϕ ⇐⇒
σ−1 0
1− σ
σ σ −1
= Mij wi τij ϕ̃ij ,
σ−1
´∞ σ−1 1
where ϕ̃ij ≡ 0 ϕσ−1 µij ( ϕ) dϕ captures the “average” productivity of producers
71
from i selling to j. This allows us to write the gravity equation (4.34) as:
1− σ
σ σ −1
Xij = τij1−σ wi1−σ Mij ϕ̃ij Yj Pjσ−1 . (4.38)
σ−1
Equation (4.38) resembles the gravity equation from Krugman (1980), except that we now
have to keep track of both the number of firms selling to j (Mij ) and their average pro-
ductivity ϕ̃ij . Note that as the average productivity of entrants increases, the trade flows
increase. [Class question: what is the intuition for this?].
Selection into exporting
In order to determine the equilibrium number of entrants Mij and the average productiv-
ity of entrants ϕ̃ij , we have to consider the export decisions of firms. A firm from country
i ∈ S with productivity ϕ conditional on producing will export to j if and only if:
πij ( ϕ) ≥ f ij
From equations (4.36) and (4.37) we can write this as:
1− σ
1 σ wi
τij Yj Pjσ−1 ≥ f ij ⇐⇒
σ σ−1 ϕ
1
σ − 1 ! σ −1
σ
σ f ij σ−1 wi τij
ϕ ≥ ϕij∗ ≡ . (4.39)
Yj Pjσ−1
Hence, only firms that are sufficiently productive will find it profitable to incur the fixed
cost of exporting to destination j. This means that the model matches the empirical fact
that larger and more productive firms select into exporting.
Together, equations (4.39) and (4.47) allow us to determine the “average” productivity
of producers selling from i to j:
72
σ−1 1
ˆ
∞
1
ϕ̃ij = ϕσ−1 dGi ( ϕ) . (4.40)
1 − Gi ϕij∗ ϕij∗
and the density of firms selling from i to j:
Mij = 1 − Gi ϕij∗ Mi , (4.41)
so that the gravity equation (4.38) becomes:
1− σ ˆ ∞
!
σ
Xij = τij1−σ wi1−σ Mi ϕ σ −1
dGi ( ϕ) Yj Pjσ−1 . (4.42)
σ−1 ϕij∗
4.5.3 The Pareto Distribution
In this section, we show that when the distribution of firm productivities is a Pareto dis-
tribution, the model above simplifies nicely. This insight is due to Chaney (2008). The
assumption that the distribution of firm’s productivities is Pareto can actually be micro-
founded as follows: Let µ (Ω) ⊂ (0, +∞) is the set of available varieties. Let Ii the mea-
sure of ideas that fall randomly into goods. In some sense Ii /µ (Ω) ideas correspond to
each good. In the probabilistic context we described above, the monopolistic competition
model arises in a very natural way. Let the distribution of the lowest cost for a good to be
Frechet such that
Ii
F1 (c1 ) = 1 − exp − cθ .
µ (Ω) 1
The measure of firms with unit cost less than C (1) ≤ c1 , is µ (Ω) F1 (c1 ) . Taking the limit
of this expression for the number of potential varieties µ (Ω) → +∞ we can show that
the distribution of the best producer’s cost of a variety is Pareto. More details are given in
appendix (13.1).
73
Suppose that ϕ ∈ [1, ∞) and:
Gi ( ϕ) = 1 − ϕ−θi , (4.43)
where θi is the shape parameter of the distribution. We assume that θi > σ − 1 (this paramet-
ric assumption is necessary in order for trade flows to be finite). Note that as θi increases,
the probability that the productivity is below any given ϕ increases, i.e. the heterogeneity
of producers is decreasing in θi .
If the productivities are Pareto distributed, then we can write:
ˆ ∞ ˆ ∞
!
d 1 − ϕ − θi
ϕσ−1 dGi ( ϕ) = ϕ σ −1 dϕ ⇐⇒
ϕij∗ ϕij∗ dϕ
ˆ ∞
= θi ϕσ−θi −2 dϕ ⇐⇒
ϕij∗
θi σ − θ i −1
= ϕij∗
θi + 1 − σ
Recall from equation (4.39) above that we can write the export threshold ϕij∗ as a func-
tion of the fixed cost of export so that:
ˆ ∞ σ − θ i −1
θi
ϕσ−1 dGi ( ϕ) = ϕij∗ ⇐⇒
ϕij∗ θi + 1 − σ
σ − θ i −1
σ
σ −1 ! σ −1
θi σ f ij σ−1 wi τij
= (4.44)
θi + 1 − σ Yj Pjσ−1
74
Substituting expression (4.44) into the gravity equation (4.42) above then yields:
1− σ ˆ ∞
!
σ
Xij = τij1−σ wi1−σ Mi ϕ σ −1
dGi ( ϕ) Yj Pjσ−1 ⇐⇒
σ−1 ϕij∗
σ − θ i −1
1− σ σ
σ −1 ! σ −1
σ f ij σ −1 wi τij
σ θi
Xij = τij1−σ wi1−σ Mi σ −1
Yj Pj ⇐⇒
σ−1 θi + 1 − σ Yj Pjσ−1
− θi σ − θ i −1 σθ−i 1
Xij = C1 τij wi f ijσ −1
Mi Yj Pjσ−1 (4.45)
σ − θ i −1 − θi
θi
where C1 ≡ σ σ −1 σ
σ −1 θ i +1− σ .
4.5.4 Trade with firm heterogeneity
Armed with the gravity equation (4.45) we have calculated, we now turn to the implica-
tions of a trade model with heterogeneous firms.
Extensive and intensive margins of trade
Equation (4.45) bears a resemblance to the gravity equation derived by Krugman (1980),
but the elasticity of trade flows with respect to variable trade costs is related to the Pareto
shape parameter instead of the elasticity of substitution! Since we have assumed that
θi > σ − 1, this means that trade flows have become more responsive to changes in trade
costs than in the Krugman (1980) model.
What gives? Intuitively, as trade costs fall two things happen: first, the firms already
producing will export more (this is known as the intensive margin); second, smaller firms
who were not exporting previously will begin to export (this is known as the extensive
margin). Both of these effects will tend to increase trade; since the Krugman (1980) model
only had the first effect, the model with heterogeneous firms will predict larger responses
75
of trade flows to changes in trade costs.
We can actually determine the elasticity of both margins of trade separately to see the
relative importance of both effects. (This was the central point of Chaney (2008)To do so,
recall the Leibnez rule:
ˆ b(z) ˆ b(z)
∂ ∂ f ( x, z) ∂b (z) ∂a (z)
f ( x, z) dx = dx + f (b (z) , z) − f ( a (z) , z)
∂z a(z) a(z) ∂z ∂z ∂z
Combining the original gravity equation (4.34) with the threshold exporting decision
(4.39) we have:
ˆ ∞
Xij = Mi xij ( ϕ) dGi ( ϕ)
ϕij∗
Hence we can write the elasticity of trade flows with respect to variable trade costs as:
´∞ ∂ϕ∗
∂ ln Xij ∂Xij τij ϕij∗
∂
∂τij xij ( ϕ ) τij dGi ( ϕ ) xij ϕij∗ τij ∂τijij dGi ϕij∗
− =− =− ´∞ + ´∞ ,
∂ ln τij ∂τij Xij ϕ∗ xij ( ϕ ) dGi ( ϕ ) ϕ∗ xij ( ϕ ) dGi ( ϕ )
ij ij
where the first term reflects the effect of a change in trade costs on the intensive mar-
gin and the second term reflects the change on the extensive margin. From the revenue
equation (4.36) we have:
1− σ
xij ( ϕ)
∂ ∂ σ wi
xij ( ϕ) = τij Yj Pjσ−1 = (1 − σ) .
∂τij ∂τij σ−1 ϕ τij
so that: ´∞ ∂
ϕij∗ ∂τij xij ( ϕ ) τij dGi ( ϕ )
− ´∞ = σ − 1,
ϕij∗ xij ( ϕ ) dGi ( ϕ )
i.e. a decline in trade costs will cause all firms currently producing to increase their pro-
duction with an elasticity of σ − 1 (this is the original Krugman (1980) effect). [Class
question: Why is the intensive margin increasing with σ?]
76
From equation (4.39) governing the threshold productivity:
1
σ − 1 ! σ −1
∂ϕij∗ ∂ σ f ij σ
σ−1 wi τij
ϕij∗
= =
∂τij ∂τij Yj Pjσ−1 τij
so that:
∂ϕ∗
xij ϕij∗ τij ∂τijij dGi ( ϕ) xij ϕij∗ ϕij∗ dGi ϕij∗
´∞ = ´∞ ⇐⇒
ϕij∗ xij ( ϕ ) dGi ( ϕ ) ϕij∗ xij ( ϕ ) dGi ( ϕ )
1− σ σ −1− θ i
σ −1
σ
σ −1 w i τij Yj Pj ϕij∗
= 1− σ ´∞ ⇐⇒
σ
σ −1 wi τij Yj Pjσ−1 ϕ∗ ϕσ−2−θi dϕ
ij
σ −1− θ i
ϕij∗
= σ−1−θi ⇐⇒
1 ∗
θi −σ+1 ϕij
= θi − σ + 1,
i.e. on the extensive margin, a decline in trade costs will induce less productive firms to
enter the market. When the elasticity of substitution is low (i.e. σ is low), even the less
productive firms will be able to capture relatively large market share, so that the difference
in size between the entering firms and the existing firms is small, meaning that the effect
on the extensive margin will be larger. With a Pareto distribution, the extensive margin
dominates the intensive margin.
Free entry and the allocation of factors across firms
Up until now, we have taken the mass of producing firms Mi to be exogenous. We now
consider what would happen if it was endogenously determined by a free entry condition
(much as in Krugman (1980)). Suppose now that firms have to incur an entry cost f ie > 0
77
prior to learning their productivity. Then the free entry condition will require that the
expected profits are equal to the entry cost:
" #
f ie = E ϕ ∑ max
πij ( ϕ) − f ij , 0 (4.46)
j∈S
Since more productive firms are (weakly) more profitable in every market, this implies
there will be an equilibrium productivity threshold ϕi∗ where firms, upon drawing a pro-
ductivity, will choose to produce only if their productivity exceeds ϕi∗ . This implies that
we can write re-write equation (4.46) as:
ˆ ∞
f ie = ∑ max
πij ( ϕ) − f ij , 0 dGi ( ϕ) ⇐⇒
ϕ∗ j∈S
ˆ ∞
f ie = ∑
n o πij ( ϕ) − f ij dGi ( ϕ) , (4.47)
j∈S max ϕi∗ ,ϕij∗
i.e. the fixed entry cost is simply equal to the sum across all destinations of the profits
in those destinations for firms who are sufficiently productive to both pay the fixed entry
and export costs.
What would happen to the profits of firms of different productivities if we were to
lower the variable trade cost τij for some j ∈ S? First, consider a firm whose productivity
is greater than the threshold productivity necessary to export to j, i.e. ϕ ≥ ϕij∗ . From
equation (4.37), its profits are:
1− σ
1 σ wi
πij ( ϕ) = τij Yj Pjσ−1 − f ij ,
σ σ−1 ϕ
so that holding all else equal, its total profits must increase. Furthermore, the more pro-
∂2 πij ( ϕ)
ductive this firm is, the greater the increase in its profits since − ∂τij ∂ϕ > 0. If the total
profits for all firms ϕ ≥ ϕij∗ are increasing, then the expected profits of entering the market
78
will also increase; because of the free entry condition, this will induce a greater number
of firms to enter the market, increasing the demand for local labor and driving up wages.
As a result, the profits of firms with ϕ < ϕij∗ will go down (as will the firms with produc-
tivities ϕ < ϕij∗ + ε), so that in equilibrium only the profits of the most productive firms
productivities will increase. In addition, as the wages increase, the minimum productiv-
ity required to produce anything at all (i.e. ϕi∗ ) will increase, forcing the least productive
firms in the model to exit. Hence, the model implies that greater openness to trade will
increase the average productivity of producing firms and will allocate labor toward the
more productive firms.
4.5.5 Next steps
The Melitz (2003) model provides the backbone for many (most?) of the major trade
papers written in the past ten years. While we will not have time to discuss its many
extensions in detail, we should note a few. Helpman, Melitz, and Yeaple (2004) endog-
enizes a firm’s decision whether to export or to pursue foreign direct investment (FDI).
Melitz and Ottaviano (2008) derive a version of the model with linear demand (instead of
CES) to analyze how mark-ups endogenously respond to trade liberalizations. Helpman,
Melitz, and Rubinstein (2008) discuss how the model (with a bounded distribution of
productivity) can be used to explain the zero trade flows observed in bilateral trade data
and what it suggests for the estimation of empirical gravity models. Helpman, Itskhoki,
and Redding (2010) incorporate labor market frictions into a Melitz (2003) framework.
Arkolakis (2010b) extends the Melitz (2003) framework to incorporate market penetration
costs. Eaton, Kortum, and Kramarz (2011) use the Melitz (2003) framework to structurally
estimate the exporting behavior of French firms.
79
4.6 Summary
• Established a macroeconomic framework where the concept of the firm had a mean-
ing while the model was tractable and amenable to a variety of exercises. In this
framework it is possible to think about trade liberalization and firms in GE.
• Explicitly modeled the importance of reallocation of production through the death
of the least productive firms.
The Melitz (2003) model provides the backbone for many of the major trade papers writ-
ten in the past years. Here we note just a few. Helpman, Melitz, and Yeaple (2004) endo-
genizes a firm’s decision whether to export or to pursue foreign direct investment (FDI).
Melitz and Ottaviano (2008) derive a version of the model with linear demand (instead of
CES) to analyze how mark-ups endogenously respond to trade liberalizations. Helpman,
Melitz, and Rubinstein (2008) discuss how the model (with a bounded distribution of
productivity) can be used to explain the zero trade flows observed in bilateral trade data
and what it suggests for the estimation of empirical gravity models. Helpman, Itskhoki,
and Redding (2010) incorporate labor market frictions into a Melitz (2003) framework.
Arkolakis (2010b) extends the Melitz (2003) framework to incorporate market penetration
costs. Eaton, Kortum, and Kramarz (2011) us
The Eaton and Kortum (2002) model remains the primary framework for the study
of trade in perfectly competitive markets (especially agriculture). However, because of a
lack of a real concept of a firm, it has proven less popular than the Melitz (2003) model for
the study of firm-level data. In Bernard, Eaton, Jensen, and Kortum (2003), the authors
did extend the framework to one where there is Bertrand competition between firms. The
basic idea is straightforward: the price charged by the single firm that exports a variety is
the marginal cost of the second best firm. While this allows for endogenous (non-constant)
80
markups, the extension required somewhat more complicated probability tools and has
turned out to be slightly less tractable than Melitz (2003) extensions such as Melitz and
Ottaviano (2008), where there are endogenous markups due to non-CES preferences.
4.7 Problem Sets
1. General properties of gravity trade models. Consider the model developed in section
4.3-4.5.
(a) Argue that in all these models profits are a constant fraction of production.
(b) Argue that in the model consider in Section (4.5) payments to fixed costs are a
constant fraction of production.
2. The Frechet distribution. For all n ∈ {1, ..., N }, suppose that the random variable
zn ≥ 0 is distributed according to the Frechet distribution, i.e.:
Pr{zn ≤ z} = exp − Tn z−θ ,
cn
where Ti > 0 and θ > 0 are known parameters. Define the random variable pn = zn .
Calculate:
πin ≡ Pr pn ≤ min pk
k6=n
3. The Pareto distribution. Suppose that the random variable zn ∈ [bn , ∞) is distributed
according to the Pareto distribution, i.e.:
−θ
bn
Pr{zn ≤ z} = 1 − ,
z
where bn > 0 and zn > 0 are known parameters. What is the distribution of zn
81
conditional on being greater than cn > bn ?
82
Chapter 5
Closing the model
In order to determine the solution of the endogenous variables of the models we con-
structed above in the general equilibrium. The individual goods markets are already as-
sumed to clear since we replaced the consumer demand directly in the sales of the firm
for each good.
In this section we will make use of three general assumption.Consider the
Aggregate profits are a constant share of revenues. Let Π j denote country j’s aggregate
profits gross of entry costs (if any). The first macro-level restriction states that Π j must be
a constant share of country j’s total revenues:
R1 For any country j, Π j /Yj = π̄ where π̄ ≥ 0.
Under perfect competition, R1 trivially holds since aggregate profits are equal to zero.
Under monopolistic competition with homogeneous firms, R12 also necessarily holds be-
cause of Dixit-Stiglitz preferences; see Krugman (1980). In more general environments,
however, R1 is a non-trivial restriction.
The second restriction
83
R2 For any country pair,i, j, the share of spending on fixed exporting cost to bilateral sales
is constantγij = γ̄ where γ̄ ≥ 0.
The third restriction is that the value of imports of goods must be equal to the value of
exports of goods:
R3 For any country j ∑k Xkj = ∑k X jk
In general, total income of the representative agent in country j may also depend on
the wages paid to foreign workers by country j’s firms as well as the wages paid by foreign
firms to country j’s workers. Thus, total expenditure in country j, X j ≡ ∑i Xij , could be
different from country j’s total revenues, Yj ≡ ∑in=1 X ji . R3 rules out this possibility.
5.1 General Equilibrium
We now show how we can determine the wages, wi , and spending, Xi that solve for the
model’s general equilibrium (income, yi , can be written always as a straightforward func-
tion of spending in the cases we will analyze). It turns out that under the technologi-
cal distributions and demand structures that we introduced above, we can first solve for
wages and spending and all the rest of the variables can be written as simple functions
of these two variables. To create a formal mapping to the data, where trade deficits are a
commonplace, we can also allow for exogenous transfer payments to countries, Di , (fol-
lowing Dekle, Eaton, and Kortum (2008)) which in a static model will imply an equal
amount of trade deficit. Of course, these trade deficits have to sum up to zero across
countries, ∑i Di = 0.
In principle, to solve for wi , Xi we need to consider two sets of equations.
84
i) the budget constraint of the representative consumer
∑ Xki = ∑ λki Xi = Xi ≡ wi Li + πi + Di , (5.1)
k k
where πi is total profits earned by firms from country i net of fixed marketing costs (if
any), and
ii) the current account balance (there are no capital flows) and consists of exports,
imports and related payment to labor for fixed marketing costs but can be equivalently
written as total expenditure equals total income and transfers
0= ∑ Xik − ∑ Xki + ∑ γki Xki − ∑ γik Xik + Di =⇒
k 6 =i k 6 =i k 6 =i k 6 =i
| {z } | {z } | {z }
exports imports net foreign income
0= ∑ Xik − ∑ Xki + ∑ γki Xki − ∑ γik Xik + Di =⇒
k k k k
Xi = ∑ Xki = ∑ Xik + ∑ γki Xki − ∑ γik Xik + Di , (5.2)
k k k k
where γij is the share of bilateral sales from i to j that accrues to labor for payments of
fixed marketing costs, and trade flows Xij . In our context, the current account balance is
effectively equivalent to the labor market clearing condition.
These set of equations can be used to solve for wi , Xi using an additional normaliza-
tion. Notice that is straightforward to show that with the CES demand we assume budget
85
balance is equivalent to the CES price index. In particular,
ˆ
∑ Ωk
p1ki−σ (ω ) dω = Pi1−σ ⇐⇒
k
´
Ωk p1ki−σ (ω ) dω
∑ Pi1−σ
Xi = Xi ⇐⇒
k
∑ λki Xi = Xi .
k
5.2 Solving for the Equilibrium when the Profit share is constant
From this general framework we need to specify the exact market structure to solve for
the equilibrium wages. With perfect competition, as in Sections 3.3,4.3, there are no fixed
marketing costs, γij = 0. Thus, these two sets of equations can be thought as one set of
equations with wages remaining to be solved as a non-linear equation on wages, given
equation (3.10). Another simple case is that of monopolistic competition with homoge-
neous firms, Section 3.4. In that case there are again no fixed marketing costs and all
income is ultimately accrued to labor due to free entry so that the budget constraint can
be written as Xi = wi Li + Di and given this relationship and (3.18), equation (5.2) can be
used to solve for all wages.
There are a number of cases where profits are not going to be zero in equilibrium
and thus we have to solve for those. But in all the cases considered above, it turns out
that profits are a constant fraction of country income, i.e. πi = π̄/ ∑k Xik . Notice that
in the case of Bertrand competition, Section 4.4, γki = 0 as well and since profits are
constant fraction of income, π̄ = 1/ (θ + 1) similarly wages can be solved using (3.18). To
proceed to characterize 5.2 with monopolistic competition and firm heterogeneity notice
that under the assumption of Pareto distributions and fixed marketing costs, it can be
shown that γij = (θ − σ + 1) / (σθ ) ≡ γ and that πi is a constant share of output π̄ =
86
(σ − 1) / (σθ ) (independent of entry being endogenous or not).
With R1,R2 we can develop a very general framework to determine the equilibrium of
the endogenous variables. Start by assuming that γij = γ̄, equation (5.2) can be written
as1
∑ Xki = Xi = ∑ Xik + γ̄ ∑ Xki − γ̄ ∑ Xik + Di (5.3)
k k k k
Now we can combine the above reformulation of the current account balance condition
with the budget constraint of the consumer equation (5.1) (equivalently the labor market
clearing), we obtain
wi Li + πi − γ̄ ∑ Xki = (1 − γ̄) ∑ Xik
k k
and using again the budget constraint of the individual we can rewrite this equation as
!
Di
wi Li + πi + Di = (1 − γ̄) ∑ Xik + γ̄ ∑ Xik + 1 − γ̄ + Di =⇒
k k
γ̄
wi L i + π i −
1 − γ̄
Di = ∑ Xik . (5.4)
k
Now note that with R1 we have
πi = π̄ ∑ Xik = π̄ ∑ λik Xk . (5.5)
k k
Combined with (5.4)
γ̄ πi
wi L i + π i − Di = =⇒
1 − γ̄ π̄
π̄ γ̄
wi L i − Di = π i (5.6)
1 − π̄ 1 − γ̄
1 Notice that the following equation implies that R2 with Di = 0 implies R3.
87
and thus we can write
1 π̄ γ̄
wi L i + π i = wi L i − Di .
1 − π̄ 1 − π̄ 1 − γ̄
We can thus summarize the equilibrium in all the models above as a set of wages that
solves
γ̄
wi L i + π i −
1 − γ̄
Di = ∑ Xik =⇒
k
γ̄ γ̄
wi L i −
1 − γ̄
Di = ∑ λik wk Lk + 1 − π̄ − π̄
1 − γ̄
Dk (5.7)
k
with λik defined depending on the model. Using equations (3.10), (4.17) λij is an explicit
function of wages, wi , and entry, Ni , alone. In the case of perfect competition the number
of entering varieties is fixed so that the only equation required for the solution is (3.10),
and the normalization, (4.17). Many papers assume monopolistic competition, including
the original papers of Krugman (1980) and Melitz (2003). To characterize the equilibrium
in those case we need to either assume that Ni is given, or solve for entry, Ni . We do the
latter next.
5.3 Endogenous Entry
Assume that new firms can freely choose to enter the economy and draw a productivity
from a distribution gi (z) upon paying an entry cost f e , in terms of labor units, where
the distribution gi (z) could be degenerate so that all the mass is in a single point. Two
restrictions are enough to guarantee that there is a simple solution for Ni .
We start from the free entry condtion, that indicates that the total profits of firms from
88
j gross of entry fixed costs, Π j , equal to fixed costs of entry
Πj
= w j f je .
Nj
Using R1 together with the above equation we have
ζYj = Nj w j f je ,
for some constant ζ.
Now use R3 implying Yj = X j where the last equals to total labor income from the free
entry condition, X j = w j L j . Therefore, we have
ζYj = Nj w j f je =⇒
ζw j L j = Nj w j f je =⇒
Lj
Nj = ζ, (5.8)
f je
i.e. entry is linear in population and does not depend on trade costs.
Taking stock this result implies that the equilibrium can be solved using the set of
equations defined by (5.7) and one normalization (due to the redundancy of one equation
as a result of Walras law) and entry given by (5.8). In the case of the models of Krugman
(1980) and Melitz (2003) Ni is endogenously determined and we need to introduce the zero
profit condition. Assuming R3 in this case, as we have illustrated, implies that Ni is linear
in population and the same set of equations can be used to solved for wages. Finally, Xi
can be found by utilizing the solution for wages, equation (5.6) which expresses profits as
a function of wages, and the budget constraint of the representative consumer, equation
(5.1).
89
5.4 Labor Mobility
To introduce economic geography in this general setup we assume that workers freely
move across regions. In equilibrium this results to welfare equalization across locations.
Because of the equivalence of the budget constraint with the price index we can simply
consider the same two equations for equilibrium as above. For simplicity we will consider
henceforth the case Di = 0. Using equation 5.7, given assumptions R1, R2, we have
Xi = ∑ λik Xk (5.9)
k
ˆ
pik (ω ) 1−σ
L i wi = ∑ Ωk Pk1−σ
wk L k . (5.10)
k
ˆ
(cwi τik / (zAk )) 1−σ
L i wi = ∑ Ni ∗
zik Pk1−σ
wk Lk g (z) dz. (5.11)
k
with Armington
w i 1− σ 1− σ σ −1
L i wi = ∑ Pk1−σ ik
τ Ai Lk w1k −σ .
k
Li wiσ = ∑ Wi 1−σ τis 1−σ Aiσ−1 Ls wσs . (5.12)
k
The price index implies
ˆ
P1i −σ =∑ pki (ω ) 1−σ dω (5.13)
k Ωk
ˆ
P1i −σ = ∑ Nk ∗
zik
c1−σ w1k −σ τki 1−σ zσ−1 Aσk −1 g (z) dz (5.14)
k
with Armington
w1i −σ = ∑ W 1i −σ T1ki−σ Aσk −1 w1k−σ . (5.15)
k
90
When welfare is equalized so that Wi = W for all i ∈ S equations (5.9) and (5.13) are lin-
ear operators whose eigenfunctions are Li wiσ and w1i −σ and whose eigenvalues are W σ−1 ,
respectively. Note that the kernels of the two equations are transposes of each other. A
spatial economy equilibrium is defined as wi , Li and W that solve equations (5.9), (5.13).
These two results allow us to prove the following theorem:
Theorem 6. Consider a geography model characterized by equations (5.9), (5.13). Then:
i) there exists a unique spatial equilibrium and this equilibrium is regular; and
ii) this equilibrium can be computed as the uniform limit of a simple iterative proce-
dure.
Proof. See Allen and Arkolakis (2014).
5.5 Problem Sets
1. Consider the monopolistic competition model with heterogeneous firms and Pareto
distribution, considered in section (4.5).
(a) Prove that the share of fixed costs and net profits are a constant fraction of the
overall firm revenues. Compute those fractions.
2. Show that when bilateral trade costs are symmetric a spatial equilibrium can be
written as a single non-linear integral equation, which will allow to provide a sim-
ple characterization of the equilibrium system. Prove this both for a trade and a
geography model.
3. Consider a geography model and assume that we are in a spatial equilibrium where
Wi = W for all countries. Prove that the solution of both equations (5.12) and (5.15)
results to the same eigenvalue.
91
Chapter 6
Model Characterization
6.1 The Concept of a Model Isomorphism
It turns out that in the simple monopolistic competition framework with Pareto distri-
bution of productivities of firms, the assumption of endogenous entry has little bite (see
Arkolakis, Demidova, Klenow, and Rodrı́guez-Clare (2008)): the model with free entry is
mathematically equivalent to one with a predetermined number of entrants (essentially
the Chaney (2008) version of Melitz (2003)) where the number of entrants is proportion-
ate to the population. The only difference between the two models is that all the profits
are accrued to labor allocated for the production of the fixed cost of entry. In addition, the
model delivers the same predictions for trade and welfare gains from trade as all the other
gravity models we studied so far (Armington (1969), Eaton and Kortum (2002), Bernard,
Eaton, Jensen, and Kortum (2003)). To understand these results we need to formally de-
fine an analyse the concept of a model isomorphism.
Consider two models, E1 = (α1 , V1 ) E2 = (α2 , V2 ) where αi , i = 1, 2 are a combination
of model parameters (and fundamentals i.e preference and production structure) and Vi is
92
a set of equilibrium outcomes. Formally, we define an isomorphism between two models
as a bijective mapping F between the parameters α1 ←→ α2 that leads to a bijective
F
mapping to the model outcomes (V1 ←→ V 2 ) such that given α1 ←→ α2 we have that each
element v1i ∈ V1 , v2i ∈V2 are such that v1i = f i v2i where f ∈ F up to a normalization,
i.e. there exist parameter choices such that you can redefine the model outcomes in an
equivalent way. The second requirement is very general in the sense that the variables
of the one model are transformations of the variables of the other model. However, an
isomorphism is not a mathematical formalism in this case for two reasons. First, models
that are isomorphic, under certain environments yield the same policy prescriptions for
a given change in policy (in our case, for example, that could be a percentage change in
trade costs). Second, the mathematical and statistical apparatus that can be used for the
solution of one model can be, as a result, used for the other model as well.
Below we illustrate two examples of isomorphisms that are key in our analysis. The
first is a precise isomorphism, where f i = Ai i.e. the transformation involves only a con-
stant. The second, is a more general isomorphism where the transformation potentially
involves more than a constant transformation, but the policy prescriptions of the models
are the same.
6.1.1 An Exact Isomorphism
We will use the concept of the isomorphism to illustrate that a model with exogenous en-
try (as in Chaney (2008)) is isomorphic to a model with endogenous entry (as in Melitz
(2003), with the specification as done by Arkolakis, Demidova, Klenow, and Rodrı́guez-
Clare (2008)). These models are summarized in the analysis of Section 4.5 without ex-
plicitly specifying the labor market clearing condition or the zero profit condition in the
endogenous entry case. In the derivations below we assume that profits accrue to domes-
93
tic consumers.
Assuming zero profits in expectation the expected profits of a firm must be equal to en-
try costs.1 Using the free entry condition and a Pareto distribution with shape parameter
Aiθ
θ > σ − 1, c.d.f. G (z; Ai ) = 1 − zθ
, and support [ Ai , +∞) we have2
ˆ
∑ ∗
zik
πik (z) dG (z; Ai ) = wi f e =⇒
k
ˆ σ τik wi 1−σ
ˆ
Aiθ Aiθ
∑ ∗
zik
σ −1 z
Pk1−σ σ
w L
k k θ
z θ +1
dz −∑
∗
zik
wk f ik θ
z θ +1
dz = wi f e =⇒
k k
θ Aiθ Aiθ
∑ wk fik θ − σ + 1 ∗
θ
− ∑ w k
∗
e
θ f ik = wi f =⇒
k zik k zik
w Aiθ σ−1
∑ wki fik = f e. (6.1)
zik θ − σ + 1
∗
θ
k
We now combine the free entry condition and a reformulation of the labor market
clearing condition to compute the equilibrium of the model. Notice that the equilibrium
number of entrants in country i, Ni , is determined by the following labor market clearing
1 Essentially, we assume that there exists a perfect capital market, which requires firms to pay a fixed entry
cost before drawing a productivity realization. Consequently, we multiply the LHS by 1 − G zii∗ , bi , the
probability of obtaining the average profit, since firms with profits below this average necessarily exit the
market. Alternatively, we could have specified a more general case with intertemporal discounting, δ. In this
case the expected profits from entry should equal the discounted entry cost in the equilibrium.
2 An implication of free entry is that in the equilibrium all the profits are accrued to labor for the production
of the entry cost.
94
condition:
ˆ
σ τik wi −σ
∗ θ
τik zik Aiθ Aiθ
Ni ∑ σ −1 z
w k L k θ θ +1 dz + f ∑ k
e
+ N f ki = Li =⇒ (6.2)
Pk1−σ z z ∗ θ ∗ θ
k z∗
ik zik k z
| {z } | {zki }
labor used into production labor for fixed costs
!
w Aiθ θ Aiθ
Ni ∑ (σ − 1) wki fik ∗
θ
− + 1
+ fe + ∑ Nk θ f ki = Li . (6.3)
k zik θ σ k z∗ki
Substituting out equation (6.1), we obtain
Ni (θ f e + f e ) + ∑ Nk Aiθ (z∗ki )−θ f ki = Li ,
k
which, together with the price index, and the definition of zij∗ and the fact that all income
accrues to labor, Yj = w j L j ,
ˆ 1− σ
σ τki wk
Pi1−σ = ∑ Nk z∗ki σ−1 z
θ Aiθ z−θ −1 dz =⇒
k
!
θσ
wi L i =
θ−σ+1 ∑ Nk Aiθ (z∗ki )−θ wi f ki ,
k
implies that
σ−1
Ni = Li , (6.4)
θσ f e
giving an explicit parametric form to equation 5.8.3
Notice that total export sales from country i to j are given by expression (4.45).4 Define
the fraction of total income of country j spent on goods from country i by λij . Using the
3 With a slightly altered proof the same results hold under the assumption that fixed costs are paid in terms
of domestic labor.
4 Average sales of firms from i conditional on operating in j are the same in the model with free entry and
the one with a predetermined number of entrants.
95
definition of total sales from i to j and equations (4.39) and (6.4), we have that
Xij
λij = =⇒
∑k Xkj
−θ 1−θ/(σ−1)
Li Aiθ τij wi f ij
λij = −θ 1−θ/(σ−1)
. (6.5)
∑k Lk Aθk τkj wk f kj
and the equilibrium wages can be simply determined using the labor market clearing con-
dition (5.7) given the fact that fixed marketing costs are a constant proportion of bilateral
sales, γ.
6.1.2 A Partial Isomorphism
Now consider all the models that we discussed above and assume that their parameters
are such that the models yield the same level of domestic share of spending λ jj . As we will
argue at a later point this can be done with different ways in different models. However,
insofar as the model generate the welfare equation of the form Wj = Cj λ−
jj
1/e
where Cj , ε >
0 are model dependent constnats, the welfare gains from the expansion of trade are the
same and equal to Ŵj = λ̂−
jj
1/e
., as long as all the models are specified with the same
parameter ε. In other words, the model may even give different implications for overall
trade but they give isomorphism implications for the welfare gains from trade. This point
is made in detail in Arkolakis, Costinot, and Rodrı́guez-Clare (2012).
6.2 General Equilibrium: Existence and Uniqueness
Because the various setups that we have studied share a common gravity form, their
equilibrium analysis turns out to be simpler than it initially appears. The starting point
is gravity models that yield the relationship between aggregate bilateral trade flows and
96
model variables and parameters equation (3.5),
Condition 1. For any countries, i and j ∈ S the value of aggregate bilateral flows is given
by
Xij = Kij γi δj , (6.6)
To consider these models in general equilibrium two conditions have to hold: labor mar-
ket clearing conditions and current account balance. The first condition implies that in-
come generated in a country has to equal to total sales to all destinations,
Condition 2. For any location i ∈ S,
Yi = ∑ Xij . (6.7)
j
In addition, current account, if there are no capital flows or transfer implies trade
balance,
Condition 3. For any location i ∈ S,
Yi = ∑ Xji . (6.8)
j
Notice, that in addition all these models have to satisfy Warlas law, so that one of our
equations is redudandant.5 For that reason we add a normalization that world income
5 To
see this note that summing these two equations over all i 6= N and equating them we obtain
∑i6= N ∑ j Xij = ∑i6= N ∑ j X ji . By the definition of gross world income being total trade across all markets
we obtain trade balance for the Nth location which implies Warlas’ law.
97
equals to one:6
∑ Yi = 1. (6.9)
i
Inspecting the above equations, it is clear that they do not impose sufficient structure
to solve the model since there is no restriction on the form that Yi can essentially take. To
these essential conditions of the model we add one more, that restricts furthermore the
class of models that we focus on. This additional restriction differs across gravity trade
and geography models.
Relationship between income and the shifters in gravity trade models. Our last
condition for a trade model postulates a log-linear parametric relationship between gross
income and the exporting and importing shifters:
Condition 4. For any location i ∈ S, Yi = Bi γiα δi , where we define α ∈ R and β ∈ R to be
β
the gravity constants and Bi > 0 is an (exogenous) location specific shifter.
We now provide sufficient conditions for establishing existence and uniqueness in
a general equilibrium gravity model. We start by formulating the equilibrium system
implied by our assumptions. Using equations (6.7) and (6.8) and substituting equations
Xij and Yi with (6.6) and C.4, respectively, yields:
Bi γiα−1 δi = ∑ Kij δj
β
(6.10)
j
and
= ∑ K ji γ j
β −1
Bi γiα δi (6.11)
j
6 This is a valid normalization as long as α 6= β. When α = β, a suitable alternative normalization is
∑i∈S γi = 1. None of the following results, unless explicitly noted, depend on the normalization chosen.
98
and using C.4, the normalization equation (6.9) becomes:
∑ Bi γiα δi
β
= 1. (6.12)
i
Thus, given model fundamentals Bi , Kij and gravity constants α, β, equilibrium is defined
as γi and δi for all i ∈ S such that equations (6.10), (6.11) and (6.12) are satisfied. In
the special case where α = β = 1, it is immediately evident from equations (6.10) that
Kij K ji
(6.11) have a solution only if the matrices with elements Bi and Bi both have a largest
eigenvalue equal to one. Since this will not generally be true, in what follows we exclude
this case.
Based on this formulation we can prove the following theorem:
Theorem 7. Consider any general equilibrium gravity model. If α + β 6= 1, then:
i) The model has a positive solution and all possible solutions are positive;
ii) If α, β ≤ 0 or α, β ≥ 1, then the solution is unique.
Proof. See Allen, Arkolakis, and Takahashi (2014).
Our approach can also be naturally extended to allow for labor mobility as in eco-
nomic geography models. To do so, we slightly alter condition C.4 to allow for the gross
income in a location to depend additionally on an endogenous constant λ, which can be
interpreted as a monotonic transformation of welfare (which is equalized across locations
in economic geography models). It is straightforward to show that the economic geogra-
phy model of Allen and Arkolakis (2014) (which under certain parametric configurations
is isomorphic to the economic geography models of Helpman (1998), Redding (2012), and
?) satisfies the following condition
1 α β
Condition 5. For any location i ∈ S, Yi = λ Bi γi δi , where λ > 0 is an endogenous variable
99
and all other variables are as above. Furthermore, we require that λc = ∑i Ci γid δie for some
c, d, e ∈ R.
Given this alternative condition, we modify part (ii) of Theorem 7 slightly to prove the
following Corollary:
Corollary 1. Consider any economic geography model that satisfies conditions C.1, C.2, C.3, C.4,
and C.5. Then (i) there exists a solution as long as α + β 6= 1; and (ii) the equilibrium is unique if
α, β ≤ 0 or α, β > 1.
Proof. See Allen, Arkolakis, and Takahashi (2014).
Notice that the additional normalization is required to determine the level of the en-
dogenous variable λ. In the economic geograhy example we consider in the previous
section this endogenous variables corresponds to a monotonic transformation the overall
welfare level.
6.3 Analytical Characterization of the Gravity Model
Thus far, we have provided various microeconomic foundations for the gravity trade
model, defined the general equilibrium conditions, characterizes when an equilibrium
exists and when it is unique, and discussed some general equilibrium properties of grav-
ity trade models. Next, we are going to take our tools “out for a spin” and see what exactly
the general equilibrium properties imply for the Armington model. Remember that it is a
reasonably straightforward task to reinterpret the Armington model in other frameworks
(i.e. there exist formal isomorphisms), so the choice of the Armington model is not partic-
ularly important. While most of this class will be applying the tools we have developed
to a particular example, we think doing so both reinforces the power of the tools we have
and provides new insights into the mechanisms at play.
100
It turns out that we can extend the range in which uniqueness is guaranteed if we
constrain our analysis to a particular class of trade frictions which are the focus of a large
empirical literature on estimating gravity trade models. We call these trade frictions quasi-
symmetric.
Definition 2. Quasi Symmetry: We say the trade frictions matrix K is quasi-symmetric if
there exists a symmetric N × N matrix K̃ and N × 1 vectors K A and K B such that for all
i, j ∈ S we have:
Kij = K̃ij KiA K jB , where K̃ij = K̃ ji
Loosely speaking, quasi-symmetric trade frictions are those that are reducible to a
symmetric component and exporter- and importer-specific components. While restrictive,
it is important to note that the vast majority of papers which estimate gravity equations
assume that trade frictions are quasi-symmetric; for example Eaton and Kortum (2002)
and Waugh (2010) assume that trade costs are composed by a symmetric component that
depends on bilateral distance and on a destination or origin fixed effect.
When trade frictions are quasi-symmetric it can show that the system of equations
(6.10) and (6.11) can be dramatically simplified, and the uniqueness more sharply charac-
terized.
Theorem 8. Consider any general equilibrium gravity model with quasi-symmetric trade costs.
Then:
i) The balanced trade condition is equivalent to the origin and destination shifters being equal
up to scale, i.e.
γi KiA = κδi KiB (6.13)
for some κ > 0 that is part of the solution of the equilibrium.
ii) If α + β ≤ 0 or α + β ≥ 2, the model has a unique positive solution.
101
Part (i) of the Theorem 8 is particularly useful since it allows to simplify the equilib-
rium system into a single non-linear equation:
1− β β
= κ β −1 ∑ K
α + β −1 eij B−1 KiA
γi i KiB K jA γ j . (6.14)
j
For any gravity trade model where trade frictions are quasi symmetric, if trade is bal-
anced, the goods market clearing condition holds, and the generalized labor market clear-
ing condition holds, then the equilibrium origin fixed effects satisfy the following set of
non-linear equations:
1
γ̃i = λ ∑ Fij γ̃ jα+β−1 , (6.15)
j∈S
K jA
− β
α + β −1 KiA
where γ̃i ≡ γi , λ≡ κ β −1 > 0, and Fij ≡ Kij K jB KiB
1
Bi > 0. This implies that
there will always exist a solution and the solution will be unique if α + β ≥ 2 or α + β ≤ 0.
Notice that the normalization implies with quasisymmetry that
∑ Yi = 1 =⇒
i
∑ Bi γiα δi
β
= 1 =⇒
i
!β
γi KiA
∑ Bi γiα κKiB
= 1 =⇒
i
!β
KiA
κ − β ∑ Bi
α+ β
γi =1
i KiB
so that the normalization can be used to pin down κ.
With economic geography this relationship holds but in addition, to determine the
level of λ ≡ κ β−1 , we make use of the additional condition 5.
102
Example: Armington model with quasi-symmetry Consider now an Armington model
with intermediate inputs, but now assume that trade costs are quasi-symmetric. From
part (i) of Theorem 8, we have γi KiA = κδi KiB , which implies:
!1− σ
wiδ Pi1−δ
KiA = κPiσ−1 wi Li KiB ,
Ai
or equivalently:
! 1
(1−σ)(2−δ)
1+(σ −1)δ
KB
Pi = wi (1−σ)(2−δ)
κLi A1i −σ iA . (6.16)
Ki
Equation (6.16) provides some intuition for the uniqueness condition presented in Theo-
rem 8: when σ < 12 , it is straightforward to show that the elasticity of the price index with
respect to the wage is less than one. This implies that the wealth effect may dominate the
substitution effect, so that the excess demand function need not be downward sloping.
In addition, combining equation (6.16) with equation (6.14), assuming δ = 1, and
wi
noting that welfare Wi = Pi yields the following equation:
∑ Kij Ai
(σ−1)σ̃ −(σ−1)σ̃
κWiσσ̃ Liσ̃ = Aσj σ̃ Lσ̃j Wj , (6.17)
j
σ −1 7
where σ̃ ≡ 2σ−1 . Equation (6.17) holds for both trade models (where labor is fixed) and
economic geography models (where labor is mobile); in the former case, Li is treated as
exogenous parameter and Wi solved for; in the latter case Li is treated as endogenous and
7 When there are only two countries (so that trade costs are necessarily quasi-symmetric), we can use
equation (6.17) to derive a single non-linear equation that yields the relative welfare in the two countries
W 1 (1−σ)σ̃
σσ̃
W 1 σ̃
W1
K22 − K11 + K21 = K12 .
W2 W2 W2
Comparative statics for welfare with respect to changes in Kij can be characterized using the implicit function
theorem in this case.
103
Wi is assumed to be constant across locations. Hence, Theorem 8 highlights the funda-
mental similarity between trade and economic geography models.
We now discuss the equilibrium when there are just two countries. Note that when
there are only two countries, all possible trade costs are quasi-symmetric. Define the
A 2σσ−1
K KB (σ−1)σ̃ σ̃ −σ̃
kernel Fij ≡ Kij KjB K iA Aσj σ̃ Ai L j Li (note that the kernel now includes all the
j i
exogenous variables in the model). Then the equilibrium conditions from equation (6.17)
can then be written (for an arbitrary number of countries) as:
∑ Fij Wj
(1−σ)σ̃
Wiσσ̃ = .
j∈S
With two countries, this becomes:
(1−σ)σ̃ (1−σ)σ̃
W1σσ̃ = F11 W1 + F12 W2
(1−σ)σ̃ (1−σ)σ̃
W2σσ̃ = F22 W2 + F21 W1
Dividing the first equation by the second yields:
σσ̃ (1−σ)σ̃ (1−σ)σ̃
W1 F11 W1 + F12 W2
= (1−σ)σ̃ (1−σ)σ̃
⇐⇒
W2 F22 W2 + F21 W1
W1 σσ̃
(1−σ)σ̃ (1−σ)σ̃ (1−σ)σ̃ (1−σ)σ̃
F22 W2 + F21 W1 = F11 W1 + F12 W2 ⇐⇒
W2
!
W1 σσ̃ W1 (1−σ)σ̃ W1 (1−σ)σ̃
F22 + F21 = F11 + F12 ⇐⇒
W2 W2 W2
W1 σσ̃ W1 (1−σ)σ̃ W1 σ̃
F22 − F11 + F21 = F12 (6.18)
W2 W2 W2
Equation (6.18) shows that with two countries, the equilibrium relative welfare in the two
104
regions is just the root of a polynomial equation! Furthermore, note that if if σ > 12 ,
(1−σ)σ̃
σ σ̃ σ̃
then ∂(W ∂/W2 ) F22 WW2
1
− F W1
11 W2 + F W1
21 W2 − F12 > 0 so that the implicit
1
function theorem implies:
∂ W1 ∂ W1 ∂ W1 ∂ W1
> 0, < 0, > 0, < 0.
∂F11 W2 ∂F21 W2 ∂F12 W2 ∂F22 W2
[Class question: what is the intuition for these comparative statics?]
6.4 Computing the Equilibrium (Alvarez-Lucas)
Using the methodology of Alvarez and Lucas (2007) it can be proven that the model with
this gravity structure has a unique equilibrium. To show existence Alvarez and Lucas
(2007) define the analog of an excess demand function, which in our context and with
zero exogenous deficits is given by,
−θ 1−θ/(σ −1)
1 Li Aiθ τij wi f ij
wi ∑
f i (w) = −θ 1−θ/(σ−1) w j L j − wi Li ,
j ∑k Lk Ak τkj wk f kj
θ
where w is the vector of wages, and they show that it satisfies the standard properties of
an excess demand function.8 To show uniqueness, the gross substitute property has to be
proven
∂ f i (w)
> 0 for all i 6= k
∂wk
∂ f i (w)
> 0 for all i
∂wi
and uniqueness follows from Propositon 17.F.3 of Mas-Colell, Whinston, and Green (1995).
8 These properties are continuity, homogeneity of degree zero, Warlas’ law, boundness from below and
infinite excess demand if one wage is 0. See Mas-Colell, Whinston, and Green (1995), Chapter 17.
105
To compute the equilibrium, notice that for some κ ∈ (0, 1] we can define a mapping
Ti (w) = wi [1 + κ f i (w) /Li ] (6.19)
Now if we start with wages that satisfy ∑i wi Li = 1, we have
∑ Ti (w) Li = ∑ wi Li + ∑ wi κ f (w)
i i i
−θ 1−θ/(σ −1)
Li Ai τij wi f ij
= 1 + κ ∑ ∑ −θ 1−θ/(σ−1)
w j L j − wi L i
i j ∑k Lk Ak τkj wk f kj
= 1 − κ ∑ wj Lj + κ ∑ wj Lj = 1
j j
so that Ti (w) is mapping w such that it maps ∑i wi Li = 1 to itself. By starting with an
initial guess of the wages, and updating according to (6.19) the system is guaranteed to
converge to the solution Ti (w) = wi (see Alvarez and Lucas (2007)).
6.5 Conducting Counterfactuals: The Dekle-Eaton-Kortum Pro-
cedure
Dekle, Eaton, and Kortum (2008) have established a methodology for calculating counter-
factual changes in the equilibrium variables with respect to changes in the iceberg costs
or technology parameters. The merit of this approach is that it does not require prior
information on the level of technology Ai and bilateral trade costs τij , but rather only per-
centage changes in the magnitudes of these parameters. The idea is to use data for the
endogenous variables λij , y j to calibrate the model in the initial equilibrium, and exploit
the fact that the level of technology Ai and bilateral trade costs τij are perfectly identified
106
given the values for λij , y j .
The procedure can be applied to most of the frameworks above, and in fact delivers
robust predictions for changes in trade and welfare as argued by Arkolakis, Costinot,
and Rodrı́guez-Clare (2012), under the simple assumption that the elasticity of trade with
respect to wages and trade costs is the same.
Denote the ratio of the variables in the new and the old equilibrium, e.g. ŵ j = w0j /w j .
We use labor in country j as our numeraire, w j = 1. We will make crucial use of the fact
that either profits are a constant fraction of income or that labor income is the only source
of income in the models above so that we also obtain that ŷi = ŵi for all i = 1, ..., n.
Under the assumption that the elasticity of trade with respect to wages and trade costs
is the same, and equal to ε, the shares of expenditures on goods from country i in country
j in the initial and new equilibrium, respectively, are given by
ε
χij · Ni · wi τij
λij = (6.20)
∑in0 =1 χi0 j Ni0 · wi0 τi0 j
ε
where χij is some parameter of the model, other than τij (e.g. bilateral fixed marketing
costs). Thus, for example, ε = −θ in the Eaton and Kortum (2002) model whereas ε =
− (σ − 1) in the Armington (1969) setup. Notice that an essential simplifying assumption
is that Ni is a constant and does not depend on technology or bilateral trade costs.
Combining this observation with the above two equations we obtain
ε
ŵi τ̂ij
λ̂ij = (6.21)
∑in0 =1 λi0 j ŵi0 τ̂i0 j
ε
From the previous expression and the fact that ŵ j τ̂jj = 1 by our choice of numeraire we
have that
1
λ̂ jj =
∑in0 =1 λi0 j
ε
ŵi0 τ̂i0 j .
107
For the models illustrated above, we can use the trade balance condition as argued by
Arkolakis, Costinot, and Rodrı́guez-Clare (2012) so that in the new equilibrium:
n
wi0 Li0 = ∑
0
λij0 w0j L0j =⇒
j =1
n
ŵ j L̂ j w j L j = ∑ ŵ j0 L̂ j0 λ̂ij0 λij0 w j0 L j0 (6.22)
j 0 =1
If population is exogenous, equations (6.21) and (6.22) constitute a system on ŵ j with
the additional normalization of one wage. The equilibrium changes in wages, wi , and
market shares, λij , can be computed given expression (6.21) and (6.22), which completes
the argument.
6.6 Problem Sets
1. Isomorphisms. Define Xij to be the value of trade flows from i to j. Consider the
following generalized gravity equation:
Xij = Kij γi δj , (6.23)
where Kij is a bilateral trade friction, γi is an origin fixed effect, and δj is a destination
fixed effect.
(a) For each of the following trade models, show how equilibrium trade flows can
be expressed as equation (6.23). That is, write down the mapping between the
generalize gravity equation and model fundamentals.
i. Armington model (Anderson ’79).
ii. Monopolistic competition with free entry (Krugman ’80).
108
iii. Perfect competition with heterogeneous technologies (Eaton and Kortum
’02).
iv. Heterogeneous firms (with Pareto distribution) (Melitz ’03 / Chaney ’08).
(b) Suppose we only observe trade flows in the data. Can we empirically distin-
guish between the above models? If not, what other data would you need to
observe in order to distinguish between the models?
2. Uniqueness. Consider the system defined by equations (6.22) and (6.21) the solution
of which is a vector of wage changes ŵi given a set of changes in trade costs, τ̂ij ,
population L̂ j , an elasticity, e, and initial levels for population and wages. Show
what are the conditions on the parameters of that system, χij , τij , and e, so that the
solution of that system,, exists and is unique.
109
Chapter 7
Gains from Trade
7.1 Trade Liberalization and Firm Heterogeneity
There is a common perception that the gains from trade are larger than what quantitative
general-equilibrium models of trade can explain. A recurring goal in the trade literature
has been to find new channels through which such models can generate larger gains.
Recently, authors such as Melitz (2003) have postulated additional gains from the “selec-
tion” effect compared to the extensive margin effect already postulated by Romer (1994).
Arkolakis, Demidova, Klenow, and Rodrı́guez-Clare (2008) show that some of the key
quantitative frameworks in international trade deliver (Krugman, Eaton and Kortum, the
Chaney version of Melitz and Arkolakis) welfare expressions that are closely comparable.
Arkolakis, Costinot, and Rodrı́guez-Clare (2012) show that for a wide class of perfect and
monopolistic competition models of trade welfare gains from trade can be written as a
function of two sufficient statistics: the share of spending that goes to domestic goods, λ jj ,
and the elasticity of trade parameter, ε. Their result imply that changes in welfare can be
110
written as
W b 1/ε
bj = λ (7.1)
jj
7.1.1 Trade Liberalization and Welfare gains (Arkolakis-Costinot-Rodriguez-
Clare)
To understand the intuition for the main result of Arkolakis, Costinot, and Rodrı́guez-
Clare (2012) we start the analysis from the simplest setup, the Armington model. The
model is essentially identical to the model presented in section 3.3 assuming that the
endowment is labor so that the price of the endowment is wage and that there are no
preference shocks so that αij = 1 , ∀i, j.
We will obtain the result for the case of monopolistic competition where we assume that
exporting and importing country wages matter for marketing fixed costs through a Cobb-
1− µ
, µ ∈ [0, 1] . Denoting zij∗ the cutoff productivity determining
µ
Douglas function, f ij wi w j
the entry of firms from country i in country j; Ωij the set of goods that country j buys from
country i can be written as
σ−1 1
µ 1− µ
σ
f ij wi w j
wi τij
σ −1
σ
Ωij = ω ∈ Ω|zij (ω ) > zij∗ ≡ . (7.2)
σ−1 Pj Xj
where we assume that the production of fixed marketing costs f ij is using a mix of domes-
tic and foreign labor with respective shares µ, 1 − µ. In what follows we assume X j = w j L j
which is guaranteed under perfect competition and free entry.
Real wage is given by Wj = w j /Pj , in that model where the price index is
ˆ 1− σ
σ̃wi τij
Pj1−σ = ∑ Ni g (z) dz.
i zij∗ z
111
Taking logs of the real wage and differentiating we obtain a formula for predicting welfare
gains from trade after changes in trade costs,
d ln Ni γij
d ln Wj = d ln w j − ∑ λij d ln wi + d ln τij + + d ln zij∗ (7.3)
i
1−σ 1−σ
where σ −1
zij∗ g zij∗
γij = ´ ,
zij∗ zσ−1 g (z) dz
and
σ−1 1
µ 1− µ
σ
σ
σ −1
wi τij
f ij wi w j
zij∗ ≡ (7.4)
σ−1 Pj Xj
At this point it worths to pause to understand where the gains from trade are coming
from. Notice that in the case of Armington preferences, d ln Ni = 0 and also, effectively,
d ln zij∗ = 0 so that
n
d ln Wj = − ∑ λij d ln wi + d ln τij ,
(7.5)
i =1
i.e. in the Armington model welfare gains from trade arise only because of improvement
in terms of trade. Instead, in the monopolistic competition models, as expression 7.3
reveals there is an additional variety and entry effect. Do this extra terms imply larger
gains from trade? It turns out that under some conditions, the answer is no, and we will
investigate this below.
Using the definition of zij∗ , equation (7.4), in (7.3), the wage normalization, w j = 1, and
rearranging, small changes in real income are now given by
λij γij
d ln Wj = − ∑
1 − σ − γij d ln wi + d ln τij + d ln Ni − µ (d ln wi ) ,
i
1 − σ − γj 1−σ
(7.6)
where γ j ≡ ∑i λij γij .
112
With Dixit-Stiglitz preferences, we get that market shares are given by
´∞ 1− σ σ −1
Ni zij∗ wi τij z gi (z) dz
Xij = ´∞ 1− σ (7.7)
∑in0 =1 Ni0 zσ−1 gi0 (z) dzX j
zi∗0 j wi0 τi0 j
where the density gi (z) of goods with productivity z in Ωij is simply given by the marginal
density of g. Considering the ratio λij /λ jj = Xij /X jj and differentiating and using the def-
inition of zij∗ , equation (7.2),
d ln λij − d ln λ jj = d ln Ni − d ln Nj + (1 − σ) d ln wi + d ln τij − γij d ln zij∗ + γ jj d ln z∗jj
(7.8)
γij
µd ln wi − γij − γ jj d ln z∗jj + d ln Ni − d ln Nj .
= 1 − σ − γij d ln cij +
1−σ
(7.9)
Combining the expression (7.6) and (7.8) reveals that
h
λij i
d ln Wj = − ∑ d ln λij − d ln λ jj + γij − γ jj d ln z∗jj + d ln Nj
i
1 − σ − γj
d ln λ jj − γ j − γ jj d ln z∗jj − d ln Nj
= (7.10)
1 − σ − γj
using the fact that
n n
∑i=1 λij = 1 =⇒ ∑i=1 λij d ln λij = 0.
Three are the macro-level restrictions that are used to derive the result in a general perfect
competition or monopolistically competitive setup, such as the ones studies in Chapters
3.4 and 4. We have already talked about R1 and R3. Below we introduce one more restric-
tion.
The import demand system is CES. The last macro-level restriction is concerned with
113
the partial equilibrium effects of variable trade costs on aggregate trade flows. Define
the import demand system as the mapping from (w, N, τ ) into X ≡ Xij , where w≡ {wi }
is the vector of wages, N = { Ni } is the vector of measures of goods that can be pro-
duced in each country, and τ ≡ τij is the matrix of variable trade costs. This mapping
is determined by utility and profit maximization given preferences, technological con-
straints, and market structure. It excludes, however, labor market clearing conditions
as well as free entry conditions (if any) which determine the equilibrium values of w
and N. The third macro-level restriction imposes restrictions on the partial elasticities,
0
εiij ≡ ∂ ln Xij /X jj ∂ ln τi0 j , of that system:
R4 The import demand system is such that for any importer j and any pair of exporters i 6= j and
0
i0 6= j, εiij = ε < 0 if i = i0 , and zero otherwise.
0
Each elasticity εiij captures the percentage change in the relative imports from country
i in country j associated with a change in the variable trade costs between country i0 and
j holding wages and the measure of goods that can be produced in each country fixed.
. . . .
Noting that ∂ ln zij∗ ∂ ln τij = ∂ ln z∗jj ∂ ln τij − 1 and ∂ ln zij∗ ∂ ln τi0 j = ∂ ln z∗jj ∂ ln τi0 j
if i0 6= i, we can define the import demand system as the following partial derivative,
∗
1 − σ − γij − γij − γ jj ∂ ln z jj
for i0 = i
∂ ln Xij /X jj
0 ∂ ln τij
= εiij = ∗
∂ ln z jj , (7.11)
∂ ln τi0 j
− γij − γ jj
for i 0 6= i
∂ ln τ 0 i j
´∞
where γij ≡ d ln zij∗ z 1− σ g i (z) dz d ln zij∗ .
R4 implies γij = γ jj and 1 − σ − γ j = ε for all i, j, we obtain from (7.10) that d ln Wj =
d ln λ jj − d ln Nj /ε, using the fact that ∑in=1 λij d ln λij = 0. To conclude, we simply note
that free entry and R1 and R3, using the results of Section 5.3, imply d ln Nj = d ln Yj = 0.
Combining the two previous observations and integrating, we finally obtain expression
114
(7.1) which is model invariant, as long as e is chosen to be the same.
Going back to our various derivations in the previous chapters we note that all the
models deliver similar expressions for welfare gains from trade as a function of λii , and
thus the trade share of GDP. In particular, the expressions for the Armington and Krug-
man models in Chapter 3.4 is similar to the one derived in other models with heteroge-
neous firms such as the ones of Eaton and Kortum (2002), the Chaney (2008) version of
Melitz (2003) and Arkolakis (2010b) in Chapter 4. The only difference is that in the latter
cases σ − 1 is replaced by the parameter that determines the heterogeneity of the produc-
tivities of the firms or productivities of sectors.
7.1.2 Ex-ante Gains (Atkeson Burstein 2010)
Before proceeding it is worth discussing in detail an important result from Atkeson and
Burstein (2010). In the Armington model, expression 7.5 can be written under symmetry
d ln Wj = − 1 − λ jj d ln τ.
This expression is quite intuitive. Gains from trade, to a first-order, depend on the per-
centage change in trade costs, and the exposure of the country to trade. It turns out, that a
similar condition can be derived in monopolistic competition under symmetry (see Atke-
son and Burstein (2010)). To see that, you need to differentiate the free entry condition
under µ = 1
ˆ ! σ −1 ˆ
z
∑ f ij
zij∗ zij∗
g(z)dz − ∑ f ij
zij∗
g(z)dz = f e
j j
Differentiation yields
115
´
σ −1
z
f ij zij∗ g(z)dz
zij∗
∑ d ln zij∗ = 0 =⇒
´
σ −1
j
∑ j f ij z∗ zz∗
g(z)dz
ij ij
∑ vij d ln zij∗ = 0
j
where vij is the export share of i.
Using the definition of zij∗
σ σ−1 1
w j wi f ij w j wi
σ σ −1
zij∗ = τij =⇒
σ−1 Pj w j Xj wj
wj
σ wi
d ln zij∗ = d ln + d ln τij + d ln
Pj σ−1 wj
so that
wj
σ wi
∑ vij d ln zij∗ = ∑ ij
v d ln
Pj
+ d ln τij +
σ − 1
d ln
w j
=0
j j
Imposing symmetry (Wj = W, w j = w, f ij = f , τij = τ for i 6= j) and using the fact that
∑ j vij = 1
d ln Wj = − ∑ vij d ln τ = −(1 − v jj )d ln τ
j 6 =i
Next, we show that v jj = λ jj under symmetry. Using the definitions of vij and zij∗ and
imposing symmetry
116
´
σ −1
z
f jj z∗jj z∗jj g(z)dz
v jj =
´
σ −1
z
∑ j f ij z∗ zij∗ g(z)dz
ij
´
(z)σ−1 g(z)dz
z∗jj
=
´
σ −1
z∗
∑ j z∗ii zij∗ z
σ−1 g ( z ) dz
ij
´ σ −1
z∗jj z g(z)dz
= σ −1 ´ ´
∑ j6=i τ1 z∗ z
σ −1 g ( z ) dz +
z∗ z
σ−1 g ( z ) dz
ij ii
Using the definition of λij and imposing symmetry
´ 1−σ zσ−1 g ( z ) dz
Nj z∗jj ( w j )
λ jj = ´
∑i Ni zij∗ ( wi τij )
1−σ zσ−1 g ( z ) dz
´
z∗jj zσ−1 g(z)dz
=
1 σ −1
´ ´ = v jj
∑ j 6 =i zσ−1 g(z)dz + zσ−1 g(z)dz
τ zij∗ zii∗
Thus
d ln Wj = −(1 − λ jj )d ln τ
as noted.
7.1.3 Discussion: gains from trade and the trade heterogeneity
One question that arises from the above analysis is how restrictive is the assumptions R1-
R4 imposed above. The most crucial of those assumptions is assumption R4. By making
this assumption one essentially makes assumption on the distribution of firm productivi-
117
ties in the models with firm heterogeneity, essentially restricting oneselve to models with
Pareto distribution of productivities. Melitz and Redding (2014) have studied the implica-
tions of models with firm-level heterogeneity that violate this assumption. To understand
the role of firm heterogeneity, notice that as pointed out in footnote 17 of the Arkolakis
(2010a), in the general Melitz model (for general distribution and without symmetry) a
small “foreign shock” has an effect on welfare capture by
d ln Ni − d ln λii
d ln Wi = , (7.12)
σ − 1 − γ( ϕii )
´
∞
where γ( ϕ) ≡ −d ln ϕ x σ−1 dGi ( x ) /d ln ϕ, Ni denotes entry in country i, and ϕii are
the domestic productivity cut-offs. 1 With a Pareto distribution (that essentially imposes
assumption R4) γ( ϕ) = σ − 1 − θ and entry is not affected by trade costs (d ln Ni = 0), so
that d ln W = −d ln λ/θ. Whereas the elasticity of trade trade with a Pareto distribution is
θ as discussed above, the next question that naturally arises is what is the elasticity with
general distribution, and how can we measure it in a sensible way to incorporate it to our
analysis.
To do so, we follow Atkeson and Burstein (2010) and Melitz and Redding (2014) and
consider symmetric countries and a symmetric trade liberalization. Letting λ denote the
share of expenditure on domestic goods, the result in the previous subsection implies that
d ln W = (1 − λ)d ln τ.
The elasticity of trade considered above is ε(λ) ≡ d ln((1 − λ)/λ)/d ln τ. Following
Arkolakis, Costinot, Donaldson, and Rodrı́guez-Clare (2012) we can use this definition
1 This definition of γ ( ϕ ) implies that σ − 1 − γ ( ϕ ) is the elasticity of the market share of the firm at the
ii
domestic cut-off with respect to the domestic productivity cut-off in country i.
118
to change variables in the expression for d ln W and obtain
(1 − λ)d ln((1 − λ)/λ) d ln λ
d ln W = =− .
e(λ) ε(λ)
We thus see, that the welfare gains formula holds for small changes even for arbitrary
distributions.
7.2 Global Gains
We use expression (7.6). Set µ = 0 and consider changes in variable trade costs only so
that d ln f ij = 0. Using R1, R3 as above, as well as R4, we have d ln Ni = 0 and thus the
total gains from trade are
n n
∑ Xj d ln Wj = ∑ j=1 Xj ∑ d ln w j − ∑ j=1 Xj ∑ d ln Pj (7.13)
j j j
n n
= ∑ j=1 ∑in=1 X ji d ln w j − ∑ j=1 X j ∑in=1 λij d ln wi + d ln τij
(7.14)
n n
n
= ∑ j=1 ∑in=1 X ji d ln w j − Xij d ln wi − ∑ ∑ Xij d ln τij .
j =1 i =1
The first double summation term in the RHS is zero by splitting it into two terms inter-
changing i and j. Thus, total welfare changes are given by
Xj n n Xj
∑ ∑ 0 Xj0 d ln Wj = − ∑ ∑ λij ∑ 0 Xj0 d ln τij .
j j j =1 i =1 j
This expression for global gains is derived by Atkeson and Burstein (2010); Fan, Lai, and
Qi (2013). It straightforward to note that the relevant expression for the Armington and
the Eaton Kortum model is expression (7.13) so that the same conclusion holds for perfect
119
competition. In fact, the result does not even rely on labor mobility either so that it holds
under a geography environment as well.
7.3 Problem Sets
1. Pareto distribution. Prove that in the Melitz model studied in section (4.5) the Pareto
distribution implies that R1-R4 hold.
120
Chapter 8
Extensions: Modeling the Demand
Side
We now discuss a number of ways that the simple model can be extended by with as-
sumptions that have implications for the effective demand of the firm. We will discuss a
nested CES structure, endogenous marketing costs, multi-product firms, and other pref-
erences structure different than the constant elasticity demand.
8.1 Extension I: The Nested CES demand structure
We can consider a nested CES structure
σ−σ 1
ˆ ! σ−σ 1 ε−ε 1
N ε −1
Ω
∑ xk (ω ) ε dω
k =1
121
that delivers the demand
−ε −σ
pi ( ω ) P (ω )
xi ( ω ) = X,
P (ω ) P
with
" #1/(1−ε)
N
P (ω ) = ∑ pk (ω )1−ε dω ,
k =1
ˆ 1/(1−σ)
1− σ
P= P (ω ) dω .
Ω
and X being the overall spending. Serving the market incurs an entry cost Fi ≥ 0.
Serving a market incurs an entry cost
a) ε → ∞, F = 0 EK, Bertrand BEJK
b) ε = σ, F > 0 monopolistic competition Melitz-Chaney
c) ε > σ, F ≥ 0 (with either F > 0 or ε → ∞) and Cournot, Atkeson and Burstein.
8.2 Extension II: Market penetration costs
The CES benchmark proved extremely useful for many applications. Its main weakness
is in predicting the behavior of small firms-goods as Eaton, Kortum, and Kramarz (2011).
These firms-goods tend to be a very large part of trade in a trade liberalization and as
time evolves. To address this fact, a simple extension presented in Arkolakis (2010b) does
the job by modeling the fixed entry costs as cost of reaching individual consumers into
individual destinations.
Each good is produced by at most a single firm and firms differ ex-ante only in their
productivities z and their country of origin i = 1, ..., N. We denote the destination coun-
122
try by j. The preferences for consumer l are given by the standard symmetric constant
elasticity of substitution (CES) objective function:
ˆ σ−σ 1
σ −1
l
U = x (ω ) σ dω ,
ω ∈Ωl
where σ ∈ (1, +∞) is the elasticity of substitution. When a good produced with a produc-
tivity z from country i is included in the choice set of consumer l, Ωl , the demand of this
consumer is given by,
pij (z)−σ
xij (z) = y j , (8.1)
Pj1−σ
where pij (z) is the price charged in country j, y j the income per capita of the consumer,
and Pj a price aggregate of the goods in the choice set of the consumer. An unrealistic as-
sumption of the CES framework introduced by Dixit and Stiglitz is that all the consumers
have access to the same set of goods Ωl . This formulation departs from the standard
formulation of trade models with CES preferences by proposing a formulation where Ωl
can be different for different consumers. In order to be able to fully characterize the gen-
eral equilibrium of the model, we assume that consumers are reached independently by
different firms and that each firm pays a cost to reach a fraction n of the consumers. In
equilibrium, all firms z from country i will reach the same fraction of consumers in coun-
try j and thus their ‘effective’ sales will be:1
pij (z)1−σ
tij (z) = nij (z) L j y j
| {z } Pj1−σ
consumers | {z }
reached in j
sales per-consumer
where L j is the measure of the population of country j. In order to give foundations to
1 Given the existence of a continuum of firms and consumers I am making use of the Law of Large Num-
bers. This implies that nij (z) from a probability becomes a fraction. The application of the Law of Large
Numbers also implies that Pj is now a function of nij (z)’s and has a given value for all consumers.
123
the market penetration cost function as an explicit function of nij (z) we depart from the
standard formulation where there is a uniform fixed marketing cost to enter the market
and sell to all the consumers there. Instead, we consider an alternative formulation that
intends to broadly capture the marketing costs incurred by the firm in order to increase
their sales in a particular market. The marketing costs are modeled as increasing access
costs that the firms pay in order to access an increasing number of customers in each given
country. Due to market saturation, reaching additional consumers becomes increasingly
difficult once a relatively large fraction of them has already been reached. Based on a
derivation of a marketing technology from first principles the cost function of reaching a
fraction n of a population of L consumers in Arkolakis (2010b) is derived to be
1− β
Lα 1−(1−n)
if β ∈ [0, 1) ∪ (1, +∞)
ψ 1− β
f (n) = .
Lα
log (1 − n) if β = 1
ψ
1/ψ denotes the productivity of search effort and a ∈ [0, 1] regulates returns to scale of
marketing costs with respect to the population size of the destination country. The pa-
rameter β determines how steeply the cost to reach additional consumers is rising. How-
ever, for any parametrization of β the marginal cost to reach the very first consumers in a
given market j is always positive (the derivative is always bigger than zero). Thus, only
firms with productivity above some threshold zij∗ will have high enough revenues from
the very first consumers to find it profitable to enter the market.2 The case where β = 1
corresponds to the benchmark random search case of Butters (1977) and Grossman and
Shapiro (1984). If β = 0 the function implies a linear cost to reach additional consumers,
which in turn is isomorphic to the case of Melitz (2003)-Chaney (2008) given that firms
reach either all the consumers in a market or none.
2 With no additional heterogeneity across firms this implies a hierarchy of exporting destinations
124
The production side of the firm is standard. Labor is the only factor of production. The
firm z uses a production function that exhibits constant returns to scale and productivity
z. It incurs an iceberg transportation cost τij to ship a good from country i to country
j. This implies that the optimal price of the firm is a constant markup σ/(σ − 1) over
the unit cost of producing and shipping the good, w j τij /z. The equilibrium of the model
retains many of the desirable properties of the benchmark quantitative framework for
considering bilateral trade flows develop by Eaton and Kortum (2002) and particularly the
gravity structure. It also allows for endogenous decision of exporting and non-exporting
of firms as in Melitz (2003).
How can this additional feature of endogenous market penetration costs help the
model to address facts on exporters? The following version of the proposition proved in
Arkolakis (2010b) computes the responses of firm’s sales in a trade liberalization episode:
Proposition 4. [Elasticity of trade flows and firm size]
The partial elasticity of a firm’s sales in market j with respect to variable trade costs, ε ij (z) =
∂ ln tij (z) /∂ ln τij , is decreasing with firm productivity, z, i.e. dε ij (z) /dz < 0 for all z ≥ z∗ .
ij
Proof. Compute the partial elasticity of trade flows tij (z) with respect to a change in τij ,
namely ∂ ln tij (z) /∂ ln τij = |ζ (z)| × ∂ ln zij∗ /∂ ln τij , where
!(σ−1)/β −1
σ−1 z
ζ (z) = ( σ − 1) + − 1 .
| {z } β zij∗
intensive margin | {z }
of per-consumer extensive margin of
sales elasticity consumers elasticity
Notice that ζ (z) ≥ 0 for z ≥ zij∗ . ζ (z) is also decreasing in z and thus decreasing in initial
export sales. In fact, as β → 0 then ζ (z) → (σ − 1) for all z ≥ zij∗ .
The proposition implies that trade liberalization benefits relatively more the smaller
125
exporters in a market. The parameter β governs both the heterogeneity of exporters cross-
sectional sales and also the heterogeneity of the growth rates of sales after a trade liberal-
ization.
8.3 Extension III: Multiproduct firms
We now turn to an extension of the basic CES setup that can accomodate multiproduct
firms. This extension is suggested by Arkolakis and Muendler (2010) and is modeling
the idea of “core-competency” within the standard heterogeneous firms setup of Melitz
(2003).
A conventional “variety” offered by a firm ω from source country i to destination j is
the product composite
σ−σ1
Gij (ω )
σ−1
xij (ω ) ≡ ∑ xijg (ω ) σ ,
g =1
where Gij (ω ) is the number of products that firm ω sells in country d and xijg (ω ) is the
quantity of product g that consumers consume. The consumer’s utility at destination j is
a CES aggregation over these bundles
N ˆ ! σ−σ1
σ−1
Uj = ∑ xij (ω )
ω ∈Ωij
σ dω for σ > 1, (8.2)
i =1
where Ωij is the set of firms that ship from source country i to destination j. For simplicity
we assume that the elasticity of substitution across a firm’s products is the same as the
elasticity of substitution between varieties of different firms.3
3 Arkolakis and Muendler (2010) generalize the model to consumer preferences with two nests. The inner
nest contains the products of a firm, which are substitutes with an elasticity of ε. The outer nest aggregates
those firm-level product lines over firms and source countries, where the product lines are substitutes with
126
The consumer’s first-order conditions of utility maximization imply a product de-
mand −σ
pijg (ω )
xijg (ω ) = Xj , (8.3)
Pj1−σ
where pijg is the price of variety ω product g in market j and we denote by X j the total
spending of consumers in country j. The corresponding price index is defined as
− σ−11
ˆ
N Gkj (ω )
Pj ≡ ∑ ∑ pkjg (ω )−(σ−1) dω . (8.4)
k =1 ω ∈Ωkj g=1
A firm of type z chooses the number of products Gij (z) to sell to a given market j.
The firm makes each product g ∈ 1, 2, . . . , Gij (z) with a linear production technology,
employing local labor with efficiency z g . When exported, a product incurs a standard
iceberg trade cost so that τij > 1 units must be shipped from i for one unit to arrive at
destination j. We normalize τii = 1 for domestic sales. Note that this iceberg trade cost is
common to all firms and to all firm-products shipping from i to j.
Without loss of generality we order each firm’s products in terms of their efficiency so
that z1 ≥ z2 ≥ . . . ≥ zGij . A firm will enter export market j with the most efficient product
first and then expand its scope moving up the marginal-cost ladder product by product.
Under this convention we write the efficiency of the g-th product of a firm z as
z
zg ≡ with h0 ( g) > 0. (8.5)
h( g)
We normalize h(1) = 1 so that z1 = z. We think of the function h( g) : [0, +∞) → [1, +∞)
as a continuous and differentiable function but we will consider its values at discrete
points g = 1, 2, . . . , Gij as appropriate.
a different elasticity σ 6= ε. general case of ε 6= σ generates similar predictions at the firm-level and at the
aggregate bilateral country level.
127
Related to the marginal-cost schedule h( g) we define firm z’s product efficiency index
as
Gij
!− σ−11
H ( Gij ) ≡ ∑ h( g)−(σ−1) . (8.6)
g =1
This efficiency index will play an important role in the firm’s optimality conditions for
scope choice.
As the firm widens its exporter scope, it also faces a product-destination specific in-
cremental local entry cost f ij ( g) that is zero at zero scope and strictly positive otherwise:
f ij (0) = 0 and f ij ( g) > 0 for all g = 1, 2, . . . , Gij , (8.7)
where f ij ( g) is a continuous function in [1, +∞).
The incremental local entry cost f ij ( g) accommodates fixed costs of marketing (e.g.
with 0 < f ii ( g) < f ij ( g)). In a market, the incremental local entry costs f ij ( g) may increase
or decrease with exporter scope. But a firm’s local entry costs
Gij
∑
Fij Gij = f ij ( g)
g =1
necessarily increase with exporter scope Gij in country j because f ij ( g) > 0. We assume
that the incremental local entry costs f ij ( g) are paid in terms of importer (destination
country) wages so that Fij ( Gij ) is homogeneous of degree one in w j . Combined with the
preceding varying firm-product efficiencies, this local entry cost structure allows us to
endogenize the exporter scope choice at each destination j.
A firm with a productivity z from country i faces the following optimization problem
for selling to destination market j
Gij −σ
pijg
wi
πij (z) = max ∑
Gij ,pijg g=1
pijg − τij
z/h( g) Pj1−σ
X j − Fij ( Gij ).
128
The firm’s first-order conditions with respect to individual prices pijg imply product prices
pijg (z) = m̄ τij wi h( g)/z (8.8)
with an identical markup over marginal cost m̄ ≡ σ/(σ −1) > 1 for σ > 1. A firm’s choice
of optimal prices implies optimal product sales for product g
σ−1
Pj
z
pijg (z) xijg (z) = Xj . (8.9)
m̄ τij wi h( g)
Summing (8.9) over the firm’s products at destination j, firm z’s optimal total exports to
destination j are
Gij (z) σ−1
Pj
tij (z) = ∑ pijg (z) xijg (z) =
m̄ τij wi
z X j H ( Gij (z))−(σ−1) , (8.10)
g =1
where H ( Gij ) is a firm’s product efficiency index from (8.6). Expression (8.10) reveals
that firm sales in country j are strictly increasing in productivity z given that the term
H ( Gij (z))−(σ−1) weakly increases in Gij (z) and Gij (z) weakly increasing in z.
Given constant markups over marginal cost, profits at a destination j for a firm z sell-
ing Gij are
σ−1
Pj Xj
−(σ−1)
πij (z) = z H Gij − Fij Gij .
m̄ τij wi σ
The following assumption is required for the firm optimization to be well defined:
f˜ij0 ( G ) > 0 (8.11)
where f˜ij ( G ) ≡ f ij ( G ) h( G )σ−1
Under this assumption, the optimal choice for Gij (z) is the largest G ∈ {0, 1, . . .} such
that operating profits from that product equal (or still exceed) the incremental local entry
129
costs:
σ−1
Pj Xj
z
τij wi ≥ f ij ( G ) ⇐⇒
m̄ h( G ) σ
σ−1
Pj z Xj
g =1
πij (z) ≡ ≥ f ij ( G ) h( G )σ−1 ≡ f˜ij ( G ). (8.12)
m̄ τij wi σ
g =1
Operating profits from the core product are πij (z), and operating profits from each ad-
g =1
ditional product g are πij (z)/h( g)σ−1 .
Assumption 8.11 is comparable to a second-order condition (for perfectly divisible
scope in the continuum version of the model, Assumption 8.11 is equivalent to the sec-
ond order condition). When Assumption 8.11 holds we will say that a firm faces overall
diseconomies of scope.
We can express the condition for optimal scope more intuitively and evaluate the op-
timal scope of different firms. Firm z exports from i to j iff πij (z) ≥ 0. At the break-even
point πij (z) = 0, the firm is indifferent between selling its first product to market j and
remaining absent. Equivalently, reformulating the break-even condition and using the
above expression for minimum profitable scope, the productivity threshold zij∗ for export-
ing from i to j is given by
σ−1
σ f ij (1) m̄ τij wi
σ−1
zij∗ ≡ . (8.13)
Xj Pj
In general, using (8.13), we can define the productivity threshold zij∗,G such that firms
with z ≥ zij∗,G sell at least Gij products as
σ−1 f˜ij ( G ) ∗ σ−1
zij∗,G = zij , (8.14)
f ij (1)
under the convention that zij∗ ≡ zij∗,1 . Note that if Assumption 8.11 holds then zij∗ < zij∗,2 <
zij∗,3 < . . . so that more productive firms introduce more products in a given market. So
130
Gij (z) is a step-function that weakly increases in z.
Using the above definitions, we can rewrite individual product sales (8.9) and total
sales (8.10) as
!σ−1
z
pijg (z) xijg (z) = σ f ij (1) h( g)−(σ−1)
zij∗
!σ−1
z
= σ f˜ij Gij (z) h( g)−(σ−1)
(8.15)
zij∗,G
and !σ−1
z −(σ−1)
tij (z) = σ f ij (1) H Gij (z) . (8.16)
zij∗
The following proposition summarizes the findings.
Proposition 5. If Assumption 8.11 holds, then for all i, j ∈ {1, . . . , N }
• exporter scope Gij (z) is positive and weakly increases in z for z ≥ zij∗ ;
• total firm exports tij (z) are positive and strictly increase in z for z ≥ zij∗ .
Proof. The first statement follows directly from the discussion above. The second state-
ment follows because H ( Gij (z))−(σ−1) strictly increases in Gij (z) and Gij (z) weakly in-
creases in z so that tij (z) strictly increases in z by (8.16).
There are two key differences to the Melitz (2003) setup. The first is the term H ( Gij (z))−(σ−1)
that reflects multi-product choice within the firm. Adding new products make this term
higher, but with core-competency these new products are less and less important for over-
all sales. The second difference with the Melitz setup is the fixed cost term Fij Gij that
jointly with H determines the products optimization. These two features properly esti-
mated from the data can be used to evaluate the prediction of this setup for a number of
facts on multi-product exporters. We will come back to this point when we talk about the
estimation of firm-level models.
131
8.3.1 Gravity and Welfare
The market shares in this model are given by
Ni Ai (wi τij )−θ f ij (1)−θ̃ F̄ij
λij =
∑k Nk Ak (wk τkj )−θ f kj (1)−θ̃ F̄kj
where f ij (1)−θ̃ F̄ij = ∑∞
G =1 f ij ( G )
−(θ̃−1) h ( G )−θ and θ̃ = θ
σ −1 . The key new insight is that
changes in the entry cost will have a different effect on overall trade than in the Melitz
(2003) setup insofar they affect the entry costs for different products differently. Condi-
tional on overall trade flows though, the welfare gains from trade are given by an expres-
sion that is similar to the Melitz (2003) setup. Thus, the difference is in the counterfactual
predictions with respect to changes in trade costs.
8.4 Extension IV: General Symmetric Separable Utility Function
8.4.1 Monopolistic Competition with Homogeneous Firms (Krugman 79)
We now retain the monopolistic competition structure and all the notation from the pre-
vious section but consider a general symmetric separable utility function as in Krugman
(1979).
8.4.1.1 Consumer’s problem
The problem of the representative consumer from country j is
N ˆ
!
max ∑
u x j (ω ) dω ,
i =1 Ωi
N ˆ
s.t. ∑ Ωi
p j (ω ) x j (ω ) dω = w j ,
i =1
132
with u0 > 0 and u00 < 0. We assume a particular regularity condition on the utility
function that will allow us to focus on the empirically relevant cases, and in particular
x j (ω ) u00 x j (ω )
∂ ln u0 ( x )
− =− >0. (8.17)
u0 x j (ω )
∂ ln x
This condition implies
u0 x j (ω ) = λ j p j (ω )
(8.18)
where λ j is the Lagrange multiplier of the consumer in country j. The demand function
implied by this solution is given by
x j p j (ω ) = u0−1 λ j p j (ω )
(8.19)
We focus on demand functions that have the choke price property, i.e. there exists a p∗j so
that given λ j x j p j = 0 for all p ≥ p∗j . It is straightforward to show that this property
requires u0 (0) < +∞ and that p∗j = u0 (0) /λ j and thus
! !
p j (ω ) p j (ω )
xj p j , p∗j = xj =u 0−1 0
u (0) .
p∗j p∗j
8.4.1.2 Firm’s Problem
We can now incorporate this demand as a constraint to the firm’s problem. In particular,
a firm from country i with productivity z = zi choses price in country j to maximize
wi 0−1
πij (z) = p (ω ) − τij u λ j p (ω ) (8.20)
z
133
The first order condition of this problem is given by (making use of the inverse function
theorem and of expressions 8.18 and 8.19)
wi 0−1 0
u0−1 λ j p j (ω ) + p (ω ) − τij u λ j p (ω ) λ j = 0 =⇒
z
1 τij wi
p= ,
1 + ũ x j (ω ) z
where
x j (ω ) u00 x j (ω )
ũ x j (ω ) =
u0 x j (ω )
or alternatively we can simply write the price as a function of the elasticity of demand
− dd ln
ln x
p τij wi
p=
− dd ln
ln x
p −1 z
and the markup can be written as
− dd ln
ln x
!
p j (ω ) p
µ = (8.21)
p∗j − dd ln
ln x
p −1
A number of important points can be made for this expression.
First, notice that expression (8.21) depends on the demand elasticity. Thus, how the
markup changes with firm size depends on how the demand elasticity changes with size.
By simply inspecting the derivative of the markup function it is obvious that if demand
is log-convave, i.e. d ln2 x/ (d ln p)2 < 0, then markup increase with firm size and the
opposite for log-convex demand.
Second, notice that the degree of pass-through depends on how markups change with
marginal cost changes, for different levels of demand. This effectively requires taking one
more derivative of the markup function and the result ultimately depends also on the
third derivative of the demand function.
Finally, notice that the second order conditions of the optimization problem (8.20) are
134
d2 ln x
always satisfied if < 0. Now define zij∗ ≡ wi τij /p∗j then we can finally express
d(ln p)2
!
z
xj = x ,
zij∗
and !
z
µj = µ .
zij∗
In this environment we can write bilateral trade shares as
τij wi
Ni zi µ zz∗i q zz∗i
ij ij
λij = .
τkj wk z z
∑k Nk zk µ z∗k q z∗k
kj kj
Define vij as z/zij∗
µ(vij )
Ni wi τij µ vij − vij
zi q
λij =
µ(vij )
∑k Nk wk τkj
zk q µ vkj − vkj
Welfare under homogeneity (Neary-Mazrova)
Analytical expression for the welfare gains from trade are challenging to derive in this
case. We will instead work in the case of two symmetric countries around the free trade
equilibrium. The change in equivalent expenditure, e, required to keep the consumer is
written as
N ∗ p∗
I Np I I
V N, N ∗ , p, p∗ , = f Nu + N ∗
u
Y ( N p + N ∗ p∗ ) N pe ( N p + N ∗ p∗ ) N ∗ p∗ e
I
= f (N + N∗) u
( N p + N ∗ p∗ ) e
where f is some function p is the price of the domestic good and p∗ the price of the
foreign good and the two parts inside the indirect utility function represent the relative
utility obtain from domestic and foreign goods. We can set this expression equal to a
constant, the targeted level of utility, totally differentiate, divide by N + N ∗ and solve for
135
the equivalent expenditure to obtain4
N Np u0 I
( N p + N ∗ p ∗ )Y
I
( N p+ N ∗ p∗ )e
N̂ + − N̂ + p̂ +
∗ N p + N ∗ p∗
N+N u I
( N p+ N ∗ p∗ )e
∗ ∗ ∗ u 0 I I
u0 I I
N N p − N p e
∗ ∗ N ∗ p∗ e ( N p+ N ∗ p∗ )e ( N p+ N ∗ p∗ )e
N̂ ∗ ∗ ∗
∗
+ ∗ ∗
N̂ + p̂ = ê
N+N Np + N p u I
u I
( N p+ N ∗ p∗ )e ( N p+ N ∗ p∗ )e
( 1 − λ N ) + ( 1 − λY ) ξ
λ N + λY ξ ∗
N̂ + N̂ − {λY [ p̂] + (1 − λY ) [ p̂∗ ]} = ê (8.22)
ξ ξ
where
u0 I
N ∗ p∗ e
I
N ∗ p∗ e
ξ=
I
u ( N p+ N ∗ p∗ )e
and
Np
λY =
N p + N ∗ p∗
Can we express this as a function of trade. Notice that domestic trade shares are given
by
N pq ( p)
λ= =⇒
( N pq ( p) + N ∗ p∗ q∗ ( p∗ ))
λ = N̂ + p̂ + ε p̂ − N̂ + p̂ + ε p̂ λ + N̂ ∗ + p̂∗ + ε∗ p̂∗ (1 − λ) =⇒
b
λ = N̂ + p̂ (1 + ε) (1 − λ) − N̂ ∗ + p̂∗ (1 + ε∗ ) (1 − λ)
b
4
Np u0 ( N p+λI ∗ ∗
N p )e ( N p+ N p )e
I
∗ ∗
− ( N + N ∗ )
N̂N + N̂ + p̂ + ê +
N p + N ∗ p∗
I
u ( N p+ N ∗ p∗ )e
∗ p∗ 0 I
u N ∗ p∗ e N ∗Ip∗ e
N
N̂ ∗ N ∗ + − ( N + N ∗ ) N̂ ∗ + p̂∗ + ê = 0
N p + N ∗ p∗ u I
( N p+ N ∗ p∗ )e
136
where ε = q0 p/q. Around free trade and symmetry λ = 1 − λ, ε = ε∗
λ − N̂ − N̂ ∗ (1 − λ) = (1 − λ) (1 + ε) ( p̂ − p̂∗ ) =⇒
b
N̂ − N̂ ∗
λ
= ( p̂ − p̂∗ )
b
−
(1 − λ ) (1 + ε ) (1 + ε )
and replacing in (8.22) (using also that around free trade λY = 1 − λY )
( ∗
)
− ( 1 − ) N̂ − N̂
λ N + λ Y ξ ( 1 λ Y ) λ λ Y
ê = N̂ + N̂ ∗
b
− −
ξ (1 − λ ) (1 + ε ) (1 + ε )
1 − λY
∗
λ N + λY ξ λ
b
= N̂ + N̂ + + − p̂
ξ 1+ε (1 + ε )
and under initial free trade and symmetry it is also λY = λ.
8.4.1.3 Monopolistic Competition with Heterogeneous Firms (Arkolakis-Costinot-Donaldson-
Rodriguez Clare)
We now retain the monopolistic competition structure with separable utility function but
we consider heterogeneous firms following Arkolakis, Costinot, Donaldson, and Rodrı́guez-
Clare (2012). The analysis here covers utility functions considered by Behrens and Murata
(2009), Behrens, Mion, Murata, and Sudekum (2009), Saure (2009), Simonovska (2009),
Dhingra and Morrow (2012) and Zhelobodko, Kokovin, Parenti, and Thisse (2011).
Demand can again be written as
xij (ω ) = u0−1 λ j pij (ω )
(8.23)
where λ j is the Lagrange multiplier of the consumer. As long as u0 (0) < ∞ we can define
a “choke-up” price p∗j = u0 (0) /λ j so that if pij (ω ) = pij∗ , xij ( p) = 0. Given that the
distribution we use is continuous and with unbounded support there will be always a
firm from each i offering a price low enough to sell to all the markets. We maintain the
restrictions made in Section 8.4.1.
137
8.4.1.4 Firm Problem
The firm problem is the same as in the homogeneity case. We have the price choice of
τw
the firm to be pij (z) = µ pij (z) ijz i as indicated above. It can be shown that given the
assumptions for u the markup is increasing in z and 0 for z = zij∗ where
τij wi
zij∗ = (8.24)
p∗j
so that price can be expressed as
τ w
ij i
pij (z) = µ z/zij∗ . (8.25)
z
For the cross-section of firms we can offer a characterization of how the markup changes
with changes in productivity using the properties discussed above. In particular, when
demand is log-concave, d ln2 x/ (d ln p)2 < 0, markups are increase on firm relative size.
Since all the papers discussed above consider the case of the log-concave demand we
will proceed under this assumption as our benchmark in our analysis below.
8.4.1.5 Gravity
Using the expression for firm demand, equation (8.23), and firm prices (8.25), firms sales
can now be written as
τ w
ij i 0−1
tij (z) = µ z/zij∗ u µ z/zij∗ z/zij∗ . (8.26)
z
138
Using this expression we can aggregrate across firms to compute average sales of firms
from i in country j, given by
ˆ
θ
τ w zij∗
ij i 0−1
X̄ij = µ z/zij∗ u µ z/zij∗ zij∗ /z dz =⇒
zij∗ z z θ +1
ˆ
θ
τ w z∗ zij∗
ij i ij 0−1
= µ z/zij∗ u µ z/zij∗ zij∗ /z dz
zij∗ zij∗ z z θ +1
and using a standard change of variables argument and the definition of zij∗ ,
ˆ θ +1
1 0−1 µ (ν) 1
X̄ij = p∗j u dν , (8.27)
1 ν ν ν
i.e. average sales are source independent.
Therefore, trade shares are given by
Ni Aiθ X̄ij
zij∗
λij = Ak
∑k Nk θ X̄kj
z∗kj
−θ
Ni wi τij
= −θ (8.28)
∑k Nk wk τkj
the standard formula for bilateral trade shares. Ni is the measure of entrants. Arko-
lakis, Costinot, Donaldson, and Rodrı́guez-Clare (2012) that this number is independent
of trade in this model, if there is a free entry condition, so in the interest of space, we will
assume that dNi = 0 in the derivations below.
8.4.1.6 Welfare
Because the demand is non-homothetic to characterize welfare we need to explicitly solve
for the expenditure function rather than simply computing the real wage. Let e j ≡ e( p j , u j )
denote the expenditure function of a representative consumer in country j facing a vector
139
of prices p j and let u j be the utility level of such a consumer at the initial equilibrium. By
Shephard’s lemma, we know that de j /dpω,j = q( pω,j , p∗j , w j ) for all ω ∈ Ω. Since all price
changes associated with a change in trade costs are infinitesimal, we can therefore express
the associated change in expenditure as5
ˆ
de j = qω,j dpω,j dω,
ω ∈Ω
where dpω,j is the change in the price of good ω in country j caused by the change in trade
costs. The previous expression can be rearranged in logs as
ˆ h i
de j = λ( pω,j , p∗j , w j )d ln pω,j dω,
ω ∈Ω
where λ( pω,j , p∗j , w j ) ≡ pω,j qω,j /e j is the share of expenditure on good ω in country j in
the initial equilibrium. Using equation (8.26) and the fact that firms from country i only
sell in country j if z ≥ zij∗ , we obtain
ˆ ∞
d ln e j = ∑ zij∗
λij (z) d ln wi τij + d ln µ(z/zij∗ ) dGi (z)
i
where
wi τij
Ni µ z/zij∗ z q ij z/z ∗
ij
λi ( z ) = ´ w 0τ0
∑i Ni0 µ z/zi∗0 j ∗
i i j
z q i 0 j z/z 0
i j dGi ( z )
Using this expression with equation (8.28) we obtain, after simplifications,
ˆ ∞ ˆ ∞
wi τij
Ni µ z/zij∗ z q ij z/z ∗
ij d ln µ(z/zij∗ ) ∗
d ln e j = ∑ λij d ln wi τij − ∑ ´ wi0 τi0 j
dzij dGi (z)
i zij∗ i zij∗ ∑ N 0 µ z/z ∗ q 0 z/z ∗ dG ( z ) d zij∗
i i 0
ij z i j 0
ij i
5 In principle, price changes may not be infinitesimal because of the creation of “new” goods or the de-
struction of “old” ones. This may happen for two reasons: (i ) a change in the number of entrants N or (ii ) a
change in the productivity cut-off z∗ . Since the number of entrants is independent of trade costs, as argued
above, (i ) is never an issue. Since the price of goods at the productivity cut-off is equal to the choke price,
(ii ) is never an issue either. This would not be true under CES utility functions. In this case, changes in pro-
ductivity cut-offs are associated with non-infinitesimal changes in prices since goods at the margin go from a
finite (selling) price to an (infinite) reservation price, or vice versa.
140
or
ˆ ∞
d ln e j = ∑ zij∗
λij d ln wi τij − ∑ ρd ln zij∗ dGi (z)
i i
where
wτ
d ln µ(v) Ni µ (v) iz ij qij (v) v−θ −1
ρ= ´
d ln (v) ∑i Ni0 µ (v) wi0 τi0 j qi0 j (v) v−θ −1 dv
z
is a weighted average of the markup elasticities µ0 (v) across all firms, where zij∗ is defined
in expression (8.24).
Notice that the first term is a standard term and represents gains from trade from
marginal costs reductions, but movements in markups have direct effects (negative im-
pact to welfare from exporters raising markups) and GE implications (potentially positive
effects on welfare because domestic producers lower their markups).
Finally, we can use the labor market clearing and the expression for total bilateral sales
to obtain and differentiating
d ln e j = ∑ λij d ln cij − ρ ∑ λij d ln cij + ρd ln p∗j
i i
θ
d ln p∗j = ∑ λij d ln
wi τij (8.29)
1+θ i
At this point, notice that from the gravity equation, ∑i λij d ln wi τij is equal to d ln λ jj /θ.
Putting everything together we have
ρ
d ln e j = 1− d ln λ jj /θ
1+θ
These derivations imply that under standard restrictions on consumer demand and the
distribution of firm productivity, gains from trade liberalization are weakly lower than
those predicted by the models with constant markups considered in ACR. In particular
According to this result, if η < 0, then conditional on matching the same macro data,
models with variable markups will predict larger welfare gains and conversely if η > 0. It
141
worth pointing out, that if markups are constant then ρ = 0 and we go back to the original
formula that is true for the models with CES demand (see Chapter 7).
8.5 Problem Sets
1. Monopolistic Competition with non-CES preferences.
Consider a consumer that has wealth W, consumes from a set of differentiated vari-
eties ω ∈ Ω, and solves the following maximization problem:
ˆ ˆ
max u (q (ω )) dω s.t. p (ω ) q (ω ) ≤ W,
{q(ω )} Ω Ω
Firms are monopolistic competitors and each firm has a productivity z drawn from
a country specific distribution Gi (z) . Firms from country i face an iceberg trade cost
to sell to a destination j.
(a) Derive the first order conditions of the firm. What are the conditions on the
utility function so that the problem of the firm is well defined (i.e. there exists
a unique solution for the firm markup at each level of firm productivity)?
(b) Discuss the pricing-to-market implications of this model. How do they depend
on firm market shares in a given market?
(c) Assume that each country has labor endowment L j and that labor markets
clear. Define a monopolistically competitive equilibrium in this setup.
(d) Solve for the welfare gains from trade, following Arkolakis, Costinot, Donald-
son, Rodriguez-Clare (but without assuming a Pareto distribution).
(e) (optional) Write a program that computes the equilibrium for any level of the
parameters in this model and computes the welfare.
142
(f) Compare the gains from trade in this model to the gains from Bernard, Eaton,
Jensen, and Kortum model.
2. Nested CES. Consider a Nested CES model, a la Atkeson-Burstein 2008, American
Economic Review. There is a continuum of sectors, s ∈ [0, 1], and n (s) is the number
of potential entrants in sector s for each country. Final output is given by
ˆ 1
! σ−σ 1
σ −1
Y= y (s) σ ds ,
0
! ε−ε 1
ε −1
y (s) = ∑ yi ( s ) ε ds ,
i
where ε > σ > 1. Assume Cournot competition among the n (s) competitors of each
industry.
(a) Derive the optimal demand of the consumer for each firm.
(b) Derive the optimal firm markup as a function of firm market shares. How are
the parameters ε and σ affect the result?
(c) Discuss the pricing-to-market implications of this model. How do they depend
on firm market shares?
(d) Discuss the limit cases when there are many firms in a sector (n (s) → ∞) or
when there is only a monopolist (n (s) = 1). (Hint: the first one is tricky).
(e) Pick a few parameter combinations for ε and σ and simulate the effect of a
market change from a monopoly to a duopoly. What happens to markups and
firm sales?
(f) Now redo i)-ii) with Bertrand competition. Show that there is no discontinuity
as ε → ∞. Which is the model that arises in that case?
143
(g) Discuss also the limit case e = σ. Which model arises in that case?
144
Chapter 9
Modeling Vertical Production
Linkages
In this section we will present some straightforward ways of introducing intermediate
inputs into the heterogeneous firms models. We will comment on the different ways in
which the production theory developed above can be used.
9.1 Each good is both final and intermediate
In their heterogeneous sectors framework Eaton and Kortum (2002) have used the inter-
mediate inputs structure initially proposed by Krugman and Venables (1995). The idea is
that the production of each good requires labor and intermediate inputs, with labor hav-
ing a constant share ι. Intermediates comprise the full set of goods that are also used as
finals and they are combined according to the same CES aggregator. Therefore, the overall
price index in country i, Pi (derived in previous sections), becomes the appropriate index
of intermediate goods prices in this case. The cost of an input bundle in country i is thus
ci = wiι Pi1−ι (9.1)
145
The overall changes in the predictions of the model are small, but the main effect is that
trade shares are now affected by ι and thus
−θ
Ai τij−θ wiι Pi1−ι
λij = −θ .
N −θ ι P 1− ι
∑k=1 k kj
A τ w k k
9.2 Each good has a single specialized intermediate input
Yi (2003) develops a model where endogenous vertical specialization into different stages
of production is allowed. The output y2 (ω ) for a final good ω ∈ Ω is produced using in-
put from a uniquely specialized intermediate good y1 (ω ). The corresponding production
functions are
y2i (ω ) = z2i (ω ) li2 (ω )ι y1i (ω )1−ι , i = 1, 2
y1i (ω ) = z1i (ω ) li1 (ω ) , i = 1, 2
where the output of each one of the stages can be produced by either countries and 1 − ι
is the share of intermediates into production. The model is essentially a two stages Dorn-
busch, Fischer, and Samuelson (1977) model with Perfect Competition in all the markets.
The interesting feature of the Yi (2003) model is that the degree of specialization in either
stage of production for a given country is endogenous and depends on trade barriers and
the comparative advantage of the two countries.1 When for a given good both stages of
the production are performed abroad, trade of that good is more sensitive to trade cost
changes. Yi (2003) uses this feature of the model to offer an explanation of the rapid
growth of world trade during the past decades.
1 While in the Eaton and Kortum (2002) intermediates good framework there are two possible production
patterns for the good that is sold in a given market (either home or foreign is the producer of the sold good)
in the model of Yi (2003) there are 4 for the two stages of a given variety. These are (HH) Home (country)
produces stages 1 and 2, (FF) Foreign produces stages 1 and 2, (HF) Home produces stage 1, Foreign produces
stage 2 and (FH) Foreign produces stage 1, Home produces stage 2.
146
The main drawback of his approach is that calibration is constrained by the usage of
the Dornbusch, Fischer, and Samuelson (1977) framework. Thus, Yi (2003) can use gen-
eral monotonic functions for the relative productivity of one of the stages of production
between the two countries but not of both. Of course, this setup is very difficult to be
generalized in more than two countries.
9.3 Each good uses a continuum of inputs
Arkolakis and Ramanarayanan (2008) propose a different intermediate inputs structure by
merging and generalizing the two approaches described above. Goods are produced in
two stages with the second stage of production (production of “final goods”) using goods
produced in the first stage (“intermediate goods”). Production is vertically specialized to
the extent that one country uses imported intermediate goods to produce output that is
exported. There is a continuum of measure one of goods in the first stage of production,
and in the second stage of production. We index both intermediate and final goods by ω,
although they are distinct commodities.
Each first-stage intermediate input ω can be produced with a CRS labor only technol-
ogy given by
y1i (ω ) = z1i (ω ) li1 (ω ) , (9.2)
with efficiency denoted by z1i (ω ). The technology for producing output of final good ω
is:2
ˆ (1σ−−ι1)σ
ι 0
σ−σ 1 0
y2i (ω ) = z2i (ω ) li2 mi ω, ω dω , (9.3)
where mi (ω, ω 0 ) is the use of intermediate good ω 0 in the production of final good ω. The
parameter σ is the elasticity of substitution between different intermediate inputs.
We use the probabilistic representation of Eaton and Kortum (2002) for good-specific
2 Unless otherwise noted, integration is over the entire set of goods in the relevant stage of production.
147
efficiencies. For each country i and stage s, zis in (9.2) and (9.3) is drawn from a Fréchet
distribution characterized by the cumulative distribution function
s −θ
Fis (z) = e− Ai z ,
for s = 1, 2 and i = 1, 2, where Ais > 0 and θ > 1. Efficiency draws are independent
across goods, stages, and countries. The probability that a particular stage-s good ω can
be produced in country i with efficiency less than or equal to zis is given by Fis zis . Since
draws are independent across the continuum of goods, Fis zis also denotes the fraction of
stage-s goods that country i is able to produce with efficiency at most zis .
Following Eaton and Kortum (2002), it is straightforward to show that the distribution
of prices of stage-1 goods that country i offers to country j equals
−θ
ij
− Ais qs pθ
Gijs ( p) = 1 − e ,
ij
where qs is the unit cost of producing and shipping the good. This means that the overall
distribution of prices of stage-s goods available in country j is
j θ
Gs ( p) = 1 − e−Φs p ,
j
(9.4)
where
kj −θ
∑ Aks
j
Φs ≡ qs . (9.5)
k
The probability that country j buys a certain good from country i, as Eaton and Kortum
(2002) show, equals
−θ
ij
ij
Ais qs
λs = j
. (9.6)
Φs
As in Eaton and Kortum (2002), it is also true that, because the distribution of stage-s
goods actually purchased by country j from country i is equal to the overall price distri-
148
j ij
bution Gs , the fraction λs of goods purchased from country i also equals to the fraction of
country j’s total expenditures on stage-s goods that it spends on goods from country i.
The interesting feature of this intermediate inputs structure is that the specialization
patterns introduced by Yi (2003) still hold. However, the model is much easier to calibrate
given that the function that determined comparative advantage can be easily linked to
observable trade shares for each stage of production.
149
Chapter 10
The gravity estimator and trade cost
estimates
In the previous few classes, we have seen how to calibrate gravity models to coincide with
observed trade data and perform counterfactuals for any change in bilateral frictions. One
lesson from these procedures is that the gravity constants play an enormously important
role in determining the general equilibrium forces of the model. A natural follow-up
question is how to estimate these gravity constants. This is the question we turn to in this
lecture. In much of this section we use the notation for the general gravity framework that
we considered in Section XXX.
10.1 The Head and Ries procedure
The Head and Ries (2001) procedure is a very simple method of estimating the parameters
on distance that dispenses of the need of computing the equilibrium of the model. If one
looks at the relationship
Xij X ji 1− σ
= τij τji (10.1)
Xii X jj
then this relationship is an adjustment that takes care of the critique of Anderson and
vanWincoop of neglecting the impact of parameters on general equilibrium variables. Pa-
150
rameters can be estimated through a linear regression.
10.2 Gravity estimators
Let us first do a little algebra using conditions 1-3 of the universal gravity framework in
order to contrast some traditional estimators with more structural approaches. By com-
bining the gravity equation with the balanced trade condition, we can write the destina-
tion fixed effect δj as a function of its income Yj and the origin fixed effects in all other
countries:
Yj = ∑ Xij ⇐⇒
i ∈S
Yj = ∑ Kij γi δj ⇐⇒
i ∈S
Yj
δj =
∑i∈S Kij γi
Substituting this expression back into the gravity equation yields an expression for bilat-
eral trade flows that depends only on the origin fixed effect:
Xij = Kij γi δj ⇐⇒
Kij γi
Xij = Yj . (10.2)
∑k∈S Kkj γk
Substituting equation (10.2) into the goods market clearing condition allows us to write
the origin fixed effect γi as a function of the origin income Yi and the origin fixed effects
151
in all other countries:
Yi = ∑ Xij ⇐⇒
j∈S
Kij γi
Yi = ∑∑ Yj ⇐⇒
j∈S j∈S Kkj γk
Yi
γi = Kij
.
∑ j∈S Y
∑k∈S Kkj γk j
Finally, substituting this expression back into the gravity equation (10.2) allows us to write
bilateral trade flows as a function of the (exogenous) bilateral frictions Kij , the income in
the origin and destination, and measures of the “bilateral resistance”:
Kij γi
Xij = Yj ⇐⇒
∑k∈S Kkj γk
Yi Yj
Xij = Kij × × ⇐⇒
Kij
∑ j∈S ∑k∈S Kkj γk Yj ∑k∈S Kkj γk
Yi Yj
Xij = Kij × × , (10.3)
Y
∑k∈S Kik Πkk Πj
where we define Π j ≡ ∑k∈S Kkj γk . Let us call equation (10.3) the structural gravity equa-
tion.
10.2.1 The traditional gravity estimator
Until about a decade ago, almost all estimation procedures based on the gravity equation
assumed that trade frictions Kij were a linear function of observed bilateral covariates (e.g.
distance, common language, shared border, etc.) Tij , i.e.:
Kij = Tij β + ε ij
and estimated β by running the following regression:
152
ln Xij = Tij fi + ln Yi + ln Yj . (10.4)
Call equation (10.4) the traditional gravity estimator. Comparing the traditional estimat-
ing gravity equation to the structural gravity equation (10.3), it is immediately obvious
that the traditional estimating gravity equation is missing (i.e. not controlling for) Πi or
Y
∑k∈S Kik Πkk . Since we can write the structural gravity equation as:
Y
Kij Πjj
Xij = Y
× Yi ,
∑k∈S Kik Πkk
we can see that the structural gravity equation implies that the share of trade flows from
Y Yk
i to j depends on how large Kij Πjj is to Kik Π k
in all other countries. Since Π j ≡ ∑k∈S Kkj γk ,
destinations that are more economically remote (in terms of having lower Kkj than aver-
age) will tend to have lower Π j , which will cause country i ∈ S to export a greater share
of its total trade to those destinations. Intuitively, this is because more remote countries
will have higher price indices, and hence will be willing to pay more for any given good.
This is what Anderson and Van Wincoop (2003) refer to as “multilateral resistance.”
Because Π j will (generically) varies across destinations, the traditional estimating grav-
ity equation will suffer from omitted variable bias. Furthermore, because Π j depends on
the average trade friction between j and the rest of the world, it will be correlated with Kij ,
which will result in biased estimates of β. This means that you should never use the tradi-
tional estimating gravity equation to estimate trade costs. Indeed, Baldwin and Taglioni (2006)
award papers doing this with the “gold medal error” of estimating gravity equations.
153
10.2.2 The fixed effects gravity estimator
An alternative to the traditional estimator is to take logs of the gravity equation (??):
ln Xij = ln Kij + ln γi + ln δj .
If we assume that ln Kij = Tij β + ε ij , this equation becomes:
ln Xij = Tij β + ln γi + ln δj + ε ij . (10.5)
Call equation (10.5) the fixed effects gravity estimator. Since Tij are observed and ln γi
and ln δj can be estimated by including dummy variables for each origin country and each
destination country (note: there are two dummy variables for country), β can be estimated
consistently by applying the fixed effects estimator to equation (10.5) and γi and δj can be
consistently estimated (to scale) from the coefficients on the dummy variables.
Given the estimates from equation (10.5), the fixed effects gravity estimator allows us
to recover the multilateral resistance terms (to scale) in the structural gravity equation as
follows. By taking exponents of the estimates, we can back out predicted values (up to
scale) of the origin fixed effect and the bilateral trade costs:
γ̂i ≡ exp lnˆγi
K̂ij ≡ exp Tij β̂
Since Π j ≡ ∑k∈S Kkj γk , we then can construct an estimate of the destination multilateral
resistance term:
Π̂ j ≡ ∑ K̂kj γ̂k = ∑ exp T kj β̂ + lnˆγ k
k∈S k∈S
Yk
which then allows you to construct the origin multilateral resistance term ∑k∈S K̂ik Π̂ .
k
Note that nowhere in these derivations did we use the estimated destination fixed effect,
154
which suggests that the destination fixed effects are “nuisance parameters,” i.e. they are
unnecessary (given the equilibrium conditions) to fully derive the gravity equation. This
is because, as we saw above, balanced trade implies that the destination fixed effect is
pinned down by the origin fixed effects:
Yj
δj = ,
∑i∈S Kij γi
a restriction that is not made in equation (10.4). This is the major drawback of this esti-
mation procedure: by relying just on the gravity structure of the model, the fixed effects
estimator imposes no equilibrium conditions on the estimation, and as such, the result-
ing estimates will (generically) not ensure that the goods market clears or that trade is
balanced. The other major drawback of the fixed effect estimation procedure is that there
may be computational difficulties to including so many dummy variables, especially if
one is interested in estimating γi and δj . (However, if one is simply interested in esti-
mating β, there exist new ways of doing so without having to invert the large dependent
variable matrix, see ?).
That being said, the fixed effects gravity estimator is probably the most common es-
timator of gravity equations today, as it is simple to implement and model consistent
(with the caveat above). It is also straightforward to extend the fixed effects gravity esti-
mator to include multiple years (by adding origin-country-year and destination-country-
year dummies), multiple industries (by adding origin-country-industry and destination-
country-industry dummies), etc.
155
10.2.3 The ratio gravity estimator
An alternative approach is to consider as the dependent variable the (log) trade shares
rather than the (log) trade levels. From gravity equation (??) we have:
Xij Kij γi δj Kij γi
= = =⇒
X jj K jj γ j δj K jj γ j
Xij Kij
ln = ln + ln γi − ln γ j .
X jj K jj
Kij
If we assume that ln K jj = Tij β + ε ij , then this equation becomes:
Xij
ln = Tij β + ln γi − ln γ j + ε ij . (10.6)
X jj
Following Head and Mayer (2013), we call the (10.6) the ratio gravity estimator. Because
the destination fixed effect is constrained to be the negative of the origin fixed effect, the
ratio gravity estimator no longer has any nuisance parameters in the estimation, which
makes the estimation easier to implement using dummy variables. However, while the
ratio gravity estimator has N (where N is the number of countries) fewer parameters
to estimate than the fixed effect estimator, it also has N fewer observations since any
time i = j equation (10.6) simplifies to the trivial equation 0 = ε ii ; hence the degrees
of freedom remains unchanged. In addition, as with the fixed effect estimator, the ratio
gravity estimator is based only on the gravity equation (??), so it too does not impose that
the general equilibrium conditions hold.
Furthermore, since unlike the fixed effects estimator, the ratio gravity estimator only
identifies the relative trade frictions rather than the absolute trade frictions. That is, taking
K
exponents of ln Kijjj = Tij β + ε ij yields:
K̂ij = exp Tij β̂ K̂ jj .
156
This means that we are unable to recover the multilateral resistance term Π j since:
1 Π̂ j
∑ exp Tkj β̂ + lnˆγk =
K̂ jj
∑ K̂kj γ̂k = K̂ .
k∈S k∈S jj
In order to call this expression the multilateral resistance, one would have to assume (as
is often done) that K jj = 1, i.e. internal trade is costless.
10.2.4 The Anderson and van Wincoop procedure
Anderson and Van Wincoop (2003) develop a general equilibrium methodology to obtain
estimates of the costs of trade in the model as a function of distance proxies.
Using the equation (3.8) we have
1− σ
Xj
Xi τij
= wi1−σ ∑ αij
XW j
Pj XW
where XW is total world spending (income). Using the bilateral demand
1− σ
τij
Xij = pij xij = αij wi1−σ Xj
Pj
we then have 1− σ
Xi X j τij
Xij =
W
αij 1/(1−σ)
X 1− σ
Xk
∑k αik
τik
Pk XW
Pj
while by summing up over ı́’s we can compute the price index,
1− σ
Xk 0 X j τk0 j
∑ Xk 0 j = ∑0 =⇒
α 0
k j
XW 1/(1−σ)
1− σ
k0
k Xk
∑k αik
τik
Pk XW
Pj
1/(1−σ)
1− σ
Xk 0
αk0 j τk0 j
Pj = ∑ XW
1− σ
Xk
k0 ∑k αik Pikk
τ
XW
157
where we used the fact that balanced trade implies X j = ∑k Xkj .
If we define
1− σ
Xj
τkj
Ξ1k −σ = ∑ αkj
j
Pj XW
then
" 1− σ #1/(1−σ)
τkj Xk
Pj = ∑ αkj Ξ1k −σ XW
k =1
under symmetric trade barriers, τij = τji , αij = αij , from the last equations it turns out that
Ξ j = Pj , so that
1− σ
Xi X j
τij
Xij = W
X Pi Pj
Anderson and Van Wincoop (2003) estimate the stochastic form of the equation
Xij
ln = k + a1 ln τ̃ij − a2 Dij − ln Pi1−σ − ln Pj1−σ + ε ij (10.7)
Xi X j
(1− σ )
where Dij is a dummy variable related to borders and a1 = (1 − σ) ã1 , τ̃ij = τij . The
innovation of Anderson and vanWincoop was to perform this estimation expressing Pi , Pj
as an explicit function of the model parameters, σ and ã1 , a2 as well as (observable) mul-
tilateral resistance terms. The authors cannot separately estimate σ since its effect on dis-
tance cannot be separately identified from ã1 with their methodology. Nevertheless, their
method delivers much more sensible effects for the coefficient on borders. Estimation
without considering Pi , Pj as a function of the parameters to be estimated overstates the
effect of distance of trade. The intuition is that smaller countries are likely to have higher
price indices since they impose trade barriers to larger countries.
158
10.3 The no arbitrage condition
If prices are proportional to marginal costs and there are iceberg trade costs then for any
product ω ∈ Ω produced in any origin i ∈ S and for any destinations j ∈ S and k ∈ S:
pij (ω ) τij
= .
pik (ω ) τik
If we assume that τii = 1 for all i ∈ S then setting k = i yields the no-arbitrage condition:
pij (ω )
=τij . (10.8)
pii (ω )
The no-arbitrage condition provides an exceedingly simply and surprisingly powerful
way of identifying the iceberg trade costs. The simplicity of the identification is self-
evident: if an origin sells a good to itself and sells it to another destination, then the
iceberg trade costs is simply equal to the ratio of the destination price to the origin price.
Why is the result surprisingly powerful? It is because no-arbitrage condition should
hold regardless of the model. To see this, suppose that the no-arbitrage condition did
pij (ω )
not hold and instead that pii (ω )
> τij . This should not be an equilibrium, because any self-
interested arbitrageur could purchase the good in i ∈ S, resell the good in j ∈ S, and make
pij (ω )
a profit. Conversely, suppose that pii (ω )
< τij . This implies that whoever was selling the
good from i to j ought to have just sold locally. In the first case, money was being left on
the table, while in the latter case, money was being thrown away, both of which tend to
make us economists nervous.1
Despite the simplicity and power of using the no-arbitrage condition to identify the
iceberg trade costs, there are three major difficulties in empirically implementing the esti-
mation strategy:
1 In p
my job market paper, I argue that it can be the case that pijii > τij if it is costly for arbitrageurs to
discover what the price is in other locations. I believe that such “information frictions” are quantitatively
important in real world markets.
159
1. The observed prices in both the origin and destination have to be for the same good
ω. Observing prices of identical goods is especially difficult for differentiated vari-
eties; for example, prices of t-shirts across locations may vary because of the quality
of the t-shirts rather than because of trade costs.
2. Even if the goods are identical, we need to know that the good was produced in i ∈ S
and sold to j ∈ S. If j ∈ S purchased a good from another location (or produced it
locally), there is no reason that the no-arbitrage equation must hold with location
i. Note the inherent tension between the first difficulty and this difficulty: if one is
able to find a good that is truly identical across locations (e.g. a commodity), there
is a high likelihood that it is produced in many locations.
3. The no-arbitrage condition only holds if the price of the good is proportional to
the marginal cost of production. This assumption would be violated, for example, if
producers had market power and were able to charge variable mark-ups in different
destinations. Note in the case of CES, producers do not charge variable mark-ups,
but this result is particular to the CES case (and likely unrealistic).
Let us now discuss some of the approaches taken in the trade literature that have at-
tempted (more or less successfully) to navigate these three difficulties. We now consider
four methods used to estimate the no-arbitrage conditions.
10.3.1 The Eaton and Kortum (2002) approach
Eaton and Kortum (2002) observe 50 manufactured products across the 19 countries in
their data set. They note that if a product ω ∈ Ω is not traded between the two countries,
then it must be the case that producers of ω found it more profitable to sell domestically,
160
i.e.:
p j (ω ) p j (ω )
pi ( ω ) > ⇐⇒ τij > ,
τij pi ( ω )
i.e. the iceberg trade costs exceed the price gap. Conversely, if a product is traded, then
the no arbitrage equation holds with equality. These two facts imply that the price ratio
of all products is bounded above by the iceberg trade cost.
Since they do not observe which of the 50 manufactured products are actually traded
between any pair of countries, they employ a “brute force” method of estimating the trade
cost by taking the maximum price ratio observed across all products as their measure of
the bilateral iceberg trade costs:
p j (ω )
τ̂ijEK ≡ max . (10.9)
ω ∈Ω pi ( ω )
Equation (10.9) is a valid estimator of the true iceberg trade costs if at least one of the
observed products is traded, prices are measured perfectly, the products observed are
identical, and prices are proportional to marginal cost.2
Using this estimator, Eaton and Kortum (2002) find a trade elasticity (i.e. α1 ) of roughly
eight; i.e. a 10 percent increase in trade costs is associated with an 80% decline in trade
flows. More recently, ? have argued that because it is possible for none of the observed
products to have actually been traded, an estimator based on equation (10.9) will be biased
downwards. Because observed trade flows can be rationalized equally well with a higher
trade elasticity and lower trade costs or a lower trade elasticity and higher trade costs,
if the estimated trade costs are biased downwards, the implied elasticity of trade will be
biased upwards. They develop a simulated method of moments estimator that corrects
this error, and find an elasticity of trade of approximately four, which currently is the
standard in the trade literature.
2 Recognizingthat prices likely are measured with error, Eaton and Kortum (2002) actually use the second
highest observed price ratio as their preferred estimator of the iceberg trade cost.
161
10.3.2 The Donaldson (2014) approach
In Donaldson (2014) (which we will see in detail in a few lectures), the author had a clever
solution to the three difficulties mentioned above of estimating the no-arbitrage condition.
He found a homogeneous good where the unique location of production was known: salt!
As he writes:
“Throughout Northern India, several different types of salt were consumed,
each of which was regarded as homogenous and each of which was only ca-
pable of being made at one unique location.”
In the simplest case, having such a good would allow one to construct bilateral trade
pij (ω )
costs immediately from the no-arbitrage equation, as τij = pii (ω )
. However, even with
“perfect” good for which to apply the no-arbitrage condition, Donaldson (2014) faced
two additional difficulties. First, it turned out that he did not observe the price of a variety
ω ∈ Ω of salt at the origin. Second, since not every location i ∈ S produced its own unique
variety of salt, at best, he could only apply the no-arbitrage condition to find a subset of
the bilateral iceberg trade costs.
To solve both problems, Donaldson (2014) made a parametric assumption that ln τij =
Tij β + ε ij . With this assumption, the no arbitrage condition becomes:
pij (ω )
τij = ⇐⇒
pii (ω )
ln pij (ω ) = ln pii (ω ) + ln τij ⇐⇒
ln pij (ω ) = ln pii (ω ) + Tij β + ε ij . (10.10)
By including a salt-variety ω fixed effect, β can be estimated using just the observed vari-
ation in prices of a particular variety across destinations. Once β is estimated, the trade
costs between any origin and destination can be imputed from the parametric assump-
162
tion. Furthermore, α can be estimated by regressing bilateral trade flows on Tij β̂ using
the gravity regression in equation (??). Donaldson (2014) estimates the elasticity of trade
flows for each commodity in his data set separately and finds a mean of roughly four,
consistent with Simonovska and Waugh (2013).
The Donaldson (2014) approach assumes that salt is traded in perfectly competitive
markets, which, because salt is a commodity, seems reasonable.
10.3.3 The Allen (2012) approach
In Allen (2012), I use the spatial dispersion in prices of agricultural commodities (which,
unlike Donaldson (2014), were produced in many regions) in order to infer the size of
trade costs.
The insight of the approach in this paper is to note that even when two countries
do not trade, the no arbitrage condition provides information about the size of the trade
costs. The intuition is the same as in the Eaton and Kortum (2002) above: if a particular
commodity is not observed to be traded between two locations, then this must mean that
the trade cost exceeded the price gap between the two locations, i.e. Xij (ω ) = 0 =⇒
p j (ω )
τij > pi ( ω )
, whereas when trade does occur between the two locations, then this must
p j (ω )
mean that the no-arbitrage equation holds, i.e. Xij (ω ) > 0 =⇒ τij = pi ( ω )
.
Suppose we observe the price of a particular commodity ω ∈ Ω in each location i ∈ S
in each period t ∈ {1, ..., T }, i.e. pit (ω ). Furthermore, suppose for any pair of origin i ∈ S
and destination j ∈ S, we observe whether or not trade flows occur in each period t ∈
{1, ..., T }, i.e. 1 Xijt (ω ) > 0 . Finally, suppose that the bilateral trade cost of commodity
ω in time t depend on a time invariant bilateral trade cost and an idiosyncratic error that
is i.i.d. across time periods:
ln τijt (ω ) = ln τij (ω ) + ε ijt (ω ) ,
163
where ε ijt (ω ) ∼ N 0, σ2 .
We can then estimate ln τij (ω ) using a maximum likelihood routine. The log likeli-
hood function can be written as:
T
p jt (ω )
1
l ln τij , σ = ∑ (1 Xijt (ω ) > 0 ln φ
ln − ln τij +,
t =1
σ pit (ω )
p jt (ω )
1
1 Xijt (ω ) = 0 ln 1 − Φ
ln − ln τij )
σ pit (ω )
which bears a very close resemblance to a Tobit estimator. Using this estimator to identify
trade costs and then regressing trade flows on these trade costs to identify the trade elas-
ticity yields an elasticity of a little bit more than two in the context of agricultural trade
flows between islands in the Philippines.
In the presence of information frictions where positive trade flows merely indicate that
p j (ω )
the price ratio exceeds the bilateral trade costs (i.e. Xij (ω ) > 0 =⇒ τij < pi ( ω )
), the the
log likelihood function can be written as:
T
p jt (ω )
1
l ln τij , σ = ∑ (1 Xijt (ω ) > 0 ln Φ
ln − ln τij +,
t =1
σ pit (ω )
p jt (ω )
1
1 Xijt (ω ) = 0 ln 1 − Φ
ln − ln τij ).
σ pit (ω )
In this case, the log likelihood function is identical to the following Probit regression:
p jt (ω )
1 Xijt (ω ) > 0 = β ln + αij ,
pit (ω )
1
where β = σ and αij = − σ1 ln τij (ω ). Hence, identifying the iceberg trade cost is straight-
forward: you regress (using a Probit) whether or not trade flows occurred on the observed
log price ratio and a constant. The coefficient on the price ratio identifies the variance of
the distribution of measurement error in the trade costs and the variance, combined with
the constant, identifies the iceberg trade cost. Intuitively, as the measurement error goes
164
to zero, any increase of the log price ratio above the threshold αij will induce trade with
probability one, so that β will approach infinity.
The advantage of this approach relative to Eaton and Kortum (2002) is that it explic-
itly allows for measurement error in trade costs; the advantage of this approach relative to
Donaldson (2014) is that one does not need to observe exactly where a product was pro-
duced. The disadvantage relative to both approaches is that requires knowing whether or
not trade flows occurred at the product level.
Like in the Donaldson (2014) approach, an assumption in the Allen (2012) approach is
that prices are proportional to marginal costs, which because the focus is on agricultural
commodities, seems reasonable.
165
Chapter 11
Parameter Choices for Gravity
Models
Anderson and Van Wincoop (2003) developed a framework that delivers structural re-
lationships for trade among countries (or regions) based on the model analyzed in sec-
tion (3.3). This model is useful to identify parameters related to the cost of distance and
the border. As we showed in the previous chapters, and as elaborated in Anderson and
Van Wincoop (2004) and Arkolakis, Costinot, and Rodrı́guez-Clare (2012), that basic setup
has very similar properties in terms of bilateral aggregate trade and welfare to richer
models of trade and heterogeneity. New, heterogeneous-firm models generate a num-
ber of predictions at the firm-level which can also be used to obtain key parameters of the
model. In this chapter we will discuss the identification of key parameters of these mod-
els determining aggregated but also disaggregated trade. Alternative ways of estimating
gravity equations are summarized in a survey by Anderson and Van Wincoop (2004)
11.1 Recovering the trade elasticity from gravity estimates
In the gravity regressions above, we estimated the bilateral trade friction matrix Kij . In
1
models where Kij = τijα , if we observed the iceberg trade costs and had estimates for Kij ,
166
we could estimate the gravity constant with a simple log-linear regression:
ln τij = α ln Kij + ε ij .
Unfortunately, because trade flows alone will only allow us to identify the total bilateral
trade frictions Kij , so we need to rely additional data to recover the iceberg trade costs.
The most popular method of doing so is to rely on price data and the no-arbitrage condi-
tion.
An important subset of the models that fit into the universal framework are those for
which β = 0. This is because in many of these models the gravity constant α is related to
the partial elasticity of trade flows to iceberg trade costs. For example, in the Armington
model, we have the gravity equation:
Xij = τij1−σ wi1−σ Aiσ−1 Pjσ−1 Ej
and the labor market clearing condition
Yi = Ei = wi Li
σ −1
Ai
Because the origin fixed effect γi ≡ wi , we can write the labor market clearing
condition:
1
Yi = Ai Li γi1−σ ,
1
so that the gravity constant α ≡ 1− σ . Combining this with the gravity equation, we have
that the elasticity of trade flows to iceberg trade costs (which we will refer to as the trade
elasticity) is the inverse of the gravity constant:
∂ ln Xij 1
= 1−σ = .
∂ ln τij α
Hence, for these models, identifying the gravity constant is equivalent to identifying the
167
trade elasticity. As a result, we will turn our focus to procedures for identifying the trade
elasticity, which first requires an examination of traditional reduced form gravity equa-
tions.
11.1.1 The Eaton and Kortum procedure(s)
Another approach that gives an unbiased estimate of parameter a1 is to replace the in-
ward and outward multilateral resistance indices and production variables, Xi − ln Pi1−σ
and X j − ln Pj1−σ , with inward and outward region specific dummies. This approach is
adopted by a series of papers (e.g. Eaton and Kortum (2002)).
Eaton and Kortum also provide a variety of different methods to estimate the param-
eter that governs the elasticity of trade. In the Eaton and Kortum (2002) model this is the
parameter of the Frechet distribution that governs productivity heterogeneity, θ (whereas
in the Armington model it is σ − 1). Using a relationship similar to (4.17) and specifying
intermediate inputs as in equation (9.1), they can derive a relationship of the form
Xij0
ln = −θ ln τij + Si − S j (11.1)
X jj0
where Si = Ai / (1 − ι) − θ ln wi , S j are destination fixed effects and Xij0 = Xij − [(1 − ι) /ι] ln ( Xii /Xi )
with 1 − ι the share of intermediates in manufacturing production.1 They also use proxies
for distance, border effects etc. for the first term in order to estimate θ ln τij but while they
can distinguish the effect of the components (proxies) of that term they cannot distinguish
1 Eaton and Kortum (2002) estimate
ln τij = f + m j + δij
where f includes distance and other geographic barrier fixed effects, m j a destination fixed effect and δij an
error term. To capture potential reciprocity in geographic barriers, they assume that the error term δij consists
of two components: δij = δij1 + δij2 . The country-pair specific component δij2 (with variance σ22 ) affects two-way
trade, so that δij2 = δ2ji , while δij1 (with variance σ12 ) affects one-way trade. This error structure implies that
the variance-covariance matrix of δ diagonal elements E δij δij = σ11 + σ22 and certain nonzero off-diagonal
elements E δij δji = σ22 .
168
that effect from the effect of the multiplicative term θ. To address that problem and using
their estimates from the previous stage for Si they estimate
1
Si = ln Ai − θ ln wi
1−ι
using technology and education fundamentals to be the proxies for Ai and data for wages
adjusted for education. Using a 2SLS estimation they get θ = 3.6.
The second alternative is to estimate the bilateral trade equation (11.1) using their
proxy of ln( Pi dij /Pj ), instead of the geography terms along with source and destination
effects. The proxy for dij is constructed by looking at the (second) highest ratio of prices
of homogeneous products across different destinations and the proxy for Pi /Pj as the
average of these price ratios. Using a 2SLS and geography variables to instrument for the
proxy of ln( Pi dij /Pj ) their estimate for this procedure is a θ = 12.86.
The favorite estimate of the Eaton and Kortum (2002) is the derivation of the θ using
the trade shares equation in terms of prices
−θ
Xij /X j Pi dij
= .
Xii /Xi Pj
Xij /X j
With simple method of moments, −θ is simply the ratio of the mean of ln Xii /Xi and their
Pi dij
proxies of ln Pj . Simonovska and Waugh (2013) propose an alternative estimation of
the Eaton and Kortum (2002) by using the above equation and a simulated method of
moments approach adapted from Eaton, Kortum, and Kramarz (2011).
11.1.2 Gravity trade models where β 6= 0
All the proceeding analysis constrained itself to gravity trade models where the trade elas-
ticity provided sufficient information to recover the gravity constants, i.e. β was assumed
to be zero. We now turn to the general case.
169
From our discussion of identification, we know that trade data alone is insufficient to
estimate the gravity constants; however, if both trade data and information about trade
costs are observed, then the gravity constants can be recovered. Suppose, for example,
that (the change in) trade costs is a function of a vector of observables T̂ij , i.e. ln K̂ij =
T̂ij0 µ, where the prime denotes a transpose. Then the gravity constants can be recovered
in a two-stage estimation process. First, one estimates the (log) change in exporter and
importer shifters using the observed (log) change in trade flows, ln X̂ijo :
ln X̂ijo = T̂ij0 µ + ln γ̂i + ln δ̂j + ε ij ,
where we interpret the residual ε ij as classical measurement error. Second, one estimates
the gravity constants by projecting the observed (log) change in income, ln Ŷio , on the
n o
estimated change in exporter and importer shifters ln γ̂iE and ln δ̂jE :
ln Ŷio = α ln γ̂iE + β ln δ̂iE + νi .
While theoretically straightforward, this procedure is practically difficult, as the model
predicts that the residual νi – unless it is pure measurement error – will be correlated with
both ln γ̂iE and ln δ̂iE . This omitted variable bias arises because any unobserved change in
the income shifter Bi (which causes the income of a location to be higher than observables
would imply) will enter the residual and increase both the location’s exports (through
goods market clearing) and imports (through balanced trade). As a result, estimates of α
and β will be biased upwards.
An alternative procedure is to rely on the general equilibrium structure of the model.
By incorporating the general equilibrium effects within the estimator, there is no need for
a two stage estimation procedure. In particular, we use the structure of the model – which
incorporates both the gravity structure (corresponding to the first stage above) and the
170
generalized labor market clearing condition (corresponding to the second stage above) –
to calculate the change in the exporter and importer shifters directly. Formally, we can
estimate the gravity constants α and β and the trade cost parameter µ by minimizing the
squared errors of the observed change in trade costs and the predicted change in trade
costs:
2
(α∗ , β∗ , µ∗ ) ≡ arg ∑∑ ln X̂ijo − T̂ij0 µ − ln γ̂i T̂µ; α, β − ln δ̂j T̂µ; α, β
min ,
α,β∈R,µ∈RS i j
(11.2)
where we emphasize that the change in the origin and destination shifters will be deter-
mined in general equilibrium and depend on both the gravity constants and the trade cost
parameter.
It turns out that equation (11.2) is best solved by first estimating the µ given a set of
gravity constants α and β and then solving for the α and β. Denote µ (α, β) as the trade
cost parameter which minimizes the squared error for a given α and β, i.e.:
2
µ (α, β) = arg min ∑ ∑ ln X̂ijo − T̂ij0 µ − ln γ̂i T̂µ; α, β − ln δ̂j T̂µ; α, β
µ ∈RS i j
First order conditions are:
!!!
∂ ln γ̂i ∂ ln δ̂j
−2 ∑ ∑ T̂ij + ∑ ∑ T̂kl ln X̂ij − T̂ij0 µ∗ − ln γ̂i T̂µ∗ − ln δ̂j T̂µ∗ = 0
+
i j k l
∂ ln Kkl ∂ ln Kkl
Consider the following first order approximations of the log change in the exporter and
importer shifters:
∂ ln γi ∂ ln δj
∑ ∑ ∂ ln Kkl T̂kl0 µ and ln δ̂j ∑ ∑ ∂ ln Kkl T̂kl0 µ.
ln γ̂i T̂µ ≈ T̂µ ≈ (11.3)
k l k l
171
Substituting these approximations into the first order conditions yields:
!!!
∂ ln γ̂i ∂ ln δ̂j
0 = ∑ ∑ T̂ij + ∑ ∑ T̂kl + ×
i j k l
∂ ln Kkl ∂ ln Kkl
! !
∂ ln γ ∂ ln δ
ln X̂ij − T̂ij0 + ∑ ∑ T̂kl + ∑ ∑
i 0 j
T̂kl0 µ∗ ,
k l
∂ ln K kl k l
∂ ln K kl
or equivalently:
! −1
∂ ln X̂ij ∂ ln X̂ij
∗
µ = ∑∑∑∑ T̂ T̂ 0
∂ ln Kkl kl ij ∑ ∑ ∑ ∑ ∂ ln Kkl T̂kl T̂ij0 ln X̂ij
i j k l i j k l
It turns out that this expression has an intuitive interpretation. To see this, let us first write
it in matrix form. Let T̂ denote the N 2 × S vector whose hi + j ( N − 1)i row is the 1 × S
vector T̂ij0 , D (ff, fi) is the N 2 × N 2 matrix whose hi + j ( N − 1) , k + l ( N − 1)i element is
∂ ln Xij
∂ ln Kkl (which we know from previous lecture is a function only of the gravity constants
and observed trade flows), and ŷ denote the N 2 × 1 vector whose hi + j ( N − 1)i row is
ln X̂ijo . Then the general equilibrium gravity estimator is:
0 −1 0
µ ( a, β) = D (α, β) T̂ D (α, β) T̂ D (α, β) T̂ ŷ. (11.4)
Equation (11.4) says that, to a first order, the general equilibrium estimator is the coeffi-
cient one gets from of an ordinary squares regression of the observed hatted variables on
a “general equilibrium transformed” explanatory variable T̂ijGE :
0
ln X̂ijo = T̂ijGE µ + ε ij ,
where:
∂ ln X̂ij
T̂ijGE ≡ ∑ ∑ ∂ ln K̂ T̂kl .
k l kl
Intuitively, the general equilibrium transformed regressors capture the effect of the entire
set of explanatory variables on any particular observed bilateral trade flow. As a result,
172
µ (α, β) directly accounts for all (first-order) general equilibrium effects arising from the
network structure of trade flows.
We can then find the gravity constants α and β which minimize the total squared error.
From equation (11.4) (and the fact that a projection matrix is idempotent), the estimation
of the gravity constants can be written as:
∗ ∗ 0
0 −1 0
(α , β ) = arg min ŷ I − T̂ D (α, β) T̂ D (α, β) T̂ D (α, β) T̂ ŷ. (11.5)
α,β∈R
Equation (11.5) can be estimated using traditional optimization procedures. Using the
trade cost shock of joining the WTO, ? find gravity constants are approximately -30, con-
sistent with a trade elasticity of fourteen and a labor share in the production function of
0.08.
11.2 Calibration of a firm-level model of trade
Parameters Determining Firm Sales Advantage We now turn to techniques developed
in determining deeper structural parameters of these models, that determine the micro
behavior of the firms. An example of these parameters is the marketing parameter β
and the ratio of the Pareto parameter and the elasticity of sales, θ̃ = θ/ (σ − 1), used
in Arkolakis (2010b). Both these parameters determine the distribution of sales of firms
and can be calibrated by looking at the size advantage of prolific exporters, i.e. the size
advantage of firms that are able to penetrate more markets.
This advantage can be uncovered by looking at the following two structural relation-
ships of the model i) normalized average sales of firms from France, F, conditional on
selling to market j,
−1/θ̃ −1/( β̃θ̃ )
M Fj M Fj
M FF M FF
X̄FF| j 1−1/θ̃
− 1−1/(θ̃ β̃)
= 1 1
(11.6)
X̄FF −
1−1/θ̃ 1−1/ (θ̃ β̃)
173
and ii) exporting intensity of firms in percentile PrFj in market j,
1 − 1 − PrFj ( )
1/ θ̃β
t Fj PrFj t FF (PrFF )
/ = (11.7)
X̄Fj X̄FF MFj −1/θ̃
1/(θ̃β) MFj −1/(θ̃ β̃)
MFF − 1 − Pr Fj MFF
Notice that parameters θ and σ affect equations (11.6) and (11.7) only insofar they affect θ̃.
Higher θ implies less heterogeneity in firm productivities (and thus in firm sales), whereas
higher σ translates the same heterogeneity in productivities to larger dispersion in sales.
For the calibration, Arkolakis (2010b) uses a simple method of moments estimate. In
particular, β and θ̃ are picked so that the mean of the left-hand side is equal to the mean of
the right-hand side for both equation (11.6) and equation (11.7) evaluated at the median
percentile in each market j. The solution delivers β = .915 and θ̃ = 1.65. Notice that using
equation (11.6), a method of moments estimate for the fixed model with β = 0 gives a
θ̃ = 1.49.
To complete the calibration of the model, we need to assign magnitudes to σ and θ.
Broda and Weinstein (2006) estimate the elasticity of substitution for disaggregated cat-
egories. The average and median elasticity for SITC 5-digit goods is 7.5 and 2.8, respec-
tively (see their table IV). A value of σ = 6 falls in the range of estimates of Broda and
Weinstein (2006) and yields a markup of around 1.2, which is consistent with those values
reported in the data (see Martins, Scarpetta, and Pilat (1996)). In addition, θ̃ = 1.65 and
σ = 6 imply that the marketing costs to GDP ratio in the model is around 6.6% within the
range of marketing costs to GDP ratios reported in the data. Finally, this parameteriza-
tion implies that θ = 8.25 for the endogenous cost model which is very close to the main
estimate of Eaton and Kortum (2002) (8.28) and within the range of estimates of Roma-
lis (2007) (6.2 − 10.9) and the ones reported in the review of Anderson and Van Wincoop
(2004) (5 − 10). Since the model retains the aggregate predictions of the Melitz-Chaney
framework if θ is the same I will calibrate the two models to have θ = 8.25. For the fixed
174
cost model, given the calibrated θ̃ = 1.49, it implies a σ = 6.57.
Calibration for a multi-product firms model
Parametrizing a multi-products firm model requires to dig deeper into establishing
predictions at the within-firm level. We will now briefly go over the calibration procedure
of Arkolakis and Muendler (2010) for their model described in 8.3. Guided by various
log-linear relationships observed in their data (see, for example Figure ??) they specify
the following functional relationships
f ij ( g) = f ij · gδ for δ ∈ (−∞, +∞),
(11.8)
h( g) = gα for α ∈ [0, +∞).
This specification gives product level sales for the g-th ranked product of the firm as
!σ−1
z
pijg (z) xijg (z) = σ f ij (1) Gij (z)δ+α(σ−1) g−α(σ−1) .
zij∗,G
Using the logarithm of this structural relationship, a regression of the sales of the firm on a
constant, a firm fixed effect and the number of the products of the firm obtains α (σ − 1) =
2.66 and δ ' −1.38.
11.3 Estimation of a firm-level model
We present here the framework of Eaton, Kortum, and Kramarz (2011) that is the first
work that estimates a multi-country firm-level model of trade making use of the firm-
level data. The idea is to identify a set of micro facts on exporters and to develop a con-
sistent modeling framework that would explain these micro observations using model
relationships. Then the authors estimate the fundamental parameters of the model using
the micro data. In this respect the paper of Eaton, Kortum, and Kramarz (2011) is parallel
to the Eaton and Kortum (2002) framework.
175
11.3.1 The model
Sales of the firm are given by
1− σ
pij
tij (ω ) = a j (ω ) nij ,
Pj
derived by asymmetric CES utility function with preference for each good affected by
a j (ω ) (these could be interpreted as Armington type bias in a particular good). The term
a j (ω ) reflects an exogenous demand shock specific to good ω in market j. The term Pj is
the CES price index that will be analyzed in a moment.
Producers are heterogeneous and the unit cost for a producer from i in producing a
good and shipping it to country j is
wi τij
cij (ω ) =
zi ( ω )
where τij is an iceberg cost. The measure of potential producers who can produce their
good with efficiency at least z is
µi ( z ) = Ai z −θ .
Given the unit cost this implies that the measure of goods that can be delivered to country
j from anywhere in the world at unit cost c or less in j is
N
µ j (c) = ∑ µkj (c)
k =1
N −θ
= ∑ Ak wk τkj cθ
k =1
N
≡ ∑ Φkj cθ
k =1
≡ Φ j cθ
Conditional on selling in a market the producer makes the profit from producer from
176
i in j
1− σ
c j (ω ) 1 − (1 − n )1− β
p
πij (ω ) = max 1 − a j (ω ) n X j − ε j (ω ) f j ,
p,n p Pj 1−β
where c j (ω ) is the unit production cost, ε j (ω ) an entry cost and f j > 0. Producer charges
a constant markup
pij = m̄c j (ω ) , m̄ = σ/ (σ − 1)
Define
a j (ω )
η j (ω ) = .
ε j (ω )
Thus, we can describe seller’s behavior in market j in terms of its cost draws c j (ω ) = c,
the demand shock a j (ω ) = a, and the redefined entry shock η j (ω ) = η. It can be shown
using the results of section 8.2 combined with this framework that a firm will enter a
market j iff its cost draw c ≥ c̄ j (ω )
1/(σ−1)
Xj Pj
c̄ j (η ) = η . (11.9)
σ fj m̄
Notice that the entry threshold depends on a only through η. For the firms with c ≥ c̄ j (ω )
the fraction of buyers reached in a market will be (for β > 0)
(σ−β 1)
c
nij (η, c) = 1 −
c̄ j (η )
You can rewrite sales as
(σ−β 1) −(σ−1)
c c
tij (η ) = ε j 1 − σ fj
c̄ j (η ) c̄ j (η )
Notice that even though Eaton, Kortum, and Kramarz (2011) add these 3 levels of firm
heterogeneity they can determine easily all the aggregate variables of the model. First,
177
the price index is given by the following integration
"ˆ ˆ ˆ ! #−1/(σ−1)
c̄ j (η )
1− σ 1− σ
Pj = αnij (η, c) m̄ c dµ j (c) g (α, η ) dαdη
0
ˆ ˆ
θ θ
= m̄ Φ j − β −1
αc̄ j (η )θ −(σ−1) g (α, η ) dαdη
θ − σ + 1 θ + ( σ − 1)
β
which substituting for the entry hurdle (11.9) gives
(1/θ )−1/(σ−1)
Xj
−1/θ
Pj = m̄ κ1 Φ j ,
σ fj
where
ˆ ˆ
θ θ θ −(σ −1)
κ1 = − β −1
αη σ −1 g (α, η ) dαdη ,
θ − σ + 1 θ + ( σ − 1)
β
and g (α, η ) is the joint density of the realizations of producer-specific costs. Second, from
the model we can get a series of relationships directly related to observables. The measure
of entrants in market j is
ˆ
Mj = c̄ j (η ) g
κ2 X j
=
κ1 σ f j
where
ˆ
κ2 = η θ/(σ−1) g2 (η ) dη
Number of firms selling from i to j
κ2 λij X j
Mij = ,
κ1 σ f j
where
Φij
λij =
Φj
being the observed market share, which exactly the same as in the monopolistic compe-
178
tition model with productivity as the only source of variation. Finally, average sales are
given by
κ1
X̄ij = σ fj
κ2
It also turns out that the distribution of sales in a market, and hence mean sales, is invari-
ant to the location of the supplier.
Notice that all these relationships are derived independently of the actual distribu-
tion of demand and entry shocks. This separability allows for a very simple and generic
solution of the model that retains the forces of the previous structure while allowing for
additional levels of heterogeneity that brings the model closer to the data.
11.3.2 Estimation, simulated method of moments
There are particular steps in the estimation procedure proposed by the authors. They
match 4 sets of moments (each set of moments is denoted as m)
a) The distribution of exporting sales in individual destinations by different percentiles
in these destinations,
b) the sales of french firms in France of firms that sells in individual destinations by
different percentiles in France,
c) normalized export intensity of firms by market by different percentiles in France,
d) the fraction of firms selling to each possible combination of the top seven exporting
destinations.
These 4 set of moments contribute to the objective function
#m 2
Q (m) = ∑ wk (m) p̂k (m) − pk (m)
k =1
where p̂k (m) are the simulated observations for each moment and pk (m) the ones related
179
to the data. The authors use the following weights
wk (m) = N/pk (m)
where N is the number of firms in the data sample. With these weights each Q (m) is a chi-
square statistic with degrees of freedom given by the number of moments to be matched
(#m). Chi square is the limiting distribution of Q (m) (for N large) under the null that the
sampling error is the only source of error and, thus, observed sales follow a multinomial
distribution with the actual probabilities as parameters. Hence, the means of the Q (m)’s
equal their degrees of freedom and their variances twice their means.
The paper has a set of important contributions
• It identifies a set of statistics in the data that will be a rigorous test for all future
trade theories.
• It develops a model that is consistent with these facts and can account for different
levels of heterogeneity. In particular, it shows how the model can motivate research
to interpret and “read” the data in a way consistent to the model.
• It develops an internally consistent methodology for estimating firm-level models.
180
Chapter 12
Some facts on disaggregated trade
flows
12.1 Firm heterogeneity
• Firms appear to have huge differences in sales and measured productivities (Bernard,
Eaton, Jensen, and Kortum (2003)–BEJK–)
• In fact, only a tiny fraction of firms export to at least one market and an even smaller
fraction export to multiple destinations (only 16% of French firms sells to at least one
destination other than France, 3.3% sell to at least 10 destinations and a mere .05%
to 100 or more! See figure 14.1 drawn from Eaton, Kortum, and Kramarz (2011)).
Moreover, exporters typically earn a small fraction of their total revenues from their
exporting sales (BEJK).
• Exporters have a size advantage over non-exporters. In fact, exporters that sell to
many countries sell more in total and in the domestic market than exporters that
sell to few destinations or firms that sell only domestically (Eaton, Kortum, and
Kramarz (2004), Eaton, Kortum, and Kramarz (2011) –EKK–). This fact is illustrated
in Figure 14.1 given that the slope of the line in the plot is far less than 1 (around
181
.35): including firms less successful in exporting means less than linear increase in
total sales in France.
• The number of exporters entering a market, their average size and the total num-
ber of products sold increases with the size of the market, with an elasticity that
is roughly constant. (Klenow and Rodrı́guez-Clare (1997), Hummels and Klenow
(2005) EKK, Arkolakis and Muendler (2010)). The elasticity of entry for French ex-
porters can be seen in Figure 14.2.
• The distribution of sales of firms in a country, conditional on selling to that coun-
try, is robust across countries. It features a Pareto tail when looking at the large
firms, and large deviations from Pareto when looking at the small firms: there are
too many “too” small guys selling to each destination. Figure 14.3 illustrates the
distribution of size of firms in different destinations, grouping destinations in three
categories depending on the overall sales of French firms there.
• Firms that sell more goods sell more per good (Bernard, Redding, and Schott (2011)
and Arkolakis and Muendler (2010)). This feature is true across destinations as Fig-
ure 14.3 indicates. In fact the distribution of goods is also robust across destinations.
The above facts suggests the existence of important trade barriers, that only relatively
productive firms can overcome. In addition, the facts suggest that the costs of market
penetration have similar characteristics across markets and that the same driving forces
govern the behavior of firms.
12.2 Trade liberalization
• There is a substantial response of trade flows to price changes induced by changes in
tariffs during trade liberalizations (see for example Romalis (2007)). This response
182
is much larger than the response of trade flows to price changes over the business
cycle frequency –2-3 years–. The elasticity to changes in tariffs has been estimated in
the range of 8-10 while the one for short run adjustments around 1.5 to 2 (See Ruhl
(2009) for a review).
• A large number of new firms engage in trade after trade liberalization (see discus-
sion in Arkolakis (2010b)). Also a large number of new products are traded after
a trade liberalization (Kehoe and Ruhl (2013), Arkolakis (2010b)). New goods typi-
cally come with very small sales (Arkolakis (2010b)).
• Goods with little trade before a liberalization have higher growth rates of their trade
flows after trade liberalization. (see figure ?? and Arkolakis (2010b)).
• Trade liberalization forces the least productive firms to exit the market. (Bernard
and Jensen (1999), Pavcnik (2002), Bernard, Jensen, and Schott (2003))
The above facts on trade liberalization suggest that firms respond to short run (e.g.
exchange rate movements) changes differently than they respond to permanent changes
(e.g tariff reductions). Their response to permanent changes depends also on their initial
size. Whatever the explanation for this behavior, ultimately it should also be consistent
with the previous facts on exporting behavior of heterogeneous firms.
12.3 Trade dynamics
• A large number of firms do not export continuously to a given destination (more
than 40%). In addition a large number of new firms start exporting every year at a
given destination. These new firms and the firms that die typically have tiny sales
(Eaton, Eslava, Kugler, and Tybout (2008)).
183
• The growth rate of small exporters to a given destination is higher than the growth
rate of larger exporters (Eaton, Eslava, Kugler, and Tybout (2008), Arkolakis (2011)).
• The variance of the growth rate of small exporters to a given destination is larger
than the variance of growth of large exporters (Expected to be true: see Arkolakis
(2011) and the facts presented by Sutton (2002)).
184
Bibliography
A LLEN , T. (2012): “Information Frictions in Trade,” mimeo, Yale University. (document),
10.3.3
A LLEN , T., AND C. A RKOLAKIS (2014): “Trade and the topography of the spatial econ-
omy,” Quarterly Journal of Economics, 129(6), 1085–1140. 5.4, 6.2
A LLEN , T., C. A RKOLAKIS , AND Y. TAKAHASHI (2014): “Universal Gravity,” mimeo. 6.2,
6.2
A LVAREZ , F., AND R. E. L UCAS (2007): “General Equilibrium Analysis of the Eaton-
Kortum Model of International Trade,” Journal of Monetary Economics, 54(6), 1726–1768.
6.4, 6.4
A NDERSON , J. E. (1979): “A Theoretical Foundation for the Gravity Equation,” American
Economic Review, 69(1), 106–116. 3.3, 2, 4.3.3
A NDERSON , J. E., AND E. VAN W INCOOP (2003): “Gravity with Gravitas: A Solution to
the Border Puzzle,” American Economic Review, 93(1), 170–192. 3.3.2, 10.2.1, 10.2.4, 11
(2004): “Trade Costs,” Journal of Economic Literature, 42(3), 691–751. 11, 11.2
A RKOLAKIS , C.; C OSTINOT, A. . R.-C. (2010a): “New Trade Models, Same Old Gains?,”
NBER Working Paper, 15628. 7.1.3
185
A RKOLAKIS , C. (2010b): “Market Penetration Costs and the New Consumers Margin in
International Trade,” Journal of Political Economy, 118(6), 1151–1199. 4.5.5, 4.6, 7.1.1, 8.2,
8.2, 11.2, 11.2, 12.2
(2011): “A Unified Theory of Firm Selection and Growth,” NBER working paper,
17553. 12.3
A RKOLAKIS , C., A. C OSTINOT, D. D ONALDSON , AND A. R ODR ÍGUEZ -C LARE (2012):
“The Elusive Pro-Competitive Effects of Trade,” mimeo. 7.1.3, 8.4.1.3, 8.4.1.5
A RKOLAKIS , C., A. C OSTINOT, AND A. R ODR ÍGUEZ -C LARE (2012): “New Trade Models,
Same Old Gains?,” American Economic Review, 102(1), 94–130. 4.3.4, 6.1.2, 6.5, 6.5, 7.1,
7.1.1, 11
A RKOLAKIS , C., S. D EMIDOVA , P. J. K LENOW, AND A. R ODR ÍGUEZ -C LARE (2008): “En-
dogenous Variety and the Gains from Trade,” American Economic Review, Papers and
Proceedings, 98(4), 444–450. 3.4, 6.1, 6.1.1, 7.1
A RKOLAKIS , C., AND M.-A. M UENDLER (2010): “The Extensive Margin of Exporting
Products: A Firm-Level Analysis,” NBER Working Paper, 16641. 8.3, 3, 11.2, 12.1
A RKOLAKIS , C., AND A. R AMANARAYANAN (2008): “Vertical Specialization and Interna-
tional Business Cycles Synchronization,” Manuscript, Federal Reserve Bank of Dallas
and Yale University. 9.3
A RMINGTON , P. S. (1969): “A Theory of Demand for Products Distinguished by Place of
Production,” International Monetary Fund Staff Papers, 16, 159–178. 3.3, 6.1, 6.5
ATKESON , A., AND A. B URSTEIN (2010): “Innovation, Firm Dynamics, and International
Trade,” Journal of Political Economy, 118(3), 433–489. 7.1.2, 7.1.3, 7.2
186
B ALDWIN , R., AND D. TAGLIONI (2006): “Gravity for dummies and dummies for gravity
equations,” Discussion paper, National Bureau of Economic Research. 10.2.1
B EHRENS , K., G. M ION , Y. M URATA , AND J. S UDEKUM (2009): “Trade, Wages, and Pro-
ductivity,” CEP Discussion Papers, 0942. 8.4.1.3
B EHRENS , K., AND Y. M URATA (2009): “Globalization and Individual Gains from Trade,”
CEPR discussion paper, 7438. 8.4.1.3
B ERNARD , A., B. J ENSEN , S. R EDDING , AND P. J. S CHOTT (2007): “FIrms in International
Trade,” Journal of Economic Perspectives, 21(3), 105–130. 4.5
B ERNARD , A. B., J. E ATON , J. B. J ENSEN , AND S. K ORTUM (2003): “Plants and Productiv-
ity in International Trade,” American Economic Review, 93(4), 1268–1290. 4.3.5, 4.6, 6.1,
12.1
B ERNARD , A. B., AND J. B. J ENSEN (1999): “Exceptional Exporter Performance: Cause,
Effect, or Both,” Journal of International Economics, 47, 1–25. 12.2
B ERNARD , A. B., J. B. J ENSEN , AND P. K. S CHOTT (2003): “Falling Trade Costs, Heteroge-
neous Firms and Industry Dynamics,” NBER Working Paper, 9639. 12.2
B ERNARD , A. B., S. J. R EDDING , AND P. S CHOTT (2011): “Multi-Product Firms and Trade
Liberalization,” Quarterly Journal of Economics, 126(3), 1271–1318. 12.1
B HAGWATI , J. N., A. PANAGARIYA , AND T. N. S RINIVASAN (1998): Lectures on Interna-
tional Trade. The MIT Press, Boston, Massachussetts. 2
B RODA , C., AND D. W EINSTEIN (2006): “Globalization and the Gains from Variety,” Quar-
terly Journal of Economics, 121(2), 541–585. 11.2
187
B UTTERS , G. (1977): “Equilibrium Distributions of Sales and Advertising Prices,” The
Review of Economic Studies, 44(3), 465–491. 8.2
C HANEY, T. (2008): “Distorted Gravity: The Intensive and Extensive Margins of Interna-
tional Trade,” American Economic Review, 98(4), 1707–1721. 5, 4.5.3, 4.5.4, 6.1, 6.1.1, 7.1.1,
8.2
D EKLE , R., J. E ATON , AND S. K ORTUM (2008): “Global Rebalancing with Gravity: Mea-
suring the Burden of Adjustment,” IMF Staff Papers, 55(3), 511–540. 5.1, 6.5
D HINGRA , S., AND J. M ORROW (2012): “The Impact of Integration on Productivity and
Welfare Distortions Under Monopolistic Competition,” mimeo, LSE. 8.4.1.3
D ONALDSON , D. (2014): “Railroads of the Raj: Estimating the Economic Impact of
Transportation Infrastructure,” forthcoming, American Economic Review. (document),
10.3.2, 10.3.2, 10.3.3
D ORNBUSCH , R. S., S. F ISCHER , AND P. A. S AMUELSON (1977): “Comparative Advan-
tage, Trade, and Payments in a Ricardian Model with a Continuum of Goods,” American
Economic Review, 67(5), 823–839. 4.1, 4.3, 9.2
E ATON , J., M. E SLAVA , M. K UGLER , AND J. T YBOUT (2008): “The Margins of Entry Into
Export Markets: Evidence from Colombia,” in Globalization and the Organization of Firms
and Markets, ed. by E. Helpman, D. Marina, and T. Verdier. Harvard University Press,
Massachusetts. 12.3
E ATON , J., AND S. K ORTUM (2002): “Technology, Geography and Trade,” Econometrica,
70(5), 1741–1779. (document), 4, 4.1.1, 4.3, 4.3.1, 4.3.1, 4.3.3, 4.3.3, 4.3.4, 4.6, 6.1, 6.3, 6.5,
7.1.1, 8.2, 9.1, 1, 9.3, 9.3, 9.3, 10.3.1, 10.3.1, 2, 10.3.3, 11.1.1, 1, 11.1.1, 11.2, 11.3
188
(2011): Technology in the Global Economy: A Framework for Quantitative Analysis.
Manuscript, Penn State Univesity and Yale University. 1.3, 4, 1
E ATON , J., S. K ORTUM , AND F. K RAMARZ (2004): “Dissecting Trade: Firms, Industries,
and Export Destinations,” American Economic Review, Papers and Proceedings, 94(2), 150–
154. 12.1
(2011): “An Anatomy of International Trade: Evidence from French Firms,” Econo-
metrica, 79(5), 1453–1498. (document), 4.5.5, 4.6, 8.2, 11.1.1, 11.3, 11.3.1, 12.1, 14.1, 14.2,
14.3
FAN , H., E. L AI , AND H. Q I (2013): “Global Gains from Reduction of Trade Costs,” Un-
published manuscript. 7.2
F EENSTRA , R. (2003): Advanced International Trade. Princeton University Press, Princeton,
New Jersey. 2
G ROSSMAN , G. M., AND C. S HAPIRO (1984): “Informative Advertising with Differenti-
ated Products,” The Review of Economic Studies, 51(1), 63–81. 8.2
H EAD , K., AND T. M AYER (2013): Gravity equations: Workhorse, toolkit, and cookbook. Centre
for Economic Policy Research. 10.2.3
H EAD , K., AND J. R IES (2001): “Increasing Returns versus National Product Differentia-
tion as an Explanation for the Pattern of U.S.-Canada Trade,” American Economic Review,
91(4), 858–876. 10.1
H ELPMAN , E. (1998): “The Size of Regions,” Topics in Public Economics. Theoretical and
Applied Analysis, pp. 33–54. 6.2
189
H ELPMAN , E., O. I TSKHOKI , AND S. J. R EDDING (2010): “Inequality and Unemployment
in a Global Economy,” Econometrica, 78(4), 1239–1283. 4.5.5, 4.6
H ELPMAN , E., M. M ELITZ , AND Y. R UBINSTEIN (2008): “Estimating Trade Flows: Trading
Partners and Trading Volumes,” Quarterly Journal of Economics, 2(5), 441–487. 4.5.5, 4.6
H ELPMAN , E., M. J. M ELITZ , AND S. R. Y EAPLE (2004): “Export Versus FDI with Hetero-
geneous Firms,” American Economic Review, 94(1), 300–316. 4.5.5, 4.6
H UMMELS , D., AND P. K LENOW (2005): “The Variety and Quality of a Nation’s Exports,”
American Economic Review, 95(3), 704–723. 12.1
K EHOE , T. J., AND K. J. R UHL (2013): “How Important Is the New Goods Margin in
International Trade?,” Journal of Political Economy, 121(2), 358 – 392. 12.2
K LENOW, P. J., AND A. R ODR ÍGUEZ -C LARE (1997): “Quantifying Variety Gains from
Trade Liberalization,” Manuscript, University of Chicago. 12.1
K ORTUM , S. (1997): “Research, Patenting, and Technological Change,” Econometrica, 65(6),
1389–1419. 4, 4.3.1
K RUGMAN , P. (1979): “Increasing Returns Monopolistic Competition and International
Trade,” Journal of International Economics, 9(4), 469–479. 8.4.1
K RUGMAN , P. (1980): “Scale Economies, Product Differentiation, and the Pattern of
Trade,” American Economic Review, 70(5), 950–959. 3.4, 4.3.1, 4.3.1, 4.3.3, 4.3.4, 4.5, 4.5.1,
4.5.1, 4.5.2, 4.5.2, 4.5.2, 4.5.4, 4.5.4, 5, 5.2, 5.3
K RUGMAN , P., AND A. J. V ENABLES (1995): “Globalization and the Inequality of Na-
tions,” Quarterly Journal of Economics, 110(4), 857–880. 9.1
190
K YDLAND , F. E., AND E. C. P RESCOTT (1994): “The Computational Experiment: An
Econometric Tool,” Federal Reserve Bank of Minneapolis Staff Report, 178. 1.2
M ARTINS , J. O., S. S CARPETTA , AND D. P ILAT (1996): “Mark-Up Ratios in Manufacturing
Industries: Estimates for 14 OECD Countries,” OECD Economics Department Working
Paper, 162. 11.2
M AS -C OLELL , A., M. D. W HINSTON , AND J. R. G REEN (1995): Microeconomic Theory.
Oxford University Press, Oxford, UK. 6.4, 8
M C FADDEN, D. (1974): “Conditional Logit Analysis of Qualitative Choice Behavior,”
Frontiers in Econometrics, pp. 105–142. 4.3
M ELITZ , M. J. (2003): “The Impact of Trade on Intra-Industry Reallocations and Aggre-
gate Industry Productivity,” Econometrica, 71(6), 1695–1725. 4.3.1, 4.3.3, 4.3.5, 4.5, 4.5.1,
5, 4.5.5, 4.6, 5.2, 5.3, 6.1, 6.1.1, 7.1, 7.1.1, 8.2, 8.3, 8.3, 8.3.1
M ELITZ , M. J., AND G. I. P. O TTAVIANO (2008): “Market Size, Trade, and Productivity,”
The Review of Economic Studies, 75(1), 295–316. 4.5.5, 4.6
M ELITZ , M. J., AND S. J. R EDDING (2014): “New Trade Models New Welfare Gains,”
American Economic Review, 105(3), 1105–46. 7.1.3, 7.1.3
PAVCNIK , N. (2002): “Trade Liberalization, Exit, and Productivity Improvements: Evi-
dence from Chilean Plants,” The Review of Economic Studies, 69(1), 245–276. 12.2
P RESCOTT, E. C. (1998): “Business Cycle Research: Methods and Problems,” Federal Re-
serve Bank of Minneapolis Working Paper, 590. 1.1, 1.2
R EDDING , S. J. (2012): “Goods Trade, Factor Mobility and Welfare,” mimeo. 6.2
191
R OMALIS , J. (2007): “NAFTA’s and CUFTA’s Impact on International Trade,” Review of
Economics and Statistics, 89(3), 416–435. 11.2, 12.2
R OMER , P. (1994): “New Goods, Old Theory, and the Welfare of Trade Restrictions,” Jour-
nal of Development Economics, 43(1), 5–38. 3, 7.1
R UHL , K. J. (2009): “The Elasticity Puzzle in International Economics,” Manuscript, New
York University. 12.2
S AURE , P. (2009): “Bounded Love of Variety and Patterns of Trade,” mimeo, Swiss National
Bank. 8.4.1.3
S IMONOVSKA , I. (2009): “Income Differences and Prices of Tradables,” Manuscript, Uni-
versity of California, Davis. 8.4.1.3
S IMONOVSKA , I., AND M. WAUGH (2013): “The Elasticity of Trade: Estimates and Evi-
dence,” Journal of International Economics, 92, 34–50. 10.3.2, 11.1.1
S UTTON , J. (2002): “The Variance of Firm Growth Rates: ’The Scaling Puzzle’,” Physica A,
312(3), 577–590. 12.3
WAUGH , M. (2010): “International Trade and Income Differences,” forthcoming, American
Economic Review. 6.3
W ILSON , C. A. (1980): “On the General Structure of Ricardian Models with a Continuum
of Goods: Applications to Growth, Tariff Theory, and Technical Change,” Econometrica,
48(7), 1675–1702. 4.3
Y I , K.-M. (2003): “Can Vertical Specialization Explain the Growth of World Trade?,” Jour-
nal of Political Economy, 111(1), 52–102. 9.2
192
Z HELOBODKO , E., S. K OKOVIN , M. PARENTI , AND J.-F. T HISSE (2011): “Monopolistic
Competition in General Equilibrium: Beyond the CES,” mimeo. 8.4.1.3
193
Chapter 13
Appendix
13.1 Distributions
This appendix explains the details of the two main distributions used in these notes1 .
13.1.1 The Fréchet Distribution
The type II extreme value distribution, also called the Fréchet distribution, is one of three
distributions that can arise as the limiting distribution of the maximum of a sequence of
independent random variables. The distribution function for the Fréchet distribution is
( )
x − µ −θ
F ( x ) = exp − ,
σ
for x > µ, where θ > 0 is a shape parameter, σ > 0 is a scale parameter and µ ∈ R is a
location parameter. The density of the Fréchet distribution is
( )
θ x − µ − θ −1 x − µ −θ
f (x) = exp − ,
σ σ σ
1 Many thanks to Alex Torgovitsky for the preparation of this appendix
194
for x > µ. If X is a Fréchet-distributed random variable then
ˆ ∞ ( )
θ x − µ − θ −1 x − µ −θ
E( X ) = x exp − dx
µ σ σ σ
ˆ ∞ ˆ ∞
− 1θ −y
=σ y e dy + µ e−y dy
0 0
θ−1
= σΓ + µ,
θ
−θ
x −µ
where y := and
σ
ˆ ∞
Γ(z) = tz−1 e−t dt
0
is the Gamma function. Now assume that µ = 0, take T := σθ and rewrite the distribution
function as
−θ
F ( x ) = e− Ax ,
so that the Fréchet distribution is now parameterized by θ, A. Notice that for any given
θ and A is increasing in the scale parameter, σ. Figure ??, shows how θ and A affect the
Fréchet distribution.
The Pareto Distribution
The Pareto distribution is parameterized by a shape parameter, θ > 0, a scale parameter
m > 0 and has support [m, ∞) with distribution function
m θ
F(x) = 1 − .
x
The density function is
θmθ
f (x) = .
x θ +1
195
The nth moment of a Pareto distributed random variable can easily be calculated as
ˆ ∞ θmn , if θ > n
n n θ − θ −1 θ −n
E( X ) = x θm x dx =
m +∞, if θ ≤ n
which shows that the shape parameter controls the number of existent moments. Direct
computation yields
θm
E( X ) = , if θ > 1,
θ−1
θm2
Var ( X ) = , if θ > 2.
( θ − 1)2 ( θ − 2)
The Pareto distribution is an example of a power law distribution, which can be seen
by observing that
m θ
Pr [ X ≥ x ] = .
x
This implies that
log (Pr [ X ≥ x ]) = θ log(m) − θ log( x ),
so that the log of the mass of the upper tail past x is linear in log( x ). For example, if
the number of employees in a randomly sampled firm, X, is Pareto distributed, then the
proportion of firms in the population that have more than x employees is linear with the
number of employees on a log-log scale. This is related to a useful self-replicating feature
of the Pareto distribution, which is that the distribution of X conditional on the event
[ X ≥ x̄ ], where x̄ ≥ m, is given by
θ
Pr [ X ≥ x ] x̄
Pr [ X ≥ x | X ≥ x̄ ] = = ,
Pr [ X ≥ x̄ ] x
for x ≥ x̄. That is, truncating the Pareto distribution on the left produces another Pareto
distribution with the same shape parameter! Figure ??, shows how θ and the initial point
196
m affect the Pareto distribution.
197
Chapter 14
Figures and Tables
198
Figure 14.1: Sales in France from firms grouped in terms of the minimum number of
destinations they sell to. Source: Eaton, Kortum, and Kramarz (2011).
199
Figure 14.2: French entrants and market size. Source: Eaton, Kortum, and Kramarz (2011).
DistributionsData.wmf XXX need to replace XXX
Figure 14.3: Distribution of sales for Portugal and means of other destinations group in
terciles depending on total sales of French firms there. Each box is the mean over each
size group for a given percentile and the solid dots are the sales distribution in Portugal.
Source Eaton, Kortum, and Kramarz (2011).
200
You might also like
- (EMA) Prato G.D., Zabczyk J. - Stochastic Equations in Infinite Dimensions-CUP (2014)100% (1)(EMA) Prato G.D., Zabczyk J. - Stochastic Equations in Infinite Dimensions-CUP (2014)514 pages
- Elements of Advanced International Trade 贸易理论总结No ratings yetElements of Advanced International Trade 贸易理论总结210 pages
- Econ 702 Macroeconomic Theory Spring 2016No ratings yetEcon 702 Macroeconomic Theory Spring 201659 pages
- MACROECONOMICS For PHD (Dynamic, International, Modern, Public Project)No ratings yetMACROECONOMICS For PHD (Dynamic, International, Modern, Public Project)514 pages
- Intro To Economic Growth - Daron Acemoglu PDFNo ratings yetIntro To Economic Growth - Daron Acemoglu PDF453 pages
- Lecture Notes For Econ 101A: David Card Dept. of Economics UC BerkeleyNo ratings yetLecture Notes For Econ 101A: David Card Dept. of Economics UC Berkeley162 pages
- Real Macroeconomic Theory Per Krussel March 2014No ratings yetReal Macroeconomic Theory Per Krussel March 2014277 pages
- Physic of Socioeconomic System (Cours Polytechnique)No ratings yetPhysic of Socioeconomic System (Cours Polytechnique)47 pages
- Contents 28-09-2014-MasterBook-pdf-M2-2014-2No ratings yetContents 28-09-2014-MasterBook-pdf-M2-2014-214 pages
- Lecture Notes in Macroeconomics: Christian Groth August 30, 2015No ratings yetLecture Notes in Macroeconomics: Christian Groth August 30, 2015124 pages
- Midea Air Conditioning Units Installation GuideNo ratings yetMidea Air Conditioning Units Installation Guide12 pages
- Development of Vehicle Chassis From Novel Materials For Light Weight Electric Shuttles Using Finite Element AnalysisNo ratings yetDevelopment of Vehicle Chassis From Novel Materials For Light Weight Electric Shuttles Using Finite Element Analysis13 pages
- Object Detection in Pytorch Using Mask R-CNNNo ratings yetObject Detection in Pytorch Using Mask R-CNN4 pages
- Half Yearly 2022 Answer Key With Masrking Scheme Sample Paper 1No ratings yetHalf Yearly 2022 Answer Key With Masrking Scheme Sample Paper 115 pages
- Eurocode EN 1998-2 - Seismic Design of Bridges PDFNo ratings yetEurocode EN 1998-2 - Seismic Design of Bridges PDF45 pages
- Lab Manual Electronic Devices and Circuits Practical 2No ratings yetLab Manual Electronic Devices and Circuits Practical 210 pages
- Microstrip Patch Antenna Based UHF Sensor For Detection of Partial Discharge in High Voltage Electrical EquipmentsNo ratings yetMicrostrip Patch Antenna Based UHF Sensor For Detection of Partial Discharge in High Voltage Electrical Equipments6 pages
- DBMS - Unit 4 - Question Bank - Reg 2021No ratings yetDBMS - Unit 4 - Question Bank - Reg 202116 pages
- Properties of Saturated Steam - Pressure in BarNo ratings yetProperties of Saturated Steam - Pressure in Bar6 pages
- Materials: Electrochemical Polishing of Austenitic Stainless SteelsNo ratings yetMaterials: Electrochemical Polishing of Austenitic Stainless Steels25 pages
- Chapter 9: The Biomechanics of The Human Spine100% (1)Chapter 9: The Biomechanics of The Human Spine6 pages
- Internship Presentation Professional Training (Level-2) Submitted in Partial Fulfillment of Requirements For The Award of The Degree ofNo ratings yetInternship Presentation Professional Training (Level-2) Submitted in Partial Fulfillment of Requirements For The Award of The Degree of15 pages
