productivity, quality and exporting behavior under minimum...
TRANSCRIPT
Productivity, Quality and Exporting Behavior under Minimum
Quality Requirements
Juan Carlos Hallak∗ and Jagadeesh Sivadasan†
First draft May 2006
Revised April 2007
Preliminary and Incomplete – Comments Welcome
Abstract
We develop a model of industry equilibrium with heterogeneous firms that explicitly incor-
porates quality and productivity in one framework. In our model, productivity reduces marginal
costs, whereas quality shifts out demand and increases marginal and fixed costs. We propose
that to export firms need to cross a threshold quality level. We examine exporting behavior of
manufacturing establishments in Chile, Colombia and India and find that our model’s predic-
tions are consistent with a number of regularities in the data relating exporting status to size,
output prices, capital intensity, skill intensity and a number of proxies for quality.
JEL codes: F10, F12, F14,
Keywords: Productivity, quality, exports, firm heterogeneity∗University of Michigan and NBER, email: [email protected].†University of Michigan, email: [email protected].
1
1 Introduction
The field of international trade is witnessing a radical change in paradigm about what determines
export performance. In contrast to the exclusive focus placed on variation in export behavior
across sectors since the work of Ricardo, a growing new literature focuses on the variation in
export behavior across firms within sectors. Among the most important theoretical results of this
literature [e.g. Melitz (2003)] is the prediction that there are systematic differences between firms
in productivity, size, and export status, even when the firms operate in the same sector. This
prediction matches overwhelming empirical evidence showing that more productive firms are larger
and are more active in export markets [Clerides, Lach and Tybout (1998) and Bernard and Jensen
(1999)]. Almost universally, this literature points at differences in productivity as the sole source
of heterogeneity in firms’ export behavior.
Despite the literature’s emphasis on productivity differences, an increasingly held view is that
firms’ ability to produce quality is also key for explaining export success [Rauch (2006)]. Quality
and productivity are intimately related. While high productivity is typically understood to be
the ability to produce equal quantities with lower costs, it could alternatively be thought of as
the ability to produce higher quality with equal costs.1 The latter view is modelled in Verhoogen
(2004), where firms’ different ability to produce quality is the only source of heterogeneity. In that
model, as in models with standard productivity differences, more “productive” firms (those more
able to produce high quality) are larger and are those that enter the export market.
Quality and productivity, however, need not go together. Firms producing high quality goods
might have low productivity if their costs are sufficiently high while firms producing low quality
goods might have high productivity if their costs are sufficiently low. Despite the absence of a
one-to-one correspondence between quality and productivity, models that highlight either standard
productivity or ability to produce quality as alternative sources of firm heterogeneity generate
similar predictions for the relationship between (broadly-defined) productivity, size, and export
status. A looming question is then whether the simultaneous presence of these two sources of
heterogeneity generates qualitatively different predictions. We show in the paper that the answer
to this question depends on the trading environment that is assumed. In particular, we analyze
trading environments without and with minimum export requirements. The first is a benchmark1This argument is made, for example, by Melitz (2000).
2
case. In the second case, we find theoretical predictions that, even though broadly consistent with
previous theoretical results, characterize more closely empirical regularities that are present in the
data.
The paper builds a model with two sources of heterogeneity. The first source of heterogeneity is
standard; it is the ability to produce output with low marginal costs. Following the literature, we
term such ability “productivity” (ϕ). The second source of heterogeneity is the ability to produce
high quality at low cost, which we term “caliber” (ξ). High quality in the model benefits the firm by
shifting out the demand for the firm’s output. Quality, on the other hand, increases marginal costs
of production and fixed costs of quality development. High caliber allows firms to produce high
quality paying a lower fixed cost.2 Even though the caliber of a firm is the primary determinant
of its quality choice, productivity is also important as it reduces the impact of quality on marginal
costs. Therefore, both caliber and productivity increase the level of quality that is optimally chosen.
First, we solve for industry equilibrium in a closed economy. The model has two sources of
heterogeneity, productivity (ϕ) and caliber (ξ). In the closed economy, the two can be lumped
together in a single “combined productivity” parameter (η), a scalar funcion of the underlying
productivity and caliber factors (η = η(ϕ, ξ)). The equilibrium and all key variables of interest (such
as firm revenues and profits) can then be characterized in terms of this single parameter – although
firms with different combinations of ϕ and ξ that result in the same η choose different quality levels.
Furthermore, in the closed economy the model allows for a representation that is isomorphic to
Melitz’s model. To survive in equilibrium, firms need to possess combined productivity above a
certain cutoff level η∗.
Next, we examine the open economy. We start by considering a case in which only two obstacles
constrain firms’ export activity: fixed costs of exporting (fx) and “iceberg” transport costs (τ). As
was the case in the closed economy, here also the model can be collapsed into a model with only
one source of heterogeneity – isomorphic to Melitz’ – making this case a transparent benchmark for
evaluating the different implications of caliber (the ability to produce high quality) and standard
productivity. However, the isomorphism disappears once we include more realistic features of the2This approach to modelling quality production captures the trade-offs in Yeaple (2005) and Bustos (2005), where
adoption of high technology requires firms to incur a fixed cost. In those models, investing in high technology shifts
down the marginal cost, which is isomorphic – under certain assumptions on demand – to shifting out the demand
curve from the perspective of the firm.
3
relationship between quality production and export behavior. In particular, this is true in the case
we analyze, where firms need to satisfy minimum quality requirements in order to export.
The assumption of a minimum quality export requirement, although a drastic simplication, is
based on a wealth of anecdotal evidence documenting firms’ need to upgrade quality to enter export
markets and on empirical evidence showing that exporters produce higher quality than domestic
firms [Brooks (2006) and Verhoogen (2004)]. While the empirical literature emphasizes the demand
for higher quality products of rich countries, several alternative explanations can motivate this
assumption. First, as conjectured by Alchian and Allen (1964) and showed by Hummels and Skiba
(2004), transportation costs are proportionally lower for high quality goods. The assumption of
a minimum quality for export reduces this fact to the existence of a quality level below which
transport costs are prohibitive. Second, low quality might lead to haggling and negotiations over
defective items, which are costlier to undertake in international transactions. Firms might refrain
to engage in trade of low quality goods if they anticipate a sufficiently high probability of returns.
Third, some export markets directly impose minimum quality standards. Fourth, in the presence of
incomplete contracts and given costs of international monitoring, high quality standards (e.g. ISO
9000) might be a way to prevent cheating on the contract, as those standards commit the firm to
established procedures. It is not the purpose of the paper to identify the relative importance of these
different explanations. Rather, the aim is to analyze how the usual predictions of firm-heterogeneity
models are modified in a context in which minimum quality requirements are present.
Once a minimum quality export requirement is imposed, productivity and caliber become dis-
tinct characteristics of a firm that cannot perfectly substitute for one another. Conditional on the
same value of combined productivity η, firms with lower caliber (and hence higher productivity)
are inclined to choose lower quality levels. Those firms are forced to choose between upgrading
their quality to meet the requirements of the export market or instead focus on the domestic mar-
ket. Therefore, firms that would have been of equal size (i.e. revenue) in the closed economy (or
were so at a pre-liberalization stage) might differ in size and export status in the open economy, a
prediction that departs from the standard prediction in theoretical models that size in the closed
economy is a perfect predictor of export status.3
3In the extremes, the results of our model are standard. If the productivity (ϕ) and caliber (ξ) parameter
realizations are both low, then the firm either exits or serves only the domestic market. If both parameters are high,
4
Similarly, the model demonstrates that, again in contrast to the predictions of most theoretical
models, size in the open economy does not perfectly predict export status either. A firm that is
large in the domestic market but does not export might have a volume of sales equal to another
firm that sells smaller volumes in the domestic market but is able to access foreign markets. The
former is a firm with high ϕ and low ξ. The latter is a firm with low ϕ and high ξ. In the firm-level
datasets for Chile, Colombia, and India that we examine, we observe that exporting behavior is
not perfectly explained by firm size. Although the fraction of exporters is higher for larger size
percentiles, there are significant number of exporters in almost all size percentiles and a significant
number of non-exporters even among the largest firms.
The non-sufficiency of size for explaining export status is the central piece of our theoretical
results. Conditional on size, we find that exporters sell products of higher quality. This predic-
tion is substantially different from the prediction that exporters sell products of higher quality
unconditionally. In the latter case, size is sufficient for explaining export status; information on
quality levels, for example, would be redundant. In contrast, our model predicts that, conditional
on size, the probability of exporting increases with quality. In addition, to the extent that higher
quality products are more expensive and require a more intensive use of capital and skilled-labor,
the model also predicts that, conditional on size, the probability of exporting increases with price,
the capital-labor ratio and the skilled-labor share of the labor force. Those are the predictions that
we take to the data.
We use firm-level data for manufacturing in three developing countries: India, Chile and Colom-
bia. For Chile and Colombia, we have data from census surveys of all manufacturing plants that
employ more than 10 workers. These datasets include information on output, inputs and exports.
For India, we have a unique dataset covering all manufacturing plants for the year 1998. In addi-
tion to data on inputs, output and exports also available for Chile and Colombia, the India dataset
includes information on price per unit of output for each of the product lines produced by a plant.
Our analysis of the data shows that, consistent with the predictions of our model, exporters
are more capital intensive, more skilled-labor intensitve and charge higher prices even after size is
controlled for, a reflection of the fact that they export higher quality goods. These results imply
that quality has significant explanatory power, in addition to size, for explaining export status.
then they “naturally” choose a quality level above the threshold and serve both domestic and foreign markets.
5
The findings are remarkably consistent across countries.
The rest of the paper is organized as follows. Section 2 describes our theoretical model. Section
3 describes the data. Section 4 presents our empirical results. Section 5 concludes.
2 Productivity and quality in a two-factor heterogenous-firm model
This section develops a two-factor heterogenous-firm model of industry equilibrium. In Section 2.1,
we characterize the equilibrium in a closed economy. In Section 2.2.1, we examine the case of an
open economy with no quality constrains on exports. In both of these cases, we show that the
model can be collapsed to a model with only one source of heterogeneity. Finally, in Section 2.2.2,
we consider the case in which exports require the attainment of minimum quality levels.
2.1 The closed economy
The model is developed in partial equilibrium. We assume a monopolistic competition framework
with constant-elasticity-of-substitution (CES) demand. The demand system here is augmented to
account for product quality variation across varieties (as in Hallak and Schott 2005):
Dj =p−σ
j λσ−1j
(P ′)1−σ E (1)
where j indexes products, and pj and λj are, respectively, the price and quality of product j. We
assume that each firm produces only one product; therefore, j also indexes firms. Finally, E is the
(exogenously given) level of expenditure and
P ′ =
∫
j
p1−σj λσ−1
j dj
11−σ
is a cost-of-utility price index. For notational simplicity, we define P = (P ′)1−σ. 4
Product quality is modelled here as a demand shifter. It captures all attributes of a product
– other than price – that consumers value. This demand system solves a consumer maximization
problem with a Dixit-Stiglitz utility function defined in terms of quality-adjusted units of consump-
tion, q̃j = qjλj , and quality adjusted prices p̃j = pj
λj. Thus, firm revenues, rj = pjqj = p̃j q̃j , can be
4It is important to keep in mind for the remainder of the paper that, since σ > 1, P is inversely related to the
price index. Therefore, higher P implies lower (average) prices.
6
expressed as
rj =p̃j
1−σ
PE. (2)
There are two sources of firm heterogeneity. Following standard models (Melitz 2003, Bernard
et al. 2003), a first source of heterogeneity is “productivity” (ϕ). Productivity indexes firms’
ability to produce output with low marginal costs. In addition, we introduce a second source of
heterogeneity, which we denote “caliber” (ξ). Caliber indexes firms’ ability to develop a high quality
product paying a low fixed cost. We follow Shaked and Sutton (1983) in assuming that attaining
higher quality levels requires paying higher fixed costs.5
Marginal costs are assumed to be constant in the scale of production. They are also increasing
in the quality (λ) of the product and decreasing in the productivity (ϕ) of the firm. Marginal costs
are given by
c(λ, ϕ) =c
ϕλβ (3)
where c is a constant parameter. Fixed costs are increasing in the quality (λ) of the product and
decreasing in the caliber (ξ) of the firm. They are given by
F (λ, ξ) =f
ξλα + F (4)
where f is a constant parameter and F is a fixed cost of plant operation.
2.1.1 Firm’s optimal choice of price and quality
Firms choose price and quality to maximize profits, which are given by
Π (ϕ, ξ) = D(p− c(λ))− F (λ) = p−σλσ−1 E
P
(p− c
ϕλβ
)− f
ξλα − F.
The first order condition with respect to price yields the standard CES result of a constant mark-up
over marginal cost:
p =σ
σ − 1c(λ) =
σ
σ − 1c
ϕλβ. (5)
5Verhoogen (2004) also considers heterogeneity in the ability to produce higher quality, but this heterogeneity only
affects variable costs. As a result, this source of heterogeneity plays essentially the same role as typical productivity
heterogeneity (except for the fact that quality has a differential impact on demand according to importer per-capita
income).
7
Using this result and the first order condition with respect to quality yields:
λ(ϕ, ξ) =[1− β
α
ξ
f
(σ − 1
σ
)σ (ϕ
c
)σ−1 E
P
] 1α′
(6)
where α′ = α − (1− β) (σ − 1). To ensure that the first order conditions identify a maximum,
we need to impose two restrictions on parameter values. First, we assume that 0 < β < 1, so
that marginal costs are increasing in quality but not excessively fast. Second, we assume that
α′ > 0, so that fixed costs grow fast enough with quality. Equation (6) indicates that both caliber
(ξ) and productivity (ϕ) have a positive effect on quality choice; caliber and productivity reduce,
respectively, the fixed costs and marginal costs of producing quality.
Substituting the result for λ into the solution for price in equation (5) results in:
p(ϕ, ξ) =(
σ
σ − 1
)α−β−(σ−1)
α′(
c
ϕ
)α−(σ−1)
α′[1− β
α
ξ
f
E
P
] βα′
. (7)
Equation (7) shows that high caliber firms sell their products at a higher price, as they produce
higher quality. The effect of productivity on price, however, is ambiguous, depending on the sign
of the term α− (σ − 1). On the one hand, higher productivity induces lower prices conditional on
quality, as in all standard models. On the other hand, higher productivity induces higher quality,
which in turn increases production costs and prices. Whether one or the other effect dominates
depends on parameter values. It is interesting to note that, as opposed to models with only one
source of heterogeneity, productivity and prices here do not display a monotonic relationship. Since
prices depend on the values of two parameters, the correlation between productivity and prices
observed in the data will depend on the correlation between the productivity and caliber draws.
2.1.2 The cut-off function
We can substitute the solutions for quality and price in equations (6) and (7), respectively, into the
expression for revenues (2) to obtain:
r(ϕ, ξ) =(
σ − 1σ
) (ασ−α′)α′
(E
P
) αα′ (ϕ
c
)α(σ−1)
α′[1− β
α
ξ
f
]α−α′α′
. (8)
Revenues increase with both productivity (ϕ) and caliber (ξ). As with price, the relationship
between revenues and any one of these variables alone is not necessarily monotonic. A convenient
8
way of summarizing information about the two heterogeneity draws is to define a firm’s “combined”
productivity η as
η(ϕ, ξ) =[ϕ
αα′ ξ
1−βα′
]σ−1.
Revenues can then be expressed as a function of η:
r(ϕ, ξ) = η(ϕ, ξ)(
σ − 1σ
)−(ασ−α′)α′
(E
P
) αα′
(1c
)α(σ−1)
α′[1− β
αf
]α−α′α′
. (9)
Relative revenues (size) between any two firms i and j is given by the ratio of their combined
productivities: rirj
= ηi
ηj. From standard results of CES demand, we know that operative profits
equal rσ . Therefore, Π(ϕ, ξ) = 1
σ r(ϕ, ξ) − F (ξ, λ). Using this expression and the expression for
revenues in equation (8) we obtain firm profits:
Π(ϕ, ξ) = J(ϕ
c
)α(σ−1)
α′(
ξ
f
)α−α′α′
(E
P
) αα′ − F (10)
where J =(
σ−1σ
)ασα′
(1−βα
) αα′
(α′
α−α′
).
Profits, as revenues, increase with productivity and caliber and can also be written as a function
of η:
Π(ϕ, ξ) = η(ϕ, ξ)J(
1c
)α(σ−1)
α′(
1f
)α−α′α′
(E
P
) αα′ − F. (11)
Since combined productivity η is a summary statistic for both revenues and profits, it follows
that in a graph on ϕ–ξ space, iso-revenue curves coincide with iso-profit curves.
Firms remain in the market only if they can make non-negative profits, given their productivity
and caliber draws. Since profits depend on these two sources of heterogeneity, this condition results
in a cut-off function, ξ(ϕ); for each productivity level, there is a minimum caliber that allows a
firm to make non-negative profits. The cut-off function is easy to derive by equating (10) to zero:
Π(ϕ, ξ) = 0 =⇒
ξ (ϕ) = f
(F
J
) α′α−α′ (ϕ
c
) −α1−β
(E
P
) −αα−α′
. (12)
The most important feature of (12) is that ξ(ϕ) is a negative function of ϕ. Therefore, more
productive firms can afford to be of lower caliber (and vice versa). This negative relationship is
easy to understand as an implication of the fact that combined productivity η is a summary statistic
9
for the effect of the two heterogeneity factors on profits. In fact, given this property, the model
can be collapsed into a one-factor model, iso-morphic to Melitz’, where a single productivity draw
determines entry-exit decisions: firms stay iff η is above a cut-off value η, determined by equating
(11) to zero.
The solutions for price, quality, revenues, profits, and cut-off function all depend on P , which is
an endogenous variable. To complete the characterization of the equilibrium, we need to determine
P using the free-entry condition.
2.1.3 The price index and the free-entry condition
Before entering, firms do not know their productivity or caliber. They have to pay a fixed entry cost
of fe > 0 to learn them. Once they pay this cost, they draw ϕ and ξ from a bivariate probability
distribution with density v (ϕ, ξ). Denote the probability of surviving by Pin. This probability is
given by
Pin =
∞∫
0
∞∫
ξ(ϕ)
v (ϕ, ξ) dξdϕ (13)
The joint distribution of ϕ and ξ conditional on surviving can then be characterized by the density
function
h(φ, ξ) =
1Pin
v(ϕ, ξ) if ξ > ξ(ϕ)
0 otherwise
(14)
Denoting by M the mass of surviving firms, we can now rewrite P as the aggregation across
productivity and (surviving) caliber levels instead of across firms:
P =∫
j
p1−σj λσ−1
j dj = M
∫
ϕ
∫
ξ(ϕ)
r (ϕ, ξ)(
E
P
)−1
h (ϕ, ξ) dξdϕ.
Using (8) to substitute for revenues, and solving for P , we obtain:
P = Mα′α
(σ − 1
σ
)ασ−α′α
(1− β
α
E
f
)α−α′α
(1c
)σ−1
∫
ϕ
∫
ξ(ϕ)
ϕ(σ−1)α
α′ ξα−α′
α′ h (ϕ, ξ) dξdϕ
α′α
= Mα′α
(σ − 1
σ
)ασ−α′α
(1− β
α
E
f
)α−α′α
(1c
)σ−1
(η̃)α′α (15)
10
where
η̃ ≡∫
ϕ
∫
ξ(ϕ)
η(ϕ, ξ)h (ϕ, ξ) dξdϕ (16)
is the (weighted) average combined productivity of surviving firms.
There is free entry into the market. Firms choose to pay the fixed cost fe and learn their
productivity and caliber only if the expected profits are greater or equal than the entry cost.
Expected profits Π are given by
Π =∫
ϕ
∫
ξ
Π(ϕ, ξ) v (ϕ, ξ) dξdϕ = Pin
∫
ϕ
∫
ξ(ϕ)
Π(ϕ, ξ) h (ϕ, ξ) dξdϕ (17)
where the last equality uses equation (14) and the fact that firms with ξ < ξ (ϕ) earn zero profits.
The free-entry condition imposes Π = fe. Using (10), (15), and (17), we obtain:
M =(
PinE
fe + PinF
)(α′
ασ
). (18)
The free-entry condition determines the mass of entrants as an increasing function of Pin. The
intuition is simple. Since expected profits have to equal the exogenously given fe, the easier it is to
survive (higher Pin) the larger will be the mass of entrants, so that the competition effect of entry
exactly cancels out with the higher survival rate.
Finally, we can substitute (18) into (15), divide both sides by P and elevate them to the power
of αα′ to obtain
1 =
(η̂/P
αα′
fe + PinF
)B
αα′ (19)
where B = Jα′α E
(1f
)α−α′α
(1c
)σ−1
and
η̂ ≡ Pinη̃ =∫
ϕ
∫
ξ(ϕ)
η(ϕ, ξ)v (ϕ, ξ) dξdϕ. (20)
The cut-off condition (12) and the free-entry condition (19) form a system in P and ξ (ϕ).
We cannot solve this system in closed form unless we assume a particular shape of the bivariate
distribution. However, we can prove that there exists a solution and that it is unique.
Proposition 1. In the closed economy, an equilibrium exists and it is unique.
Proof. See Appendix.
11
2.2 The open economy
This section analyzes the open-economy equilibrium. First, we follow standard firm-heterogeneity
models by assuming that exporting requires paying iceberg transport cost τ and fixed exporting
cost fx. Then, we introduce a minumum quality requirement for export and characterize the cross-
section of firms in equilibrium. We also consider how differences in the distributions of productivity
and caliber across countries affect the benefits of opening up an economy to trade.6
2.2.1 Unconstrained export quality
For a firm that exports, we define total demand Dw as the sum of domestic and foreign demand:7
Dw = D + D∗ = p−σλσ−1
[E
P+ τ−σ E∗
P ∗
].
The maximization problem of an exporting firm is analogous to the problem of a firm in the
closed economy except for the fact that EP + τ−σ E∗
P ∗ is their relevant market size. For a firm that
exports, optimal quality is then given by
λu(ϕ, ξ) =[1− β
α
ξ
f
(σ − 1
σ
)σ (ϕ
c
)σ−1(
E
P+ τ−σ E∗
P ∗
)] 1α′
(21)
while profits are given by
Πx.u(ϕ, ξ) = J(ϕ
c
)α(σ−1)
α′(
ξ
f
)α−α′α′
(E
P+ τ−σ E∗
P ∗
) αα′ − F − fx (22)
= η(ϕ, ξ)J(
1c
)α(σ−1)
α′(
1f
)α−α′α′
(E
P+ τ−σ E∗
P ∗
) αα′ − F − fx (23)
where the subindex “x.u” referes to an exporting firm in the unconstrained equilibrium. A firm
that chooses not to export but remains active in the domestic market has profits given by equation
(10). For expositional clarity, we now denote those profits by Πd. Finally, denote by ∆u(ϕ, ξ) the
difference in profits between exporting and not exporting. Using (11) and (22), we obtain
∆u(ϕ, ξ) = Πx.u(ϕ, ξ)−Πd(ϕ, ξ) (24)
= η(ϕ, ξ)JA
(1c
)α(σ−1)
α′(
1f
)α−α′α′ − fx
6We do not include this part in the paper because it is not completed yet.7Foreign variables are denoted with an asterisk.
12
where A =[(
EP + τ−σ E∗
P ∗) α
α′ − (EP
) αα′
]. Firms choose to export if Πx.u(ϕ, ξ) > Πd(ϕ, ξ), i.e. if
∆u(ϕ, ξ) ≥ 0 (indifferent firms are assumed to export). The function ∆u(ϕ, ξ) is increasing in η,
which in turn is increasing in both of its arguments, ϕ and ξ. This implies that a firm will export
if its caliber is above a threshold value
ξx.u
(ϕ) =fx
JA
α′α−α′
f(ϕ
c
)−α(σ−1)
α−α′, (25)
determined by setting ∆u(ϕ, ξ) = 0 and solving for ξ.
As in Melitz (2003), there are two possible scenarios. Either all surviving firms export or there
is a subset of them who do not enter the foreign market. Whether one or the other scenario prevails
can be determined by evaluating the profits that cut-off firms – those located on the curve ξ(ϕ)
– would make if they exported. If cut-off firms make negative profits by entering foreign markets
(Πx.u(ϕ, ξ(ϕ) < 0), they will prefer to be active only domestically. Therefore, a non-empty subset
of non-exporting survivors will exist iff
fx >FA
(EP
) αα′
, (26)
which we assume holds for the remainder of the paper. This assumption then implies that ξx.u
(ϕ) >
ξ(ϕ).
Combined productivity η is a summary statistic for profits and revenues, both for exporters and
for non-exporters. Thus, firms with the same η have identical profits and revenues regardless of their
particular realizations of ϕ and ξ. This property has several implications, which are shared by most
trade models of firm heterogeneity. First, firm size perfectly explains export status. In particular,
there is a cut-off size such that all firms larger (smaller) than the cut-off size are exporters (non-
exporters). Second, since η summarizes all relevant information determining relative size (revenue)
both before and after trade, the ranking of firms by size is invariant to trade regime. In addition,
the relative size of any two exporters or any two non-exporters is not affected by trade liberalization;
only the relative size between exporters and non-exporters changes.
Since both exporter and non-exporter profits depend only on η, the function ∆u(ϕ, ξ) also
depends on the value of this variable alone. This implies that the value of η that defines the function
ξx.u
(ϕ) is also an iso-profit curve.8 Even though exporters with the same combined productivity η
8It is easy to show that firms located on this curve earn strictly positive profits: substitute (25) into (22) to find
that Πx.u(ϕ, ξx.u
(ϕ)) > 0.
13
have equal size and profits, the quality that they produce is not identical. Equating Πx.u(ϕ, ξ) to any
constant, solving for ξc(ϕ) (superscript c denotes the level of the iso-profit curve) and substituting
into equation (21), we obtain:
λu(ϕ, ξc(ϕ)) = B(ϕ
c
)− σ−1α−α′ (27)
where B is a function of constant parameters. Equation (27 ) shows that, along any iso-profit curve
(including ξx.u
(ϕ)), quality is decreasing in ϕ.9
In the equilibrium with no export quality requirements, even though the particular draws of ϕ
and ξ – conditional on η – matter for prices and quality choice, only η matters as a determinant
of firm size, profits and export status; the two heterogeneity factors’ influence on those outcomes
occurs only through η. As a result, this benchmark model of the open economy can also be collapsed
into a model isomorphic to Melitz’. Figure 2 shows the equilibrium configuration of firms. Firms
with productivity values below the cut-off function ξ(ϕ) (those with η below the cut-off combined
productivity η) exit the market. Firms with productivity values between this function and the
export cut-off function ξx.u
(ϕ) (those with η between the two relevant cut-off values) sell in the
domestic market but do not export. Finally, firms with productivity draws above ξx.u
(ϕ) are active
in the export market.
2.2.2 Constrained export quality
Firm-level evidence in many developing countries suggests that exporters produce higher quality
products than non-exporters (e.g. Brooks 2006 and Verhoogen 2004). The standard explanation
for this regularity is that quality for export is higher because the largest importers have higher
income levels and demand higher quality. However, a number of other factors, in addition to
income differences between domestic and foreign markets, can explain the regularity. First, as
conjectured by Alchian and Allen (1964), transportation costs could be lower for higher quality
goods. Second, low quality could lead to haggling and negotiations over defective items, which
are costlier to undertake in international transactions. Third, some export markets might impose
minimum quality standards. Fourth, in the presence of incomplete contracts and given the costs
of international monitoring, high quality standards (e.g. ISO 9000) might be a way to prevent9The property that quality is decreasing with ϕ along an iso-profit curve is also true for non-exporters, both in
the open-economy and in the closed-economy case.
14
cheating on the contract. For some or all of these reasons, quality for export may be higher than
quality for the domestic markets. If these factors – instead of income differences – are important,
the finding that quality for export is higher than quality for the domestic market would be true not
only for developing countries but also for developed countries, even when they export to developing
countries.
In our model, we capture the idea that quality for export is higher than quality for the domestic
market in a direct and simple manner. We assume that a firm has to attain a minimum quality
λ to be able to export. Therefore, compared to the unconstrained export quality equilibrium, the
requirement of a minimum quality forces firms who would otherwise export with quality λu < λ to
choose between upgrading their quality (relative to their unconstrained choice) or not participating
in the export market. As a result, the unconstrained export cut-off function ξx.u
(ϕ) derived in the
previous section is no longer relevant.
We now characterize the export cut-off function in the constrained environment, denoting it by
ξx(ϕ). First, since unconstrained quality is decreasing in ϕ along ξ
x.u(ϕ), there exists a threshold
productivity value ϕ∗λ such that ∀ϕ ≤ ϕ∗λ, λu(ϕ, ξx.u
(ϕ)) ≥ λ. For those values of ϕ, firms that
would export in the unconstrained environment will also export in the constrained environment, as
their unconstrained optimal choice of quality is above the minimum requirement. Thus for ϕ ≤ ϕ∗λ,
ξx.u
(ϕ) = ξx(ϕ). Using (21) and (25), we can solve λu(ϕ, ξ
x.u(ϕ)) = λ for ϕ∗λ to find:
ϕ∗λ = cλ− α−α′
(σ−1)
[1− β
α
(E
P+ τ−σ E∗
P ∗
)] α−α′(σ−1)α′
(fx
AJ
) 1(σ−1)
(σ − 1
σ
)σ(α−α′)(σ−1)α′
(28)
Analogously, for firms with ϕ > ϕ∗λ, λu(ϕ, ξx.u
(ϕ)) < λ. In order to export, these firms would
need to choose a quality level above their unconstrained choice. Since they were indifferent be-
tween exporting and non-exporting in the unconstrained environment and now exporting requires
deviating from their optimal choice, these firms will strictly prefer not to export. Therefore, we
can establish that ∀ϕ > ϕ∗λ, ξx(ϕ) > ξ
x.u(ϕ).
Among exporters with ϕ > ϕ∗λ, not all are forced to upgrade their unconstrained choice of
quality. In fact, firms with sufficiently high ξ will spontaneously choose quality levels above λ.
Denote by ξxλ
(ϕ) the threshold function above which firms spontaneously choose quality above
λ. This function is an “unconstrained” iso-quality curve; i.e. it represents the locus of points on
the (ϕ, ξ) plane on which firms choose the same quality level (λ). We can obtain this function by
15
equating (21) to λ and solving for ξ, which results in
ξxλ
(ϕ) = λα′fα
1− β
(σ − 1
σ
)−σ (ϕ
c
)−(σ−1)(
E
P+ τ−σ E∗
P ∗
)−1
. (29)
Using the fact that λ > λ(ϕ, ξ
x.u(ϕ)
)when ϕ > ϕ∗λ in equation (29), it is easy to show that
ξxλ
(ϕ) > ξx.u
(ϕ) when ϕ > ϕ∗λ. Also, it is easy to show that ξx.u
(ϕ∗λ) = ξxλ
(ϕ∗λ) when ϕ > ϕ∗λ.
Define the difference in profits between exporting and not exporting in the constrained envi-
ronment as ∆(ϕ, ξ) = Πx(ϕ, ξ) − Πd(ϕ, ξ). Then the locus of firms that are indifferent between
exporting and not exporting, ξx(ϕ), is defined implicitly by the equation ∆(φ, ξ
x(ϕ)) = 0. For
firms with ϕ > ϕ∗λ and ξ (ϕ) > ξxλ
(ϕ∗λ), ∆u = ∆, as Πx.u(ϕ, ξ) = Πx(ϕ, ξ). We found in equation
(24) that ∆u is an increasing function of both ϕ and ξ. Therefore ∆(ϕ, ξxλ
(ϕ)) = ∆u(ϕ, ξxλ
(ϕ)) >
∆u(ϕ, ξx.u
(ϕ)) = 0. This implies that firms along ξxλ
(ϕ) strictly prefer to be exporters. Hence,
firms that are indifferent between exporting and not exporting have ξx(ϕ) < ξxλ
(ϕ). Combining
this result with the result of the previous paragraph, we have that ξx.u
(ϕ) < ξx(ϕ) < ξ
xλ(ϕ).
The export cut-off function ξx(ϕ) which is implicitly defined by ∆(ϕ, ξ
x(ϕ)) = 0 cannot be
solved in closed form. To characterize – in implicit form – this function we need to derive the profit
function for firms that upgrade their quality to export (Πxλ). These firms are forced to match the
minimum quality, and therefore their price, demand, revenue and profits are as summarized below:
p =σ
σ − 1c
ϕλβ
D = p−σλσ−1
(E
P+ τ−σ E∗
P ∗
)
r =(
σ
σ − 1
)1−σ (c
ϕ
)1−σ
λα−α′(
E
P+ τ−σ E∗
P ∗
)
Πxλ =1σ
(σ
σ − 1
)1−σ (c
ϕ
)1−σ
λα−α′(
E
P+ τ−σ E∗
P ∗
)− f
ξλα − fx − F
Accordingly the difference in profits between exporting and not exporting in the constrained envi-
ronment for the firms that choose to upgrade their quality level is given by:
∆ = Πxλ −Πd =1σ
(σ
σ − 1
)1−σ (c
ϕ
)1−σ
λα−α′(
E
P+ τ−σ E∗
P ∗
)− f
ξλα − fx
−J(ϕ
c
)α(σ−1)
α′(
ξ
f
)α−α′α′
(E
P
) αα′
It can be shown that ∂∆∂ϕ > 0 and that ∂∆
∂ξ > 0. Therefore, applying the implicit function theorem
to d(ϕ, ξ) = 0, we find thatdξ
x(ϕ)
dϕ < 0.
16
The characterization of the equilibrium can be seen in Figure 3. The export cutoff function in
the constrained regime ξx
coincides with the export cutoff function for the unconstrained regime
ξx.u
(ϕ) for all values of ϕ ≤ ϕ∗λ. For ϕ > ϕ∗λ, ξx.u
(ϕ) plays a role only as a reference locus. Another
reference locus (represented by a dotted line) is the threshold (λ) iso-quality curve ξxλ
(ϕ). As shown
previously, the relevant export cut-off function ξx(ϕ) (represented by a solid line) lies between the
export cutoff function in the unconstrained regime (ξx.u
(ϕ)) and the threshold iso-quality curve
(ξxλ
(ϕ)). Firms in region I do not survive. Firms in region II only sell in the domestic market.
Firms in region III also sell only in the domestic market but they would have been exporters were
it not for the minimum quality requirement. Firms in region IV export by producing the minimum
export quality, which is above their optimal choice in the absence of the constraint. Firms in regions
V export without being forced to change their optimal quality choice.
The requirement of a minimum export quality breaks the sufficiency of η to explain size rankings
and export status, both of which now depend on the particular combinations of ϕ and ξ. This
implies that the monotonic relationship between size and export status is also broken. As opposed
to the predictions of models with only one source of heterogeneity, the export status of firms with
the same size might be different. In particular, we now prove that:
Proposition 2. Conditional on size (total revenue), expected quality is higher for exporters.
Proof. Proof: We provide a graphical proof of this proposition using Figure 4. Figure 4 reproduces
Figure 3, but adds two iso-revenue curves. The iso-revenue curve R1 represents a set of curves that
lie fully in Region II. Since there are no exporters on those curves, the proposition holds trivially
for the revenue levels that they represent. A second set of iso-revenue curves is characterized by
R2. R2 crosses Region V (point A is representative of that segment) until it intersects with region
IV at point B. At that point, it goes straight down across region IV until reaches point C.10 To
understand why firms with different ξ have identical revenue along the last segment, we can note
10Those two segments are characterized, respectively, by the equations ξ(R2, ϕ) =
0@R2
α′α−α′
ϕα(σ−1)(α−α′)
1A 1
H
and ϕ(R2) =
R2
1σ−1
λα−α′σ−1
!1
H ′ , where H and H ′ are constant functions of the parameters, H =
�σ − 1
σ
�ασ−α′α−α′
�E′
P ′
� αα−α′
�1
c
� α1−β
�1− β
αf
�, H ′ =
�σ − 1
σ
��1
c
��E′
P ′
� 1σ−1
, andE′
P ′=
E
P+ τ−σ E∗
P ∗. Note
that the second equation does not depend on ξ.
17
that, since those firms all produce the minimum export quality (λ) and have the same productivity
(ϕ), there marginal cost is identical and so they charge the same price. Therefore, with equal price
and quality, their revenue is also identical (despite the fact that they pay different fixed costs to
produce λ due to their different caliber ξ).
At point C, there is a discontinuity in the revenue curve. This point marks the threshold
between firms who choose to upgrade their quality to meet the export requirement and those who
decide to focus in the domestic market. Formally, the existence of a discontinuity can be shown by
considering a point (ϕIVC , ξIV
C ) arbitrarily close to point C belonging to Region IV, and a second
point (ϕIIIC , ξIII
C ) arbitrarily close to point C belonging to Region III. It is easy to find that:
rIIIC = (λIII
C )α−α′(
σ
σ − 1
)1−σ (c
ϕIIIC
)1−σ (E
P
)
< rIVC = (λ)α−α′
(σ
σ − 1
)1−σ (c
ϕIVC
)1−σ (E′
P ′
)
since λIIIC < λ and
E
P<
E′
P ′ =E
P+ τ−σ E∗
P ∗ . Finally, iso-revenue curve R2 continues in region III
starting in point D and continuing through point E indefinitely to the right.
Let D(ϕ, ξ) be an indicator function for export status, so that D(ϕ, ξ) = 1 when a firm exports
and D(ϕ, ξ) = 0 otherwise. From our discussion of iso-revenue curve R2, we can establish that
∀(ϕ, ξ) on R2 such that D = 1, the minimum quality level is λ. In contrast, ∀(ϕ, ξ) on R2 such
that D = 0, the maximum quality level is (infinitesimally) less than λ. Generalizing this result to
the set of iso-revenue curves represented by R2, we obtain that:
E[λ|D = 1, r = R̄] ≥ λ > E[λ|D = 0, r = R̄]
Thus, size is not the sole predictor of export status. Conditional on size, expected quality is higher
for exporters.
The data does not include measures of quality perfectly consistent with λ in our model. However, in
the model, quality affects marginal costs and fixed costs of operations. If these quality related costs
could be identified in the data, these costs could form reasonable proxies for quality, allowing us to
test proposition 1 indirectly. In the Indian dataset (described in Section 3), there is information
that allows us to form such proxies for quality. One available variable is a dummy that tracks
whether the firm has ISO 9000 certification. Since ISO 9000 is a quality certification, this dummy
18
variable would be a good proxy for quality (λ). In addition there are 5 dummy variables that track
whether the establishment uses computer and robotics technologies. Since it is reasonable that
at least part of the investments in computers and robotics may be related to helping the firm to
improve the quality of the product, these dummy variables could also serve as proxies for λ.11
A corollary of Proposition 2 establishes a similar prediction in terms of a directly observable
variable – the output price.
Corollary 1. Conditional on size, mean price levels are higher for exporting firms.
Proof. Using Proposition 2 and equations 6, 7 and 8, we derive the price charged by a non-exporting
firm:
p =(
E
P
) 1σ−1
(λ
r1
σ−1
).
Therefore,
E[p|D = 0, r = R̄] =(
E
P
) 1σ−1
(1
R̄1
σ−1
)E[λ|D = 0, r = R̄].
Similarly, for an exporter we have:
E[p|D = 1, r = R̄] =(
E′
P ′
) 1σ−1
(1
R̄1
σ−1
)E[λ|D = 1, r = R̄].
Then, Proposition 1 and the fact thatE
P<
E′
P ′ =E
P+ τ−σ E∗
P ∗ imply that:
E[p|D = 1, r = R̄] > E[p|D = 0, r = R̄].
The intuition for this result is very simple. Compare the segments to the left and to the right of
point C along R2. Firms to the left of point C produce higher quality and have lower productivity,
both of which are factors that induce them to charge a higher price.
Thus, in a world where firms need to exceed a minimum quality threshold to become exporters,
firms that have the same size (gross revenues) may differ in their export status. Hence, size the
relationship between size and export status is not necessarily a step function, and is likely to be11 In the model in Bustos (2005), exporters are more likely to adopt high technology, a result driven by the higher
productivity (size) of exporters, and the fixed costs of investing in technology. In this model, technology adoption
is driven by the single productivity factor. Hence it does not predict a higher probability of technology adoption
conditional on size. Therefore our empirical findings are not driven by the considerations modelled in Bustos.
19
smoother (as in Figure 1). Conditional on size, some firms that have relatively higher caliber (ξ)
would choose higher levels of quality and become exporters, while firms that have relatively high
productivity (ϕ) stay as purely domestic firms.
For achieving the same revenue level, exporters produce lower quantities but charge higher
prices, whereas the purely domestic firms produce greater quantities which they sell at lower prices.
This distinction between the way quality and productivity affect prices is important. If export status
was indeed driven by a supply-side productivity advantage as modelled in Melitz (2003), then it
would be the case that exporters charge lower prices than non-exporters. Thus any evidence that
exporters charge a relatively higher price would be a direct refutation of the productivity advantage
hypothesis and provide support for a demand-side quality advantage for exporters.
In the empirical section below (Section 4), we test corollary 1 using plant-level data for the
Indian manufacturing sector (described in Section 3). This dataset has data on unit quantities
and value of sales, which allows us to measure the per unit price for all the products produced by
each plant. This price data can be used to directly test corollary 1, by forming sample analogs of
E[p|D, r] (see discussion in Section 4.1).
In the next section, we model the fixed and variable costs of producing quality in greater detail.
This allows us to relate the production of quality to a number of additional observable variables –
capital and skill intensity, as well as worker ability (proxied using wage rates).12
2.3 Capital intensity, skill intensity and worker ability
Since this is a partial equilibrium model, we did not need to fully specify the source of the higher
costs required for the production of quality. We do this here by following an approach similar to12 Our model also makes subtle predictions relating to reallocation of resources following trade liberalization.
Specifically, our model suggests the ranking of firm sizes would be reshuffled with trade liberalization. For example
consider an iso-quant that crosses regions V IV and III in a closed economy (autarky) world in Figure 3. Following
trade liberalization with constrained quality, firms on the part of the iso-quant that lies in Region V (relatively high
caliber firms) would choose to export and hence increase in size considerably. However, firms on the part of the
iso-quant that lies in region III would now choose to focus on the domestic market and hence will remain at the same
size as in the closed economy case. The firm on the part of the iso-quant that lies in region IV would upgrade their
quality and serve the export markets. Thus their size (total revenue) would go up due to both the upgradation in
their quality and due to their access to export In sum, trade liberalization would lead to the reallocation of resources
towards higher caliber firms.
20
Verhoogen (2004).
Define the average wage paid by the firm for skilled workers as ws and for unskilled workers as
wu:
ws =
∑j wj
sLjs∑
j Ljs
wu =
∑j wj
uLju∑
j Lju
where j denotes types of workers by their ability and wk the market wage for that ability. Similarly,
define the average rental rate paid by the firm as wk:
wk =
∑j wj
kKj
∑j Kj
where j here denotes different vintages of capital.
To produce quality λ, we assume that it is required that ws = wsλbs , wu = wuλbu and wk =
wkλbk . Here, wj , j ∈ (s, u, k) can be thought of as the price for the minimum quality input within
each category. Thus, we assume that producing higher quality requires hiring skilled and unskilled
workers of greater ability and the use of more recent (hence more productive) vintages of capital.
Next, we assume that bu > bs > bk. That is, for the same increase in quality level, the increase in
average input price required is highest for unskilled workers, and then for skilled workers and least
for capital. Implicitly, we assume that the difference in ability (in price terms) between unskilled
workers who are required produce higher quality goods and those used to provide lower quality
goods is much more than the ability difference between skilled workers in high vs low quality good
production, which in turn is higher than the difference in vintage between capital used in low versus
high quality production.
We assume that output is produced using a three input constant returns to scale Cobb-Douglas
production function:
Y = ϕLαss Lαu
u Kαk
where αs + αu + αk = 1. This yields a constant output cost function of the form, conditional on
quality level λ as:
C(Y,ws, wu, wk, λ, ϕ) =A
ϕ· Y · wαs
s wαuu wαk
k =c
ϕ· Y · λβ
21
where A =1
ααss ααu
u ααkk
and c = A · wsαs · wu
αu · wkαk and β = αsbs + αubu + αkbk. Thus we get
marginal cost as defined:
c(λ, ϕ) =c
ϕλβ
We model the fixed cost part of quality production in a very similar fashion. Specifically, we
assume that quality production requires use of capital, skilled labor and unskilled labor which are
combined in a Cobb-Douglas CRS production function to produce quality (with same input shares
as in the output production function):
λ1κ = ξL′s
αsL′uαuK ′αk
Note that caliber ξ plays a role in quality production that is similar to the role played by pro-
ductivity ϕ in physical output production. These quality-related fixed costs can be thought of as
expenses related to systems for maintaining records, quality control equipment or worker training.
In addition, the firm incurs other fixed costs F (such as annual maintenance expenses or head-
quarters expenses, which we assume is not reflected in any of the input costs) unrelated to quality.
Accordingly, just as in the case of the output production function, we get a cost function conditional
on quality λ given by:
F (λ,ws, wu, wk, ξ) =A
ξ· λκ · wαs
s wαuu wαk
k + F =f
ξ· λα + F
where A =1
αsαsαu
αuαkαk
and f = A·wsαs ·wu
αu ·wkαk and α = κ+αsbs+αubu+αkbk. Given these
assumptions regarding the production of output and quality, we can show the following additional
corollaries to Proposition 2. 13
Corollary 2. Conditional on size (i.e. revenue), mean capital intensity (measured as the capital
to labor ratio) is higher for exporting firms.
Proof. See Appendix 3.
Corollary 3. Conditional on size (i.e. revenue), mean skill intensity (measured as the skilled
workers’ share of employment) is higher for exporting firms.13From expression for β above, we get α = κ+β. In Section 2.1 (see discussion below equation 6), we had imposed
the restriction α′ = α− (1− β)(σ− 1) > 0. This condition then translates to κ− σ + 1 + βσ > 0 or κ > σ(1− β)− 1.
22
Proof. From the assumption that bu > bs, the proof that skill intensity is increasing in quality in the
variable cost component is similar to that of Corollary 3 (see Appendix 3). The same independence
condition as in equation 47 is sufficient here too.
Corollary 4. Conditional on size (i.e. revenue), mean worker ability for both skilled and unskilled
workers (measured as the average wage rate) is higher for exporting firms.
Proof. This follows directly from our assumptions and Proposition 2 above. The observed average
wage for skilled workers ws = wsλγ . Therefore following from Proposition 2:
E[ws|D = 1, r = R̄] = wsE[λγ |D = 1, r = R̄]
> wsE[λγ |D = 0, r = R̄] = E[ws|D = 0, r = R̄]
The same reasoning applies for average wage of unskilled workers.
3 Data
3.1 Data sources
For our empirical analysis, we use data from manufacturing sector surveys for India, Chile and
Colombia.
For India, we use unit-level data from the Annual Survey of Industries for the year 1997-98. The
Annual Survey of Industries (ASI) is a survey undertaken by the Central Statistical Organization
(CSO), a department in the Ministry of Statistics and Programme Implementation. The ASI covers
all industrial units (called “Factories”) registered under the Factories Act employing more than 20
persons.14 The ASI frame is classified into two sectors: the “census sector” and the “sample sector”.
Factories employing more than 100 workers constitute the census sector. Factories in the sample
sector are stratified and randomly sampled, while all those in the census sector are surveyed. The
sample weights (multipliers) for all the establishments are provided in the dataset. In our analysis,14This limit is lower (10 employees) for plants that use electric power for production. Even though plants with
employment below this lower limit are not required to register, a significant number of plants report less than 10
employees. This happens apparently because some plants below the mandated limit voluntarily choose to register or
because some plants that initially registered when they had more than 10 employees continue to stay registered even
after employment levels fall below the cutoff.
23
we use these multipliers to adjust for the sampling weights. We focus on 1997-98 because the ASI
data for this year includes information on exports and on output quantities and values. The Indian
data is classified according to the National Industrial classification 1987 revision (NIC-87). Each
establishment is provided a 4 digit NIC code, which is a finer classification than the ISIC 4 digit
code. In fact, the NIC 3 digit code is based on the ISIC 4 digit classification. There are about 450
distinct 4 digit NIC industries in the data.
For Chile, the data we use is drawn from the annual Chilean Manufacturing Census (Encuesta
Nacional Industrial Anual) conducted by the Chilean government statistical office (Instituto Na-
cional de Estadistica). The survey covers all manufacturing plants in Chile with more than 10
employees and has been conducted annually since 1979. We use data for the years 1991-96, for
which data on exporting is available. Chilean industries are classified using the ISIC 4 digit classi-
fication. Data for Colombia comes from the Colombian manufacturing census for the years 1981 to
1991. The census covers all plants with 10 or more employees and includes data on exports for each
establishment. Like Chile, Colombia’s industries are classified using the ISIC 4 digit classification.
3.2 Definition of variables and summary statistics
The key predictions of our theory relate to output price, capital intensity, skill intensity and worker
ability. Data on capital intensity (capital to labor ratio), skill intensity (two variables – skilled
share of employment and skilled share of wages15), and worker ability (overall average wage rate,16
wage rate of skilled workers and wage rate of unskilled workers) are available in the datasets for all
three countries, so corollaries 2, 3 and 4 are testable using all three datasets.
For India, there is more information collected from the establishments, which allows us to form
additional proxies for quality. Data on quantity and sales value allow us to estimate the per-unit
prices for each product line for each firm. Hence we are able to test corollary 1 using Indian15Corollary 3 relates to skill intensity defined as the share of skilled workers in total employment. Since the ratio
of skilled wage bill to unskilled wage bill is given bywsL
totals
wuLtotalu
=αs
αu, under the following additional assumptions: (i)
αj = α0jλ
aj , j ∈ (s, u, k) and (ii) ak > as > au, it is easy to show that the mean skilled share of wage bill (as well
as capital share of value added) would also be higher (conditional on size) for exporters. Assumptions (i) and (ii)
imply that firms choice of technology depends on the quality level, and the technology parameter corresponding to
the output elasticity of skilled labor increases faster with quality than the output elasticity of unskilled labor.16Following from corollaries 3 and 4, a sufficient condition for the mean overall wage rate (conditional on size) to
be higher for exporting firms is that ws > wu for all feasible λ values.
24
data. The price variable is defined at the product category level, so that there are multiple price
observations for each firm. Since all other data, especially data on exports (value of exports), is
available only at the establishment level, in our analysis we check robustness to different assumptions
about which product lines may be exported (see discussion in Section 4.2.1 below).
As discussed in Section 2.2.2, the Indian dataset also includes information on whether the
establishment has obtained ISO 9000 certification. We use an indicator variable for ISO 9000 as
a proxy for quality to test Proposition 2 in Section 2.2.2. Also, the dataset includes five dummy
variables that reflect adoption of computer usage. One dummy indicates whether computers are
used in the establishment. A second dummy is an indicator for the use of robotics in the plant.
A third dummy variable indicates the presence of a computer network. A fourth dummy variable
indicates whether there is access to internet at the plant. Finally, a fifth dummy variable indicates
whether the establishment provided responses to the survey on a floppy – this dummy could indicate
a more active use of computers among plants that do have computers on the premises. These
variables are relevant to our model in two ways. One, investment in computers could reflect the
higher fixed costs incurred in raising quality. Two, as discussed in section 2.3, we expect higher
quality firms to invest in more sophisticated equipment (such as computers). The mean values for
these variables indicate that only 3% of the establishments has obtained ISO 9000 certification.
While 32% of the establishments used computers in some form, usage of robotics and presence of
a computer network are really low (2%). Also, only 4% of the establishments has access to the
internet and 6% responded to the survey on a floppy.
All the variables that are common across the three datasets are defined in a consistent manner.
The export dummy variable is defined to equal one for all establishments reporting positive value of
total exports. The capital variable for Chile is constructed using the perpetual inventory method.
For India and Colombia, capital is taken as the reported total fixed assets.17 For all three datasets,
the labor variable used in constructing the log capital to labor ratio is total employment.
For India and Colombia, each establishment reports its initial year of operation. Hence age is
defined as the difference between the current year and the reported initial year of operation. For17In all our analysis, we use industry-year fixed effects. Coupled with the fact that the capital variable is logged in
all the regressions, we expect our results to be invariant to using an industry level deflator to normalize the capital
variable.
25
Chile, the startup year is not reported. So we define age as the difference between current year and
the first time the firm appears in the dataset. To define the first year of entry, we use data from
1979 to 1996 (though lack of data on exports restricts us to use only data from 1991 to 1996 in the
rest of the analysis).
The average wage rate is obtained by dividing total wages by total employment for each es-
tablishment. The skilled share of the wage bill is the ratio of total wages to non-production
workers divided by the total wage bill of the establishment (which is the sum of total wages to
non-production and production workers). The skilled wage rate is obtained by dividing the total
wages to non-production workers by the number of non-production workers. Similarly, the unskilled
wage rate is defined as the total wages to production workers divided by the number of production
workers. Total sales is the total revenues for each establishment.
Panel 1, Panel 2 and Panel 3 of Table 1 present summary statistics for India, Chile and Colombia
respectively. All statistics are adjusted to account for sampling weights. The adjustment for
sampling weights is important only for India, as the datasets for Chile and Colombia are census
datasets. That is, for Chile and Colombia, each establishment is sampled with certainty so that
multipliers (inverse of the sampling weights) are equal to one for all observations.
The mean value for the export dummy shows that exporting is more common in Chile, where
about 20% of establishments export. It is less common in India (13%) and in Colombia (12%).
Since the currency units vary across countries, some of the variables are not directly comparable
across the three countries. The skilled share of the wage bill is remarkably similar across the
three countries – 0.35 for India and Chile and 0.34 for Colombia. The mean employment level is
higher in India (4.88 log points) and similar across Chile and Colombia (3.72 and 3.45 log points
respectively).
4 Empirical analysis
We wish to show that quality is an important independent factor, in addition to productivity, in
determining export status. In particular, our theory predicts that the expected value of quality
is higher for exporters, holding constant the size (i.e revenue) of the exporter. This prediction
contrasts with the results from the standard models in the literature (e.g. Melitz 2003) where
26
productivity and hence size is the sole determinant of export status. In the traditional models, all
firms above a cutoff productivity (and hence revenue) export, and conditional on revenue there are
no interesting differences between exporters and non-exporters.
Since quality cannot be directly measured in the data, we look for reasonable proxies for quality.
We show that under certain simple assumptions output price is indeed correlated with quality, and
Corollary 1 demonstrates that the mean output price will be higher for exporters relative to non-
exporters. Similarly, we show that under certain assumptions, capital intensity, skill intensity and
mean worker ability are correlated with quality, and hence the conditional mean of these variables
would be greater for exporters relative to non-exporters (corollaries 2, 3 and 4).
In the next section, we discuss the methodology and underlying identifying assumptions used
to test the predictions of our model.
4.1 Methodology
For all of our dependent variables, we assume that the conditional expectation takes a linear form
as below:
E[Y |D, r, Z] = δ ·D + f(r) + Dj + g(Z) + u (30)
where D is an indicator dummy for export status (D = 1 for exporters and D = 0 for non exporters),
f(r) is a control function for the size (revenue), Dj denotes 4 digit industry year dummy variables,
Z is a vector of other variables that could affect the dependent variable of interest, and u is classical
error term uncorrelated with other variables. For example, our theory suggests that in the case of
capital intensity, Z includes technology parameters as well as relative prices of capital and labor (see
Section 2.3). Our key identifying assumption (as discussed in footnote 32) is that conditional on
other control variables, the omitted variables Z are not correlated with the export dummy variable.
That is:
E[g(z)|D = 1, f(r), Dj ] = E[g(z)|D = 0, f(r), Dj ] (31)
In the example of capital intensity, this would mean that for the firms in the same industry and
same size, variables such as rental price of capital and technology parameters do not vary between
exporters and non-exporters.18 This may not be unreasonable as firms of similar size within the18Another omitted variable is quality, but note that this does not really induce a bias as the exporter dummy is
intended to capture quality as a dummy variable.
27
same industry could be expected to have access to capital and other inputs at similar prices.
Our model predicts that δ > 0 for all the dependent variables (price, capital intensity, skill
intensity and proxies for worker ability). To test this prediction, we adopt two broad approaches –
a linear regression framework and a non-parametric regression approach.
In the linear regression approach, we use the following baseline regression specification:
logYit = δDit + αlogXit + Djit + εit (32)
where i indexes firms, t indexes years and j indexes industries. Dit is a dummy that equals 1 if firm
i is an exporter in year t and Xit is firm size, and Djit denote industry-year dummy variables.
In our model, productivity does not directly impact the dependent variables, but our predictions
are conditional on size (revenue). Hence we take particular care to check robustness to alternative
ways of conditioning on size. To conform with the theoretical model, in our baseline specifications
we use revenues as the proxy for size. In alternative specifications, we check the robustness of our
results to the following different ways of conditioning on size: (i) linear and quadratic revenue terms;
(ii) total employment (rather than revenue) as the size proxy; (iii) linear and quadratic employment;
(iv) industry-year-size fixed effects, where size is as the decile of the size distribution within each
industry.19 This provides a much more flexible control for size, as it allows the relationship between
the dependent variable and size to vary across each of the ten size cells within each industry, rather
than constraining it to have a specific (linear or quadratic) functional relationship. Finally, we
condition on age by estimating a specification with (v)industry-year-size-age fixed effects. This
would control for differences between exporters and non-exporters that arise due to differences in
firm age. For example, input prices especially for capital may be correlated with age, and hence
affect many of our dependent variables. To keep the total number of fixed effects reasonable, we
look at age quintiles (within the industry-year-size decile cells).
Note that if the variables (Y, D, r, Z) were distributed joint normally, then the conditional
expectation of Y could be written as the linear sum of the other variables and an orthogonal error
term u, as in equation (30). Since these variables may not generally be jointly normally distributed,
the linear functional form restriction may be too narrow. To address this concern, we allow for a
very flexible relationship between dependent variables of interest and export status, conditional on19We chose ten size cells to ensure that there are sufficient exporter as well as non-exporter firms within each cell.
A finer cut on size would lead to loss of data from cells that have either the control group or exporters missing.
28
size, using a non-parametric approach. Here we condition on size using a locally weighted smoothed
regression (lowess). The lowess graph provides a non-parametric picture of E[Y |D, r], allowing for a
much richer dependence between export status (D) and the variable of interest (Y) across different
size percentiles. We use the default bandwidth of 0.8 in our analysis (which implies that 80% of the
data are used in smoothing each point).20 Also, to ease computational burden, we use bootstrapped
samples of 30,000 to form the lowess estimates.
Under the identifying condition (31), the estimated coefficient on the exporter dummy would
indeed capture the difference between the mean of the dependent variable for exporters and non-
exporters, not reflecting differences in omitted variables, except quality differences. It must be
stressed, however, that export status is indeed driven by quality differences, which in turn are
affected by the exogenous draws of caliber (and productivity).
Thus the key exogenous variables that influence individual firm choices idiosyncratically (whether
to export, as well as choices on price, capital intensity, skill intensity and worker ability) here are
caliber and productivity. Assuming that caliber and productivity are independently distributed, the
estimated coefficient on the exporter dummy can be considered to be the mean change in dependent
variable if the average caliber of non-exporters was raised to the average caliber of exporters.
Note that, as in Melitz (2003), in our model exporters self-select on previous endowments of
caliber and productivity. Since caliber and productivity are not endogenous in our model, we do not
expect a change in status of the firm from non-exporter to exporter to affect any of the dependent
variables (as the change in status does not affect the underlying productivity or caliber). Hence
we expect the differences in quality (and hence of price, capital intensity, skill intensity and worker
ability) between exporters and non-exporters to be largely explained by cross-firm differences in20Technically, the lowess smoother forms a locally weighted smoothed prediction for the dependent variable in the
following manner. Let yi and xi be two variables, and order the data so that xi ≤ xi+1 for i =1,..., N-1. For each
yi, a smoothed value ysi is calculated. The subset used in the calculation of ys
i is indexed by i− = max(1, i − k)
through i+ = min(i + k, N), where k= Int[(N.bandwidth - 0.5)/2]. The weights for each of the observations between
j = i,..., i+ are the tricube (default):
wj =
(1−
� |xj − xi|∆
�3)
where ∆ = 1.0001max(xi+ − xi, xi − xi−). The smoothed value ysi is the (weighted) regression prediction at xi.
29
initial endowments.21,22 In Appendix 4, using the panel data from Chile and Colombia, we analyze
changes in variables within firms who switch from non-exporter to exporter status and find that
the observed differences between exporters and non-exporters seem to largely reflect self-selection.
That is, we find little change in the variables of interest within firms following a change in their
export status.
4.2 Empirical results
4.2.1 Output price
First, we use data from the Indian manufacturing sector for 1997-98 to examine the relation between
per unit price and export status. As predicted by Corollary 1 in section 2.2.2, per-unit prices should
be higher for exporters relative to non-exporters within an industry, reflecting the higher quality
of exporters. As noted earlier, since productivity advantages in the traditional models (e.g. Melitz
2003) translate into lower prices for more productive firms, evidence of higher prices for exporters
would be strong evidence for a quality-based difference between exporters and non-exporters, as a
quality advantage translates to a higher price in our model.
As discussed in section 3, the price per unit is obtained by dividing value of production by
quantities per product line. There are multiple product lines for most establishments. Also, units
used to measure quantity differs from product to product. For example some products are measured
in kilograms, others in liters – altogether there are 25 different units used. Data on exports, however,
are not disaggregated by product line. For our baseline analysis, we assume that an establishment
exports all of its product lines. Since it is meaningless to compare the prices of products that are
measured in different units, all regressions include unit code fixed effects. To enhance comparability
we go further and look at products which share the same unit code within the same industry
code definition. For checking robustness, we look at two different levels of industry-item code
aggregation. The results are presented in Table 2.
In the first row of Table 2 we report the coefficient on the exporter dummy controlling for 4-digit21In our model, some firms would choose to upgrade quality following their switch to exporting, and hence some
increase in the dependent variables within firms would not be inconsistent with our theory.22In this context, an interesting question could be why firms switch from non-exporter to exporter status or vice
versa. While we do (and most others in the literature do not explicitly model this), firm decisions to move in or out
of export markets can be considered as driven by idiosyncratic shocks to their transport or fixed exporting costs.
30
industry-unit code fixed effects. That is, in these regression, the coefficient on the exporter dummy
reflects the mean difference between exporters and non-exporters in the prices for products within
the same 4 digit industry that share the same unit code. In the second row, we report results
using a more aggregated fixed effect (3-digit industry-unit code). Column one is the specification
in equation (32) run without any controls for size. Column 2 includes log revenues, column 3
includes log revenue and the square of log revenue. Column 4 includes log total employment as
the size proxy, and column 5 is the same as column 4 except with the square of log employment
as an additional control. In all specifications, we find that price per unit is significantly higher
for exporting firms. In column 2, controlling for log revenue, prices are on average 20% higher
for exporting firms. The effects are stronger when we control for size using employment (columns
4 and 5) than when we use revenue (columns 2 and 3). In column 6, we control for size using
size decile dummies within the particular industry classification – the 4 digit industry-unit code
level (in row 1) and in the 3 digit industry-unit code level (in row 3). Here we find that prices are
higher by 19.3 per cent and 15.6 per cent when we control for 4 digit-unit code-size fixed effects and
3 digit-unit code-size fixed effects respectively. When we additionally condition on age-quintiles
within industry-year-size deciles (column 7), the price premium for exporters is 17.7 percent (for
both 4 and 3 digit industry definitions).
We undertake two robustness checks of our baseline results. First, we redo our analysis retaining
only the largest product line for each establishment, thus checking for robustness to the possibility
that establishments are more likely to export only their main line of products. Two, we exclude
observations where the units are undefined. There are about 10,000 observations where the name
of the unit used is unspecified; hence the measured prices on these products could be noisy.23 We
find that our qualitative conclusions are quite robust to these checks.24
Figure 6 illustrates a locally weighted smoothed (lowess) regression of per unit price on size,
separately for domestic and export firms. This provides a more detailed view of the relationship23In the baseline specification, we assume that all product lines with unspecified units within a 4 digit industry
have the same units, so that their prices are comparable.24We also checked for differences in the premium for small (below median) size firms. Consistent with the non-
parametric analysis below, we found that the premium is larger for smaller firms. We also checked and found no
significant differences between industries that are predominantly non-differentiated and predominantly differentiated
industries (using the Rauch 1999 classification).
31
between price and size, as we allow the relationship to vary across size percentiles. Overall, prices
are higher for larger firms. But conditional on size, we find a large difference between prices
for domestic firms and exporters. Across all percentiles of the size distribution, exporters have
a higher price per unit, consistent with the predictions of our model (and the results in Table
2). Interestingly, small firms that export have a higher price premium compared to similar sized
non-exporters. The premium is high also at the top of the size distribution, and is smallest in
the middle of the distribution. The high price premium for small exporters may reflect the larger
caliber differences required to overcome the relative productivity disadvantage for small firms.
4.2.2 Capital intensity
Next, we look at differences in capital intensity (measured as the log of the capital to labor ratio)
between exporters and non-exporters within industries, conditioning on size. All regressions include
industry-year fixed effects, and have errors clustered at the plant level. The industries are defined
at 4 digit NIC level for India and at 4 digit ISIC level for Chile and Colombia. The former is a
more disaggregated than the latter and hence there are greater number of industry fixed effects in
the regressions that use Indian data. The results are presented in Table 3.
We find that capital intensity is significantly higher for exporters, conditioning on both revenue
and employment proxies for size. In column 2 (conditioning on log revenue), capital to labor
ratio is about 43% higher for exporters in India, while it is 28.3% and 13.6% higher in Chile and
Colombia respectively. As in the case for prices, effects are stronger when we control for size using
employment (In columns 4 and 5). In column 6, we control for within industry-year size decile
effects. Here we find a capital intensity premium for exporters of 34.3 percent (India), 31.1 percent
(Chile) and 17.7 percent (Colombia). In column 7, we condition on industry-size-year-age effects,
and the magnitudes of the effects are close to those in Column 6.25
These results are confirmed by the non-parametric lowess regressions illustrated in Figure 7.
Interestingly, as in the case of price, larger firms appear to have greater capital intensity, suggesting25We looked separately at small versus large firms and differentiated versus homogenous industries (as per the
Rauch 1999 classification). We found some evidence for a larger capital intensity premium for small (less than
median sized) exporters in all three countries, consistent with Figure 7. For Chile and India, the capital intensity
premium was larger in differentiated goods industries, suggesting that quality differences may be more important for
these sectors. (There was no significant difference between the two sectors in Colombia).
32
that the production function may be non-homothetic. Again, these figures show that across all size
percentiles, exporters have a higher capital to labor ratio compared to non-exporters, consistent
with the predictions of the model. For India (and for Chile), the capital intensity premium appears
to be larger for small firms, consistent with the results for price in Figure 6.
We conclude that there is strong evidence for higher capital intensity among exporters, condi-
tional on size.
4.2.3 Skill intensity
In Table 4, we examine different measures of the skill intensity of exporters relative to non-exporters
within industries (defined the same way as for Table 3).
The first measure we look at is the skilled share of employment. This measure is significantly
greater in specifications controlling for revenue (columns 2 and 3) and employment (columns 5
and 6), but is insignificant in the unconditional regression (column 1) and when we condition on
industry-size-year-age effects. For Chile, the differences are only significant unconditionally (column
1) and conditional on employment (columns 4 and 5). In the case of Colombia, the mean skilled
share of employment is higher for exporters across all seven specifications.
An alternative measure of skill intensity is the skilled share of wage bill. We find that this
measure of skill intensity is significantly higher for exporters, across all three countries and across
all specifications. Conditioning on log of total revenues (column 2), the skilled share of the wage
bill is 2.6% higher for exporters in India relative to non-exporters. This premium is 1.9% in Chile
and 2.3% in Colombia. The premium conditional on industry-size-year-age effects is 1.3% in India
and Chile, and 5% in Colombia.
As was the case with price and capital intensity, results are stronger for all the skill intensity
variables, when we control for size using employment rather than revenues. Since our theory
suggests the use of revenue as a control, we take the results that condition on revenue more seriously.
Figures 8a and 8b provides a more detailed picture of the differences in skill intensity between ex-
porters and non-exporters, conditional on size. The graphs for skilled share of employment (Figure
8a) suggests that, in India and Chile, the skilled share of employment is higher for large exporters,
while it appears to be smaller for small exporters (relative to non-exporters). In Colombia, there
appears to be a positive premium for exporters across all percentiles of the size distribution.
33
In contrast, in Figure 8b we find that across all size percentiles, the skilled share of the wage
bill is higher for exporters, in all three countries. Also, as in the case of capital intensity, there is
evidence of non-homotheticity, with this skill intensity measure be larger for bigger firms.
We conclude that there is strong evidence for higher skill intensity (especially when measured
as skilled share of the wage bill) among exporters, conditional on size.
4.2.4 Worker ability
We measure worker ability using the wage rate as a proxy, under the reasonable assumption that
wage rates are positively correlated with ability. Since we have data separately on non-production
(skilled or white collar) and production (unskilled or blue collar) employment and wages, we can
form three separate wage measures – overall average wage rate, wage rate for skilled workers and
wage rate for unskilled workers. Results are presented in Table 5.
The first measure we look at it the overall average wage rate. As predicted by our model (see
corollary 3 in section 2.3), we find that the average wage rate is higher for exporters compared
to non-exporters even conditioning on size – alternatively revenues or employment – in all three
countries (the results are only marginally significant for Colombia).
We also examine the wage rate for skilled and unskilled workers separately. As for the average
wage rate, we find a premium for exporters in both skilled and unskilled wage rates across all
specifications for India. The premium is larger for skilled (non-production) workers than production
workers. For Chile, we find a significant premium for skilled workers across all specifications. There
is also a wage premium for unskilled workers except when we condition on revenue alone (column
2). Again, the estimated premium for exporters is larger for skilled workers in all specifications.
For Colombia, the results are similar to Chile. We find that the wage premium for skilled workers
is significant across all specifications, while the premium for unskilled workers is not significant
once we condition on size using total revenues (columns 2 and 3). Again, the magnitude of the
estimated wage premia are larger for skilled workers compared to unskilled workers.
The non-parametric analysis of the difference in the overall average wages between exporters
and non-exporters is presented in Figure 9. In India and Chile, we find that the wage rate is
premium is higher for exporters across the size distribution, with the biggest differences at the top
of the distribution. The differences are smaller for Colombia; here too, the largest differences are
34
at the top of the distribution.
Overall, we conclude that there is significant evidence for higher ability of workers (especially
skilled workers) employed by exporters, conditional on size.
4.2.5 Other proxies for quality
In Table 5, we look at other proxies for quality that we can form using information in the ASI
1998 dataset for India. In the first row of Table 5, we look at a dummy variable for ISO 9000
certification. We find that, within 4 digit NIC industries, exporters are more likely to obtain ISO
9000 quality certification, and this holds across all seven of our specifications.
The rest of the Table 6 reports results for the 5 dummy variables indicating adoption of computer
technology (see section 3 for a description of these variables). For all 5 of the variables, we find that
exporters are more likely to adopt computer technology. The effects are statistically significant at
the 1% level, across all specifications. Conditioning on revenue (column 2), we find that a 17.5%
greater probability of an exporter having computers on the facility (row 2), a 5% greater propensity
to use robotics (row 3), a 4.7% greater propensity to have a computer network (row 4), and a 12%
greater likelihood of having internet access (row 5). Also, there is an 8% greater likelihood of
exporters responding to the survey on floppies. In all specifications, we include 4 digit NIC code
industry dummies.
We interpret this as supportive evidence that there are significant quality differences between
exporters and non-exporters (even conditioning on size).
4.2.6 Results for US data
Finally, in Table 7, we reproduce results for the US from Table 2 of Bernard and Jensen (1999).
Using data on manufacturing plants from the Longitudinal Research Database of the United States
Bureau of Census, Bernard and Jensen show that exporters have a premia on a number of dif-
ferent measures. The results reproduced here show that both capital intensity and skill intensity
(measured using wage rates) is higher for exporters. All the reported effects are significant at 1%
level. Also, the specification 2 for each of the years includes a control for size (measured a log
total employment). Thus, even for the US, the findings are consistent with those for the three
developing countries – exporters exhibit a significantly higher levels of capital and skill intensity
35
relative to non-exporters, even when we control for size. The results for the U.S. suggest that the
reason exporters produce higher quality than non-exporters cannot only be due to the fact that
they export to countries that consume higher quality, as this is unlikely to be the case of U.S. firms
selling in foreign markets.
We conclude that on a large number of proxies for quality, exporters exhibit a premium rel-
ative to non exporters. These effects are highly robust to the inclusion of linear and non-linear
controls for size. Even when we control non-parametrically for size, we find that exporters are gen-
erally likely to have higher quality levels compared to non-exporters. This evidence is consistent
with the predictions of our model, and suggest that the modification to the standard single-factor
heterogenous firm models provide a better fit to the data.
5 Conclusion and future work
In this paper, we present a model of industry equilibrium with heterogeneous firms where firms differ
on two dimensions. The first source of heterogeneity is productivity, which reduces the marginal
costs of the firm. The second dimension, which is novel in our model, is caliber which endows
the firm with a cost advantage in producing higher quality of output. Higher quality shifts out
the demand for the product, or equivalently enables the firm to charge higher prices. We propose
that quality plays an important role in determining the export status of a firm. In particular, we
propose that firms need to possess a minimum quality threshold in order to export, independent of
their productivity level.
Thus in our model, exporters differ from exporters in quality as well as productivity. This leads
to a number of interesting implications, including the following: (i) Size is not the sole determinant
of export status; if productivity is high but caliber is low, a firm may be large but restrict itself
to the domestic market. Similarly a high caliber but low productivity firm may enter the export
market though its size is small. (ii) The price charged by exporting firms will be larger than that
for non-exporting firms, conditional on size. This is a particularly sharp prediction of our theory,
as a productivity based explanation of exporting would suggest that exporters charge lower prices
(exploiting their lower marginal cost). (iii) Under certain additional assumptions (mainly that there
is a input price advantage to using capital rather than labor to produce quality), capital intensity
36
(defined as the ratio of capital to labor) will be higher for exporters relative to non-exporters,
conditional on size. (iv) Under the assumption that it is relatively cheaper to use skilled workers
than unskilled workers to produce quality, skill intensity (measured as the share of skilled workers
in total employment) will be higher for exporters relative to non-exporters, conditional on size.
(v) Under the assumption that producing higher quality requires workers of greater ability, worker
ability (as measured by the wage rate) would be higher for exporters relative to non-exporters,
conditional on size.
We test these predictions using establishment-level data from India, Chile and Colombia, and
find strong support for the predictions of the model. Since we wish to highlight the role of quality
distinct from productivity, we are careful about controlling for size, and test the robustness of our
results to a number of alternative ways to condition on size.
Our model proposes that firms that have high caliber (and productivity) draws would choose to
produce higher quality, and through this choice, self-select into exporting. Hence we do not expect
any change in quality or other relevant variables such as capital intensity, skill intensity or worker
ability as a result of a firm changing status from exporter to non-exporter.26 We use panel data
from Chile and Colombia to explore this question (see Appendix 4), and find that indeed there is
not much difference in these variables around the time that firms switch export status. Thus the
evidence is consistent with our model of self-selection.
Our results should be interpreted subject to a few caveats. Price is the closest proxy we have
for quality and is the variable for which a theory on productivity differences alone has opposite
predictions from our theory which considers quality in addition to productivity. Thus we interpret
the empirical finding that price levels are higher for exporters relative to non-exporters as strong
support for our theory. Unfortunately, price data is not widely available and we are restricted to
using data on India for one year where this is available. Further testing with other datasets where
price is available would provide further support.
All other proxies, such as capital intensity, skill intensity, and worker ability, may be correlated
with productivity so that these differences may reflect underlying productivity differences. While
we check the robustness of our results to conditioning on size in a variety of ways, there is a26Alternatively, the firms may learn about quality through experiences in the export market. If this was primarily
the case, we will expect to see a significant increase in variables correlated with quality subsequent to exporting.
37
possibility of some residual bias.27
We have been intentionally agnostic about the source of higher quality requirement for exporters
and kept the modelling of the minimum quality requirement simple for a couple of reasons. One, as
ours is one of the first attempts at incorporating two sources of heterogeneity, we wanted to keep
the modelling as simple as possible.28 Two, the particular source of the quality requirement is not
important for the empirical tests that we perform. However, it would be interesting to explore the
sources for higher quality requirement for exports. We have discussed a number of possible sources
of this requirement in Section 2.2.2. Further work to explore the importance of particular sources
of the quality requirement would be illuminating.29
References
[1] Melitz, M. J., 2003. “The impact of trade on intra-industry reallocations and aggregate industry
productivity,” Econometrica, 71(6), 1695-1725.
[2] Rauch, James E., 1999. “Networks Versus Markets in International Trade,” Journal of Inter-
national Economics 48(1), 7-35.
[INCOMPLETE]
27An alternative way to check if capital intensity, skill intensity and worker ability do indeed reflect quality dif-
ferences would be possible if there were instances where exporters were forced to upgrade quality, for example by
imposition of new standards by the US or the European Union. Then we could test if the proxies for quality did
indeed increase following the imposition of higher standards. Also, we could check if firms that drop out of export
market following the increase in standard have lower measures of capital intensity, skill intensity and worker ability.28Other recent papers that introduce multiple sources of heterogeneity include Bernard, Redding and Schott (2006),
Rauch (2006), Sutton (2007) and Vogel (2007).29Firm-level data on destinations of exports as well as prices charged could be useful in distinguishing between
alternative stories. We did not have access to such a dataset, but these are becoming increasingly available, albeit
under confidentiality restrictions (e.g. Kramarz and Kortum? who use such data for France, Andersson (2006) who
uses such data on Swedish firms).
38
Appendix 1: Proof of Proposition 1
The first thing to note is that the function ξ (ϕ) is increasing in P . Note that this function enters
(19) through η̂, as it is the lower limit of the second integral in (20). In (19), the left-hand-side
(LHS) is constant while the right-hand-side (RHS) can be expressed as a function of P . When
P → 0, then ξ (ϕ) → 0 ∀ϕ, which implies that Pin → 1. Also, in that case, η̂ takes some finite
positive value.30 As a result, the RHS → ∞. In contrast, when P → ∞, then ξ (ϕ) → ∞ ∀ϕ,
which implies that Pin → 0. Since η̂ → 0, the RHS → 0. Therefore, to prove existence and
uniqueness, we only need to show that the RHS of (19) is decreasing in P .
The derivative of the RHS of (19) with respect to P is:
Bαα′
(fe + PinF )2
∂
(η̂/P
αα′
)
∂P(fe + PinF )− ∂Pin
∂PF
η̂
Pα′α
=B
αα′
(fe + PinF )2
[1
Pαα′
[∂η̂
∂P− α
α′η̂
P
](fe + PinF )− ∂Pin
∂P
F η̂
Pαα′
].
The sign of this derivative depends on the sign of
− α
α′η̂
P(fe + PinF ) +
∂η̂
∂P(fe + PinF )− ∂Pin
∂PF η̂.
We will now show that the last two terms of this expression cancel out with one another, leaving
only the first term, which is negative, to determine the sign of the derivative.
Using equations (17) and (10) together with η̂ ≡ Pinη̃, we find:
fe + PinF = J
(1c
) (σ−1)α
α′(
1f
)α−α′α′
(E
P
) αα′
η̂ (33)
Then, we can calculate∂η̂
∂P= −
∫
ϕ
ϕ(σ−1)α
α′ ξ (ϕ)α−α′
α′∂ξ
∂Pk(ϕ)dϕ (34)
and∂Pin
∂P= −
∫
ϕ
∂ξ
∂Pk(ϕ)dϕ. (35)
30We assume that the bivariate distribution has finite relevant moments so that the integral on the RHS of 20 is
well defined.
39
Finally, we can show that
∂η̂
∂P(fe + PinF )− ∂Pin
∂PF η̂
= η̂
∫
ϕ
ϕ(σ−1)α
α′ ξ (ϕ)α−α′
α′ J
(1c
) (σ−1)α
α′(
1f
)α−α′α′
(E
P
) αα′ ∂ξ
∂Pk(ϕ)dϕ
+η̂
∫
ϕ
∂ξ
∂PFk(ϕ)dϕ
= η̂
∫
ϕ
ϕ
(σ−1)αα′ ξ (ϕ)
α−α′α′ J
(1c
) (σ−1)α
α′(
1f
)α−α′α′
(E
P
) αα′ − F
∂ξ
∂Pk(ϕ)dϕ
= 0
where the first equality uses (33), (34), and (35), the second equality reorganizes terms, and the
last equality uses (12).
40
Appendix 2: Existence and uniqueness of equilibrium for the asymmetric
constrained export quality case
The sources of asymmetry are: E and E∗, P and P ∗, v(ϕ, ξ) and v∗(ϕ, ξ), λ and λ∗, F and F ∗,
fx and f∗x . The marginal cost function (3) and the fixed cost function (4) are also allowed to be
asymmetric: c and c∗, f and f∗. However, these last sources of asymmetry can be thought to be
implicitly subsumed in ϕ and ξ (i.e. ϕ = ϕ̃/c and ξ = ξ̃/f where ϕ̃ and ξ̃ are the original draws).
Then, ϕ and ξ can then be interpreted as “cost-adjusted productivity and caliber”, with distribu-
tions v(ϕ, ξ) and v∗(ϕ, ξ) reflecting the degree of international competitiveness of the industry (a
combination of technology and input/labor cost advantage).
Ex-ante expected profits are given by ϕ and ξ]
Π(P, P ∗) =
ϕ∫
ϕ
ξ∫
ξ
Π(ϕ, ξ, P, P ∗) v (ϕ, ξ) dξdϕ, P > 0, P ∗ > 0
where
Π (ϕ, ξ, P, P ∗) =
0 if (ϕ, ξ) ∈ r(I) ={(ϕ, ξ) : ξ ≤ ξ(ϕ)
}
Πd(ϕ, ξ, P ) if (ϕ, ξ) ∈ r(II) ={
(ϕ, ξ) : ξ(ϕ) < ξ ≤ ξu(ϕ)
}or
(ϕ, ξ) ∈ r(III) ={
(ϕ, ξ) : ϕ > ϕλ and ξu(ϕ) < ξ ≤ ξ
x(ϕ)
}
Πc (ϕ, ξ, P, P ∗) if (ϕ, ξ) ∈ r(IV ) ={
(ϕ, ξ) : ϕ > ϕλ and ξx(ϕ) < ξ ≤ ξ
λ(ϕ)
}
Πu (ϕ, ξ, P, P ∗) if (ϕ, ξ) ∈ r(V ) ={
(ϕ, ξ) : ϕ > ϕλ and ξλ(ϕ) < ξ
}or
(ϕ, ξ) ∈ r(V I) ={
(ϕ, ξ) : ϕ ≤ ϕλ and ξu(ϕ) < ξ
}.
At the limits between regions, the function Π (ϕ, ξ, P, P ∗) does not jump – firms are indifferent
between the export status that characterize the limiting regions. Therefore, Π (ϕ, ξ, P, P ∗) is con-
tinuous in (ϕ, ξ), even though at the limits of integration – of measure zero in R2 – the function is
not differentiable.
41
Expected profits can then be written as:
Π(P, P ∗) =ϕ∫ϕ
ξ∫ξ
Π(ϕ, ξ, P, P ∗) v (ϕ, ξ) dξdϕ
=ϕ∫ϕ
ξu(ϕ,P,P ∗)∫
ξ(ϕ,P,P ∗)Πd (ϕ, ξ, P, P ∗) v (ϕ, ξ) dξdϕ +
ϕ∫ϕλ(P,P ∗)
ξx(ϕ,P,P ∗)∫
ξu(ϕ,P,P ∗)
Πd (ϕ, ξ, P, P ∗) v (ϕ, ξ) dξdϕ
+ϕ∫
ϕλ(P,P ∗)
ξλ(ϕ,P,P ∗)∫
ξx(ϕ,P,P ∗)
Πc (ϕ, ξ, P, P ∗) v (ϕ, ξ) dξdϕ
+ϕ∫
ϕλ(P,P ∗)
ξ∫ξ
λ(ϕ,P,P ∗)
Πu (ϕ, ξ, P, P ∗) v (ϕ, ξ) dξdϕ +ϕλ(P,P ∗)∫
ϕ
ξ∫ξ
u(ϕ,P,P ∗)
Πu (ϕ, ξ, P, P ∗) v (ϕ, ξ) dξdϕ
(36)
where, as previously shown,
Πd(ϕ, ξ, P ) = Jϕα(σ−1)
α′ ξα−α′
α′(
EP
) αα′ − F
Πu(ϕ, ξ, P, P ∗) = Jϕα(σ−1)
α′ ξα−α′
α′(
EP + τ−σ E∗
P ∗) α
α′ − F − fx
Πc(ϕ, ξ, P, P ∗) = 1σ
(σ
σ−1
)1−σϕσ−1λα−α′ (E
P + τ−σ E∗P ∗
)− 1ξ λα − fx − F.
The functions Πd, Πu, and Πc are continuous and differentiable in P and P ∗, as are also the
limits of integration. The continuity and differentiability of the function Π(P, P ∗) in P and P ∗
then follows directly.
By application of Leibniz rule, since Π (ϕ, ξ, P, P ∗) is continuous in (ϕ, ξ) – even though not
everywhere differentiable – the derivatives of Π(P, P ∗) with respect to P and P ∗, respectively, result
in expressions analogous to (36), except that they integrate over derivatives instead of function
values:
∂Π(P,P ∗)∂P =
ϕ∫ϕ
ξu(ϕ,P,P ∗)∫
ξ(ϕ,P,P ∗)
∂Πd(ϕ,ξ,P,P ∗)∂P v (ϕ, ξ) dξdϕ +
ϕ∫ϕλ(P,P ∗)
ξx(ϕ,P,P ∗)∫
ξu(ϕ,P,P ∗)
∂Πd(ϕ,ξ,P,P ∗)∂P v (ϕ, ξ) dξdϕ
+ϕ∫
ϕλ(P,P ∗)
ξλ(ϕ,P,P ∗)∫
ξx(ϕ,P,P ∗)
∂Πc(ϕ,ξ,P,P ∗)∂P v (ϕ, ξ) dξdϕ
+ϕ∫
ϕλ(P,P ∗)
ξ∫ξ
λ(ϕ,P,P ∗)
∂Πu(ϕ,ξ,P,P ∗)∂P v (ϕ, ξ) dξdϕ +
ϕλ(P,P ∗)∫ϕ
ξ∫ξ
u(ϕ,P,P ∗)
∂Πu(ϕ,ξ,P,P ∗)∂P v (ϕ, ξ) dξdϕ
(37)
42
and (note that ∂Πd(ϕ,ξ,P,P ∗)∂P ∗ = 0)
∂Π(P,P ∗)∂P ∗ =
ϕ∫ϕλ(P,P ∗)
ξλ(ϕ,P,P ∗)∫
ξx(ϕ,P,P ∗)
∂Πc(ϕ,ξ,P,P ∗)∂P ∗ v (ϕ, ξ) dξdϕ
+ϕ∫
ϕλ(P,P ∗)
ξ∫ξ
λ(ϕ,P,P ∗)
∂Πu(ϕ,ξ,P,P ∗)∂P ∗ v (ϕ, ξ) dξdϕ +
ϕλ(P,P ∗)∫ϕ
ξ∫ξ
u(ϕ,P,P ∗)
∂Πu(ϕ,ξ,P,P ∗)∂P ∗ v (ϕ, ξ) dξdϕ.
(38)
Since ∂Πi∂P < 0, i = d, u, c, ∀(P, P ∗, ϕ, ξ) and ∂Πi
∂P ∗ < 0, i = u, c, ∀(P, P ∗, ϕ, ξ), from (37) and (38) we
can easily establish that
∂Π(P, P ∗)∂P
< 0, ∀(P, P ∗)
∂Π(P, P ∗)∂P ∗ < 0, ∀(P, P ∗).
Analogously, for the foreign country we can also derive the expect profit function Π∗(P, P ∗),
with
∂Π∗(P, P ∗)∂P
< 0, ∀(P, P ∗) (39)
∂Π∗(P, P ∗)∂P ∗ < 0, ∀(P, P ∗). (40)
Free-entry in each country defines the following system of equations:
Π (P, P ∗) = fe (41)
Π∗ (P, P ∗) = f∗e (42)
We want to show that an equilibrium pair (P, P ∗) exists and that it is unique. First, we impose
the following assumptions:
Assumption 1 :
limP→∞
Π(P, P ∗) < limP→∞
Π∗ (P, P ∗)
Assumption 2 :
limP ∗→∞
Π(P, P ∗) > limP ∗→∞
Π∗ (P, P ∗)
The two assumptions are analogous. When P →∞ there are no profits to be made in the Home
market so firms only operate in the Foreign market. Then, the first assumption simply states that
Foreign firms’ expected profits in the Foreign market – for any P ∗ – is higher than Home firms’
43
expected profits in that market. Analogously, the second assumption states that Home firms’
expected profits in the Home market are higher than Foreign firms’ profits there.
Proposition 3. : Under assumptions 1 and 2, there exists a pair (P, P ∗) that solves the system of
equations (41) and (42).
Proof. Since Π (P, P ∗) is strictly decreasing in P ∗, for any given P the value of P ∗ that solves
equation (41) is unique and implicitly defines a function P ∗ = P ∗H(P ). Similarly, since Π (P, P ∗) is
strictly decreasing in P , we can obtain the inverse function P = PH(P ∗). Using the Implicit Func-
tion Theorem and the results in (39) and (40), we easily establish that this function is downward
sloping:dP ∗
dP
∣∣∣∣H
= − ∂Π(P, P ∗) /∂P
∂Π(P, P ∗) /∂P ∗ < 0.
Analogously, equation (42) defines P ∗ = P ∗F(P ) and P = PF (P ∗) with slope
dP ∗
dP
∣∣∣∣F
= − ∂Π∗ (P, P ∗) /∂P
∂Π∗ (P, P ∗) /∂P ∗ < 0.
The proof of existence is represented in Figure 5. Assumption 1 implies, in particular, that
fe = limP→∞
Π(P, P ∗H
(P ))
< limP→∞
Π∗(P, P ∗H
(P ))
.
Since Π∗ (P, P ∗) is decreasing in P ∗, the above inequality also implies that
limP→∞
P ∗H(P ) < lim
P→∞P ∗F
(P ). (43)
Analogously, assumption 2 implies that
f∗e = limP ∗→∞
Π∗(PF (P ∗), P ∗) < lim
P ∗→∞Π
(PF (P ∗), P ∗)
which, together with the fact that Π (P, P ∗) is decreasing in P , implies that
limP ∗→∞
PF (P ∗) < limP ∗→∞
PH(P ∗). (44)
Since both P ∗H(P ) and P ∗F
(P ) are decresing, (43) and (44) imply that these two curves must
cross at least once. Thus, an equilibrium exists.
To show uniquess we only need to demonstrate that the two curves satisfy the single crossing
property. In particular, we need to show that∣∣∣ dP ∗
dP
∣∣H
∣∣∣ >∣∣∣ dP ∗
dP
∣∣F
∣∣∣
44
Calculating the derivatives that are present in (37) and (38), we can write the absolute value ofdP ∗dP
∣∣H
and dP ∗dP
∣∣F, respectively, as
∣∣∣ dP ∗dP
∣∣H
∣∣∣ =A + E
P 2 B
τ−σ E∗P ∗2
B(45)
and ∣∣∣ dP ∗dP
∣∣F
∣∣∣ =τ−σ E
P 2 D
C + E∗P ∗2
D(46)
where
A = J αα′
(EP
) αα′ P−1
ϕ∫
ϕ
ξu(ϕ,P,P ∗)∫
ξ(ϕ,P,P ∗)η (ϕ, ξ) v (ϕ, ξ) dξdϕ +
ϕ∫ϕλ(P,P ∗)
ξx(ϕ,P,P ∗)∫
ξu(ϕ,P,P ∗)
η (ϕ, ξ) v (ϕ, ξ) dξdϕ
B = 1σ
(σ−1
σ
)σ−1λα−α′
ϕ∫ϕλ(P,P ∗)
ξλ(ϕ,P,P ∗)∫
ξx(ϕ,P,P ∗)
ϕσ−1v (ϕ, ξ) dξdϕ
+J αα′
(EP + τ−σ E∗
P ∗)α−α′
α′
ϕ∫
ϕλ(P,P ∗)
ξ∫ξ
λ(ϕ,P,P ∗)
η (ϕ, ξ) v (ϕ, ξ) dξdϕ +ϕλ(P,P ∗)∫
ϕ
ξ∫ξ
u(ϕ,P,P ∗)
η (ϕ, ξ) v (ϕ, ξ) dξdϕ
C = J αα′
(E∗P ∗
) αα′ P ∗−1
ϕ∫
ϕ
ξ∗u(ϕ,P,P ∗)∫
ξ∗(ϕ,P,P ∗)η (ϕ, ξ) v∗ (ϕ, ξ) dξdϕ +
ϕ∫ϕ∗λ(P,P ∗)
ξ∗x(ϕ,P,P ∗)∫
ξ∗u(ϕ,P,P ∗)
η (ϕ, ξ) v∗ (ϕ, ξ) dξdϕ
D = 1σ
(σ−1
σ
)σ−1λ∗
α−α′ ϕ∫ϕ∗λ(P,P ∗)
ξ∗λ(ϕ,P,P ∗)∫
ξ∗x(ϕ,P,P ∗)
ϕσ−1v∗ (ϕ, ξ) dξdϕ
+J αα′
(E∗P ∗ + τ−σ E
P
)α−α′α′
ϕ∫
ϕ∗λ(P,P ∗)
ξ∫ξ∗
λ(ϕ,P,P ∗)
η (ϕ, ξ) v∗ (ϕ, ξ) dξdϕ +ϕ∗λ(P,P ∗)∫
ϕ
ξ∫ξ∗
u(ϕ,P,P ∗)
η (ϕ, ξ) v∗ (ϕ, ξ) dξdϕ
Since A > 0, B > 0, C > 0, D > 0, it is easy to show with simple algebra that (45) > (46), ∀(P, P ∗).
Thus, there is a unique equilibrium.
45
Appendix 3: Proof of Corollary 2
Proof. In the variable cost part, conditional on output Y and quality level λ, the capital to skilled
labor ratio is given by:31
K
Ls=
ws
wk
αk
αs=
ws
wk
αk
αsλbs−bk
Similarly, in the fixed cost part we get:
K ′
L′s=
ws
wk
αk
αs=
ws
wk
αk
αsλbs−bk
So we get the overall capital to skilled labor ratio is given by:
Ktotal
Ltotals
=K + K ′
Ls + L′s=
ws
wk
αk
αsλbs−bk
Since bs > bk, we get thatd
(Ktotal
Ltotals
)
dλ> 0. Similarly, following from the assumption that bu > bk, we
get we getd
(Ktotal
Ltotalu
)
dλ> 0. These two results together imply that
d(
Ktotal
Ltotal
)
dλ=
d(
KtotalLtotal
u +Ltotals
)
dλ> 0.
Hence we can write the overall capital to labor ratio as
Ktotal
Ltotal= k(λ,Z),
where Z = (ws, wu, wk, αs, αu, αk) is a vector of other relevant variables and where k is a strictly
increasing function of λ. Then, so long as we have:
Z ⊥⊥ (D|r), (47)
where D is a dummy denoting export status, i.e. as long as the wage rates, rental prices and the
technology parameters are the same for exporters and non-exporters (conditional on revenue),32 it
follows from Proposition 2 that:
E
[(Ktotal
Ltotal
)|D = 1, r = R̄
]= E[k(λ,Z)|D = 1, r = R̄]
> E[k(λ,Z)|D = 0, r = R̄] = E
[(Ktotal
Ltotal
)|D = 0, r = R̄
]
31Note that since our underlying Cobb-Douglas production function is homothetic, the capital-labor as well as the
skilled-unskilled worker ratio is independent of output.32This condition states that all other determinants of the capital-labor ratio, other than quality, should be the
same for exporters and non-exporters. These are broad sufficient conditions. More narrowly, it would be sufficient
that
�wi
r
αk
αi
�⊥⊥ (D, r), i ∈ (s, u) This condition would also be relevant in interpreting our empirical results.
46
Appendix 4: Analysis of Switching into Export Status
In this Appendix, we analyze changes in the relevant dependent variables within firms, conse-
quent to their changing status as exporter or non-exporter. We do this by analyzing the panel data
available for Chile and Colombia, and using a firm fixed effects regression. We undertake two types
of analysis.
A Fixed effects regressions
First, we run the following fixed effects regression:
logYit = δDit + αlogXit + Di + εit (48)
This specification is similar to equation 32, except for the important difference that we include firm
fixed effects (Di) here. Consequently, here the coefficient on the exporter dummy (δ) measures the
changes in the dependent variable within firms who change status from exporter to non-exporter
(and vice versa). As in the analysis in section 4, we control for size in the following different
ways: (i) including log sales; (ii) including log sales and square of log sales; (iii) including log
employment; (iv) including log employment and square of log employment; (v) including industry-
year-size decile fixed effects; and (vi) including industry-year-size decile-age quintile fixed effects.
To overcome the computational difficulty involved in including both firm as well as other industry-
year-size or industry-year-size-age fixed effects in the same regression, we control for (v) and (vi)
using a two step procedure. In the first step we obtain residuals from regressing the dependent
variable on industry-year-size and industry-size-year-age fixed effects, and in the second stage we
use the demeaned residuals in the specification in equation 48 above. The age control is particularly
important here as some of the changes in the firms following their change from exporter to non-
exporter status (or vice versa) could arise because of independent age related effects.
The results are presented in Appendix Table 1. Panel 1 presents the results for Chile and Panel 2
does the same for Colombia. We find on Chile that the capital intensity and skill intensity variables
show no statistically or economically significant changes within firms, even when no controls are
included (column 1). There is some evidence of increases within firms conditional on employment
(4 and 5) but none when size is controlled for based on sales (2 and 3) or using dummies for size or
47
size and age (6 and 7). For worker ability (wage variables), there is some evidence of an increase
within firms unconditionally (column 1) and conditional on employment (columns 4 and 5) but
not when sales are controlled for (columns 2 and 3) or industry-year-size and age fixed effects are
included(columns 6 and 7). Taken together, the evidence for Chile suggests that there is very little
change in the dependent variables within firms following a change in their export status.
The results for Colombia (Panel 2) suggest a larger change within firms across all variables,
unconditionally (column 1), conditional on employment (columns 3 and 4), and even conditional
on sales (column 3). However most effects disappear once we condition on industry-size-year or
industry-size-year-age dummies. (The one significant variable is log of skilled wage, which contrary
to our prediction appears to decline for exporters. This result is somewhat puzzling, especially
given the unconditional positive effect in Column 1. In contrast, there is a marginally significant
increase in the unskilled wage.) For Colombia too, we conclude that there is no significant change
in the dependent variables once size effects are controlled for carefully.
Overall, conditional on size, there appears to be no significant change in the dependent vari-
ables following a change in export status of firms. This suggests that the differences across firms
documented in the main analysis in Section 4.2 largely reflects pre-existing differences between
exporters and non-exporters.
B Analysis of switchers around the time of switching
In the second type of analysis, we focus attention on changes in firms that switch from non-exporter
status to exporter status, around the time (4 years before and after) that they switch status. To
do this, we remove from the data firms that move in and out of the export markets, as well as data
more than 4 years before or later than 4 years after the firm switches into the export market. As
controls, we include all firms that remain non-exporters during the entire panel period. We then
perform the following two stage analysis.
In the first stage, we regress the dependent variables on industry-year-size decile-age quintile
fixed effects, so as to demean and remove influences of industry, year, size and age effects from the
data. Then we form an index defined as the number of years since exporting. Given the truncation
of data on exporters,the index varies from a maximum of 4 to a minimum of -4. For the firms
48
that never export, we define the index as a low negative number (-100). We then run the following
regression:
logY residualit =
4∑
k=−4
δkDkit + εit (49)
Thus, in our regressions, non-exporters would form the omitted category and the coefficients on
the index would represent the mean difference between exporters in the index year and all non-
exporters. We then plot the coefficients and the 5 percent confidence interval on the indices.
The results for Chile are plotted in Appendix figure 1a. The capital intensity variable shows
an upward trend after the switch, but the increase is small.The skill intensity variables do not
show a big change around the time of the switch to export status. There is some decline in the
skill intensity variables (skilled share of wage bill and skilled share of employment) but these are
noisy estimates (a large confidence band) and the difference with the pre-exporting mean does not
appear to be statistically significant (the confidence bands overlap considerably). On worker ability
variables (average wage rate, skilled wage rate and unskilled wage rate), there appears to be some
increase after the switch to exporter status, but they appear to be in line with pre-switch trends
and the difference in pre and post-switch means do not appear to statistically significant given the
confidence intervals.
Results for Colombia are in Appendix Figure 1b. Here there appears to be some little increase
post switch for the capital to labor ratio and for the skill intensity variables. The trends for ability
(wage) variables also appear pretty flat. In all cases the post export confidence interval overlaps
significantly with the pre-export confidence intervals.
Overall, for both Chile and Colombia, we conclude that there are no large breaks in trend
following a change in export status. Where there appear to be some increase in the mean, they
are generally do not seem to be statistically significant. 33 Note that in almost all the figures, the
intercept is positive, consistent with positive pre-existing differences between non-exporters and
firms that choose to export.
33We confirmed this using a formal statistical test, by replacing the index with an exporter dummy and including
firm fixed effects in specification 49. Except for some weak evidence of a small increase in wage rate in Colombia, in
all other cases the mean difference was small and statistically insignificant.
49
Figure 1: Percentage of establishments that are exporters, by size percentile
0.1
.2.3
.4.5
.6.7
.8
0 10 20 30 40 50 60 70 80 90 100Size ptile (Adj. for Ind. Mean)
Country: India
0.1
.2.3
.4.5
.6.7
.8
0 10 20 30 40 50 60 70 80 90 100Size ptile (Adj. for Ind. Mean)
Country: Chile
0.1
.2.3
.4.5
.6.7
.8
0 10 20 30 40 50 60 70 80 90 100Size ptile (Adj. for Ind. Mean)
Country: Colombia
The Y axis is the fraction of firms that are exporters. Size percentiles are adjusted for industry mean size.
Percentage exporters by size percentile
Figure 2: Unconstrained export quality equilibrium
ξ
ϕ
ξ (ϕ)
Non Survivors Domestic firms
Exporters (sell domestic and abroad)
ξx.u(ϕ)
ξ (ϕ): cutoff between survivors and non survivors
ξx.u(ϕ): export cutoff in the unconstrained regime
Figure 3: Constrained export quality equilibrium
ϕ
Region I: Non Survivors
ξx.u(ϕ)
ξ (ϕ): cutoff between survivors and non survivors
ξx.u(ϕ): export cutoff in the unconstrained regime
ξx.λ(ϕ): the iso-quality curve for threshold quality λ
ξx
ξx: the export cutoff in the constrained regime
Region III: Domestic firms
Region II: Domestic firms
Region IV: Exporters
Region V: Exporters
ξx.λ(ϕ)
ϕ*λ
ξ( )ϕ
Figure 4: Iso-revenue and iso-quality curves in the constrained export quality equilibrium
ξx.λ(ϕ): the iso-quality curve for threshold quality λ
ξ (ϕ): cutoff between survivors and non survivors
ξx: the export cutoff in the constrained regime
ξx.u(ϕ): export cutoff in the unconstrained regime
ξ (ϕ)
ξx.λ(ϕ)
Region I: Non Survivors
R1
ξx
E
A
R2
ϕ
ξx.u(ϕ)Region II: Domestic firms
Region III: Domestic firms
B
D
ξ R1 R2 A
Region V: Exporters
Region IV: Exporters [Iso-quality (λ)] C
Iso-revenue
Figure 5: Existence of equilibrium
P*
P
P*F (P)
P*H (P)
Table 1: Summary statistics All summary statistics are adjusted for sampling weights.
Description N Mean SD P25 Median P75 Panel 1: India (1998) Log per unit price 43,197 7.79 2.87 5.51 8.39 9.86 Export dummy (=1 if exports>0) 28,584 0.09 0.29 0 0 0 Log (capital/labor) 27,316 10.58 2.06 9.6 10.75 11.83 Skilled share of employment 27,964 0.30 0.21 0.16 0.25 0.39 Skilled share of wage bill 27,472 0.35 0.24 0.17 0.32 0.5 Log(average wage rate) 27,438 10.02 0.77 9.57 10.05 10.5 Log(skilled wage rate) 26,090 10.30 0.89 9.75 10.37 10.93 Log(unskilled wage rate) 26,891 9.91 0.73 9.52 9.94 10.32 Log(total employment) 27,964 4.88 1.24 4.14 4.85 5.68 Log(total sales) 27,159 17.41 2.05 16.05 17.44 18.84 Square of log(total employment) 27,964 25.32 12.25 17.17 23.54 32.23 Square of log(total sales) 27,159 307.24 70.71 257.7 304.24 354.92 Age in years 28,087 27.45 134.22 7.00 14.00 24.00 ISO 9000 certification dummy 28,584 0.03 0.18 0 0 0 Computer use dummy 28,584 0.32 0.47 0 0 1 Robotics use dummy 28,584 0.02 0.14 0 0 0 Computer network dummy 28,584 0.02 0.13 0 0 0 Internet use dummy 28,584 0.04 0.19 0 0 0 Survey response on floppy dummy 28,584 0.06 0.25 0 0 0 Panel 2: Chile (1991-96) Export dummy (=1 if exports>0) 34,990 0.20 0.40 0 0 0 Log (capital/labor) 33,817 4.76 1.61 3.91 4.74 5.68 Skilled share of employment 34,990 0.27 0.19 0.14 0.23 0.35 Skilled share of wage bill 34,989 0.35 0.24 0.17 0.32 0.5 Log(average wage rate) 34,989 7.27 0.75 6.74 7.26 7.78 Log(skilled wage rate) 32,998 7.58 0.99 6.87 7.64 8.32 Log(unskilled wage rate) 34,220 7.06 0.68 6.59 7.05 7.5 Log(total employment) 34,990 3.72 1.02 2.94 3.5 4.32 Log(total sales) 34,982 12.94 1.63 11.75 12.67 13.92 Square of log(total employment) 34,990 14.91 8.61 8.67 12.23 18.64 Square of log(total sales) 34,982 170.01 44.12 138.04 160.47 193.67 Age in years (imputed) 34,990 9.28 5.52 4.00 11.00 14.00 Panel 3: Colombia (1981-91) Export dummy (=1 if exports>0) 75,812 0.12 0.33 0 0 0 Log (capital/labor) 75,069 5.64 1.61 4.62 5.66 6.69 Skilled share of employment 75,770 0.32 0.19 0.18 0.28 0.41 Skilled share of wage bill 75,300 0.34 0.22 0.18 0.31 0.47 Log(average wage rate) 75,299 6.09 0.92 5.43 6.1 6.75 Log(skilled wage rate) 71,472 6.20 1.08 5.47 6.22 6.98 Log(unskilled wage rate) 74,230 6.02 0.85 5.39 6.03 6.66 Log(total employment) 75,770 3.45 1.10 2.64 3.26 4.06 Log(total sales) 75,757 11.20 1.86 9.9 10.98 12.3 Square of log(total employment) 75,770 13.13 8.69 6.96 10.62 16.49 Square of log(total sales) 75,757 128.84 43.77 98.07 120.48 151.32 Age in years 75,812 15.39 12.61 6.00 12.00 21.00
Table 2: Per unit price (India 1998) The dependent variable is log per unit price. All reported figures are coefficients on an exporter dummy which equals one for exporting establishments and is zero for non-exporters. The first row reports results from regressions that include 4-digit industry-unit code fixed effects, while the second row includes 3 digit industry-unit code fixed effects. The number of observations is about 43000. Standard errors are clustered at 4-digit industry level. In column 6 and 7, the industry definition in the industry-size-year and industry-size-year-age effects are defined for 4 digit industry-unit code level (in row 1) and in the 3 digit industry-unit code level (in row 3). + significant at 10%; * significant at 5%; ** significant at 1%.
Type of fixed effects (1) (2) (3) (4) (5) (6) (7) 4-digit industry-unit code 0.354 0.192 0.183 0.279 0.277 0.193 0.177 [0.059]** [0.062]** [0.063]** [0.061]** [0.062]** [0.063]** [0.070]* 3-digit industry-unit code 0.379 0.200 0.187 0.304 0.299 0.156 0.177 [0.065]** [0.067]** [0.067]** [0.067]** [0.067]** [0.064]* [0.066]** Other controls: Log (sales) No Yes Yes No No No NoSquare of log (sales) No No Yes No No No NoLog (employment) No No No Yes Yes No NoSquare of log (employment) No No No No Yes No NoIndustry-size effects No No No No No Yes NoIndustry-size-year-age effects No No No No No No Yes
Figure 6: Lowess regression of price on size The log per unit price and the size variable (log (sales)) are demeaned of industry effects.
-.50
.51
0 20 40 60 80 100Size percentiles: Log(total sales)
Log per unit price (Non-exporters) Log per unit price (Exporters)
Country: India
Table 3: Capital labor ratio vs export status All reported figures are coefficients on an exporter dummy which equals one for exporting establishments and is zero for non-exporters. All regressions included industry-year fixed effects, where the industry definition is 4 digit NIC for India and 4 digit ISIC for Chile and Colombia. The number of observations is approximately 27,000 for India, 35,000 for Chile and 75,000 for Colombia. Standard errors are clustered at plant level. + significant at 10%; * significant at 5%; ** significant at 1%. Dependent Variable (1) (2) (3) (4) (5) (6) (7) Panel 1: India (1998) Log (capital/labor) 0.955 0.434 0.322 0.957 0.907 0.343 0.356 [0.043]** [0.042]** [0.042]** [0.044]** [0.044]** [0.043]** [0.044]** Panel 2: Chile (1991-96) Log (capital/labor) 0.836 0.283 0.266 0.621 0.614 0.311 0.375 [0.037]** [0.037]** [0.037]** [0.039]** [0.039]** [0.037]** [0.037]** Panel 3: Colombia (1981-91) Log (capital/labor) 0.745 0.136 0.126 0.639 0.536 0.177 0.209 [0.030]** [0.030]** [0.030]** [0.032]** [0.032]** [0.031]** [0.031]** Other controls: Log (sales) No Yes Yes No No No NoSquare of log (sales) No No Yes No No No NoLog (employment) No No No Yes Yes No NoSquare of log (employment) No No No No Yes No NoIndustry-year effects Yes Yes Yes Yes Yes No NoIndustry-size-year effects No No No No No Yes NoIndustry-size-year-age effects No No No No No No Yes
Figure 7: Lowess regression of capital intensity on size The log(capital/labor) and the size variable (log(total sales)) are demeaned of industry effects
-10
12
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: India
-.50
.51
1.5
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Chile
-1-.5
0.5
11.
5
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Colombia
Dotted line represents exporters, the smooth line represents non-exporters
Log (capital/labor) vs size
Table 4: Skill intensity vs export status All reported figures are coefficients on an exporter dummy which equals one for exporting establishments and is zero for non-exporters. All regressions included industry-year fixed effects, where the industry definition is 4 digit NIC for India and 4 digit ISIC for Chile and Colombia. The number of observations is approximately 27,000 for India, 35,000 for Chile and 75,000 for Colombia. Standard errors are clustered at industry level for India and plant level for Chile and Colombia. + significant at 10%; * significant at 5%; ** significant at 1%. Dependent variable (1) (2) (3) (4) (5) (6) (7) Panel 1: India (1998) Skilled share of employment 0.002 0.026 0.009 0.061 0.046 0.010 0.002 [0.005] [0.005]** [0.005]+ [0.005]** [0.005]** [0.005]+ [0.006] Skilled share of wage bill 0.055 0.026 0.014 0.071 0.064 0.016 0.013 [0.006]** [0.006]** [0.006]* [0.006]** [0.006]** [0.006]* [0.007]* Panel 2: Chile (1991-96) Skilled share of employment 0.013 0.000 -0.006 0.029 0.026 -0.007 -0.007 [0.004]** [0.005] [0.005] [0.005]** [0.005]** [0.005] [0.005] Skilled share of wage bill 0.092 0.019 0.018 0.051 0.050 0.015 0.013 [0.005]** [0.005]** [0.005]** [0.005]** [0.005]** [0.005]** [0.005]* Panel 3: Colombia (1981-91) Skilled share of employment 0.019 0.022 0.015 0.080 0.050 0.015 0.019 [0.004]** [0.004]** [0.004]** [0.005]** [0.004]** [0.004]** [0.005]** Skilled share of wage bill 0.461 0.016 0.015 0.186 0.174 0.049 0.032 [0.013]** [0.009]+ [0.009]+ [0.012]** [0.012]** [0.009]** [0.005]** Other Controls: Log (sales) No Yes Yes No No No No Square of log (sales) No No Yes No No No No Log (employment) No No No Yes Yes No No Square of log (employment) No No No No Yes No No Industry-year effects Yes Yes Yes Yes Yes No No Industry-size-year effects No No No No No Yes No Industry-size-year-age effects No No No No No No Yes
Figure 8a: Lowess regression of skilled share of wage bill on size The skilled share of employment and the size variable (log(total sales)) are demeaned of industry effects.
-.04
-.02
0.0
2.0
4
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: India
-.02
0.0
2.0
4.0
6
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Chile
-.01
0.0
1.0
2.0
3.0
4
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Colombia
Dotted line represents exporters, the smooth line represents non-exporters
Skilled share of employment vs size
Figure 8b: Lowess regression of skilled share of wage bill on size The skilled share of wage bill and the size variable (log(total sales)) are demeaned of industry effects.
-.05
0.0
5.1
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: India
-.1-.0
50
.05
.1.1
5
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Chile
-.1-.0
50
.05
.1
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Colombia
Dotted line represents exporters, the smooth line represents non-exporters
Skilled share of wage bill vs size
Table 5: Worker ability (wage rates) vs export status All reported figures are coefficients on an exporter dummy which equals one for exporting establishments and is zero for non-exporters. All regressions included industry-year fixed effects, where the industry definition is 4 digit NIC for India and 4 digit ISIC for Chile and Colombia. The number of observations is approximately 27,000 for India, 35,000 for Chile and 75,000 for Colombia. Standard errors are clustered at industry level for India and plant level for Chile and Colombia. + significant at 10%; * significant at 5%; ** significant at 1%. Dependent variable (1) (2) (3) (4) (5) (6) (7) Panel 1: India (1998) Log(average wage rate) 0.450 0.129 0.131 0.314 0.316 0.115 0.111 [0.017]** [0.015]** [0.015]** [0.016]** [0.017]** [0.016]** [0.017]** Log(skilled wage rate) 0.515 0.160 0.144 0.332 0.327 0.128 0.147 [0.021]** [0.019]** [0.020]** [0.020]** [0.020]** [0.020]** [0.022]** Log(unskilled wage rate) 0.357 0.109 0.119 0.265 0.264 0.100 0.090 [0.016]** [0.015]** [0.015]** [0.016]** [0.016]** [0.016]** [0.017]** Panel 2: Chile (1991-96) Log(average wage rate) 0.458 0.040 0.050 0.224 0.232 0.068 0.074 [0.015]** [0.012]** [0.011]** [0.014]** [0.014]** [0.012]** [0.012]** Log(skilled wage rate) 0.672 0.095 0.124 0.254 0.269 0.139 0.135 [0.019]** [0.016]** [0.015]** [0.018]** [0.017]** [0.015]** [0.016]** Log(unskilled wage rate) 0.315 0.019 0.025 0.164 0.168 0.039 0.051 [0.013]** [0.012] [0.012]* [0.014]** [0.014]** [0.012]** [0.012]** Panel 3: Colombia (1981-91) Log(average wage rate) 0.461 0.016 0.015 0.186 0.174 0.049 0.043 [0.013]** [0.009]+ [0.009]+ [0.012]** [0.012]** [0.009]** [0.010]** Log(skilled wage rate) 0.657 0.033 0.056 0.213 0.220 0.091 0.081 [0.016]** [0.012]** [0.012]** [0.014]** [0.014]** [0.012]** [0.012]** Log(unskilled wage rate) 0.319 0.001 -0.013 0.117 0.098 0.017 0.013 [0.012]** [0.009] [0.008] [0.010]** [0.010]** [0.009]* [0.009] Other Controls: Log (sales) No Yes Yes No No No NoSquare of log (sales) No No Yes No No No NoLog (employment) No No No Yes Yes No NoSquare of log (employment) No No No No Yes No NoIndustry-year effects Yes Yes Yes Yes Yes No NoIndustry-size-year effects No No No No No Yes NoIndustry-size-year-age effects No No No No No No Yes
Figure 9: Lowess regression of average wage rate on size The log average wage rate and the size variable (log(total sales)) are demeaned of industry effects.
-.50
.51
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: India
-.50
.51
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Chile
-.50
.51
0 20 40 60 80 100Size percentiles: Log(total sales)
Country: Colombia
Dotted line represents exporters, the smooth line represents non-exporters
Log (average wage rate) vs size
Table 6: Other quality proxies, India (1998) All reported figures are coefficients on an exporter dummy which equals one for exporting establishments and is zero for non-exporters. The number of observations is approximately 27000. Standard errors are clustered at industry level. + significant at 10%; * significant at 5%; ** significant at 1%. Dependent variable (1) (2) (3) (4) (5) (6) (7) ISO 9000 certification dummy 0.147 0.116 0.099 0.125 0.118 0.095 0.097 [0.008]** [0.008]** [0.008]** [0.008]** [0.008]** [0.008]** [0.009]** Computer use dummy 0.350 0.175 0.144 0.246 0.237 0.131 0.124 [0.013]** [0.013]** [0.013]** [0.013]** [0.013]** [0.014]** [0.015]** Robotics use dummy 0.065 0.05 0.042 0.056 0.054 0.041 0.041 [0.005]** [0.005]** [0.005]** [0.005]** [0.005]** [0.005]** [0.006]** Computer network dummy 0.061 0.047 0.038 0.053 0.052 0.040 0.041 [0.006]** [0.006]** [0.006]** [0.006]** [0.006]** [0.006]** [0.006]** Internet use dummy 0.146 0.120 0.108 0.130 0.127 0.099 0.091 [0.009]** [0.009]** [0.009]** [0.009]** [0.009]** [0.009]** [0.009]** Survey on floppy dummy 0.125 0.084 0.068 0.099 0.095 0.066 0.063 [0.009]** [0.009]** [0.010]** [0.009]** [0.009]** [0.010]** [0.011]** Other Controls: Log(sales) No Yes Yes No No No No Square of log(sales) No No Yes No No No NoLog (employment) No No No Yes Yes No NoSquare of log(employment) No No No No Yes No No Industry-year effects Yes Yes Yes Yes Yes No NoIndustry-size-year effects No No No No No Yes NoIndustry-size-year-age effects No No No No No No Yes
Table 7: Capital and skill intensity premium for exporters, US (from Bernard and Jensen, 1999, Table 2) This table is based on Table 2 in Bernard and Jensen (1999). All reported figures are coefficients on an exporter dummy which equals one for exporting establishments and is zero for non-exporters. All regressions included industry-year fixed effects, where the industry definition is 4 digit SIC. Column 1 includes only industry and state fixed effects. Column 2 includes log (employment). The number of observations is 56247 in 1984, 199258 in 1987 and 224009 in 1992. All effects are significant at 1% level. 1984 1987 1992 Dependent variable (1) (2) (1) (2) (1) (2) Log(capital/labor) 0.190 0.175 0.128 0.101 0.119 0.093 Log(average wage rate) 0.179 0.148 0.112 0.093 0.202 0.136 Log(skilled wage rate) 0.188 0.160 0.092 0.072 0.090 0.066 Log(unskilled wage rate) 0.088 0.036 0.099 0.052 0.114 0.046 Industry fixed effects Yes Yes Yes Yes Yes Yes State fixed effects Yes Yes Yes Yes Yes Yes Log(employment) No Yes No Yes No Yes
Appendix Appendix Table 1: Within establishment changes -- Fixed effects regressions All reported figures are coefficients on an exporter dummy which equals one for exporting establishments and is zero for non-exporters. All regressions included establishment fixed effects. The number of observations is approximately 35,000 for Chile and 75,000 for Colombia. Standard errors are clustered at plant level. + significant at 10%; * significant at 5%; ** significant at 1%. Dependent variable (1) (2) (3) (4) (5) (6) (7) Panel 1: Chile (1991-96) Capital intensity Log (capital/labor) 0.001 0.020 0.010 0.073 0.073 -0.029 0.000 [0.025] [0.025] [0.025] [0.024]** [0.024]** [0.025] [0.023] Skill intensity Skilled share of employment 0.002 0.005 0.004 0.008 0.008 0.001 0.000 [0.004] [0.004] [0.004] [0.004]* [0.004]* [0.004] [0.003] Skilled share of wage bill 0.006 0.007 0.008 0.010 0.010 0.000 -0.001 [0.004] [0.004]+ [0.004]+ [0.004]* [0.004]* [0.004] [0.004] Worker ability Log(average wage rate) 0.021 -0.003 0.005 0.038 0.038 -0.007 -0.004 [0.008]* [0.008] [0.008] [0.008]** [0.008]** [0.008] [0.007] Log(skilled wage rate) 0.037 0.007 0.019 0.040 0.040 -0.002 0.003 [0.012]** [0.012] [0.012] [0.012]** [0.012]** [0.012] [0.011] Log(unskilled wage rate) 0.022 0.000 0.007 0.035 0.035 0.001 0.003 [0.010]* [0.010] [0.010] [0.010]** [0.010]** [0.009] [0.009] Panel 2: Colombia (1981-91) Capital intensity Log (capital/labor) 0.112 0.089 0.061 0.170 0.159 -0.017 -0.003 [0.022]** [0.022]** [0.021]** [0.020]** [0.020]** [0.021] [0.019] Skill intensity Skilled share of employment 0.010 0.014 0.012 0.023 0.017 0.003 0.005 [0.003]** [0.003]** [0.003]** [0.003]** [0.003]** [0.003] [0.003] Skilled share of wage bill 0.009 0.007 0.007 0.017 0.014 -0.004 -0.001 [0.003]** [0.003]* [0.003]* [0.003]** [0.003]** [0.003] [0.003] Worker ability Log(average wage rate) 0.039 -0.001 0.000 0.050 0.053 0.000 0.002 [0.006]** [0.006] [0.006] [0.006]** [0.007]** [0.006] [0.006] Log(skilled wage rate) 0.031 -0.024 -0.015 0.026 0.028 -0.021 -0.014 [0.008]** [0.008]** [0.008]+ [0.008]** [0.009]** [0.008]** [0.007]+ Log(unskilled wage rate) 0.036 0.005 0.006 0.049 0.050 0.012 0.011 [0.006]** [0.006] [0.006] [0.006]** [0.006]** [0.006]+ [0.006]+ Other Controls: Establishment fixed effects Yes Yes Yes Yes Yes Yes No Log (sales) No Yes Yes No No No No Square of log (sales) No No Yes No No No No Log (employment) No No No Yes Yes No No Square of log (employment) No No No No Yes No No Industry-size-year effects No No No No No Yes No Industry-size-year-age effects No No No No No No Yes
Appendix Figure 1a: Switching to exporter status – Chile (1991-96) The figure plots coefficients on a dummy for each of the 4 years before and after switching from exporter to non-exporter status and the year of exporting. All variables are demeaned of 4 digit industry-year-size decile-age quintile fixed effects. We include data only for exporting establishments that switch from non-exporting to exporting status once and stays as exporter after the switch. Data on exporting establishments for truncated to the 4 years before and after exporting (data on the year of exporting is also included). The omitted category (relative to which the coefficients on the index years are estimated) is the group of establishments that never export at any time during the panel period. (Thus data on establishments that switch in and out of the export market is excluded in this analysis.)
-0.2
00.
000.
200.
400.
600.
80
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log (capital/labor)
-0.0
4-0
.02
0.00
0.02
0.04
0.06
0.08
0.10
-4 -3 -2 -1 0 1 2 3 4
Years from switch
Mean 5% bound
95% bound
Skilled share of wage bill
-0.0
8-0.
06-0
.04
-0.0
20.
000.
020.
040.
060.
08
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Skilled share of employment
-0.0
50.
000.
050.
100.
150.
200.
25
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log(average wage rate)
-0.0
50.
000.
050.
100.
150.
200.
250.
30
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log(skilled wage rate)
-0.1
5-0
.10
-0.0
50.
000.
050.
100.
150.
20
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log(unskilled wage rate)
All regressions include controls for industry-size-year-age effects
Appendix Figure 1b: Switching to exporter status – Colombia (1981-91) The figure plots coefficients on a dummy for each of the 4 years before and after switching from exporter to non-exporter status and the year of exporting. All variables are demeaned of 4 digit industry-year-size decile-age quintile fixed effects. We include data only for exporting establishments that switch from non-exporting to exporting status once and stays as exporter after the switch. Data on exporting establishments for truncated to the 4 years before and after exporting (data on the year of exporting is also included). The omitted category (relative to which the coefficients on the index years are estimated) is the group of establishments that never export at any time during the panel period. (Thus data on establishments that switch in and out of the export market is excluded in this analysis.)
-0.1
00.
000.
100.
200.
300.
400.
500.
600.
70
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log (capital/labor)
-0.0
20.
000.
020.
040.
060.
080.
100.
12
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Skilled share of wage bill
-0.0
20.
000.
020.
040.
060.
080.
100.
12
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Skilled share of employment
-0.1
5-0
.10
-0.0
50.
000.
050.
100.
15
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log(average wage rate)
-0.1
0-0
.05
0.00
0.05
0.10
0.15
0.20
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log(skilled wage rate)
-0.2
0-0
.15
-0.1
0-0
.05
0.00
0.05
0.10
0.15
-4 -3 -2 -1 0 1 2 3 4Years from switch
Mean 5% bound
95% bound
Log(unskilled wage rate)
All regressions include controls for industry-size-year-age effects