ESTIMATING CROSS-COUNTRY

DIFFERENCES IN PRODUCT QUALITY

Juan Carlos Hallak and Peter K. Schott

June 2, 2010

We develop a method for decomposing countries’ observed export prices into quality

versus quality-adjusted components using information contained in their trade bal-

ances. Holding observed export prices constant, countries with trade surpluses are

inferred to o¤er higher quality than countries running trade de…cits. Our method

accounts for variation in trade balances induced by both horizontal and vertical dif-

ferentiation, and we use it to estimate the evolution of manufacturing quality for the

world’s top exporters from 1989 to 2003. We …nd that observed unit value ratios can

be a poor approximation for relative quality di¤erences, that countries’quality is con-

verging more rapidly than their income, and that countries appear to vary in terms of

displaying “high-quality” versus “low-price” growth strategies.

Special thanks to Alan Deardor¤ for many fruitful discussions. We also thank Steve Berry,
Keith Chen, Rob Feenstra, Cecilia Fieler, James Harrigan, Justin McCrary, Phil Haile, Beata
Javornik, Amit Khandelwal, Keith Maskus, Peter Neary, Serena Ng, Ben Polak, Marshall
Reinsdor¤, Matthew Shapiro, Walter Sosa Escudero, Alejandro Vicondoa, participants at
various seminars and four anonymous referees for many helpful comments. Alejandro Molnar
and Santiago Sautua provided superb research assistance. This research is supported by the
National Science Foundation under Grants No. 0241474 and 0550190. Any opinions, …ndings,
and conclusions or recommendations expressed in this material are those of the authors and
do not necessarily re‡ect the views of the National Science Foundation.
 I. Introduction

Theoretical and empirical research increasingly point to the importance of

product quality in international trade and economic development.1 Unfortu-

nately, relatively little is known about how countries’ product quality varies

across time, or how it is in‡uenced by trade liberalization and other aspects

of globalization. A major impediment to research in this area is lack of data

– reliable estimates of product quality for a wide range of countries, indus-

tries and years do not exist. In this paper, we introduce a method for obtaining

such estimates that incorporates information about world demand for countries’

products.

Researchers often react to the absence of information about countries’prod-

uct quality by constructing ad hoc proxies, the most common of which is ob-

served export prices (unit values).2 This measure is unsatisfactory, however,

because export prices may vary for reasons other than quality. Chinese shirts

might be cheaper than Italian shirts in the U.S. market because of lower quality,

but they might also sell at a discount because China has lower production costs

or an undervalued exchange rate. If consumers value variety and goods are hor-

izontally as well as vertically di¤erentiated, high-cost exporters can survive in

the U.S. market even in the face of cost disadvantages.

Our method for identifying countries’product quality involves decomposing

1. Flam and Helpman (1987) is representative of a line of theoretical research studying the
in‡uence of product quality on international trade. Empirically, cross-country and time-series
variation in product quality has been linked to …rms’export success (Brooks 2006, Verhoogen
2008), countries’ skill premia (Verhoogen 2008), quantitative import restrictions (Aw and
Roberts 1986, Feenstra 1988) and trade patterns (Schott 2004, Hallak 2006). The contribution
of quality growth to macroeconomic growth is investigated theoretically by Grossman and
Helpman (1991) and empirically by Hummels and Klenow (2005).
2. See, for example, Schott (2008). More generally, unit value di¤erences …gure promi-
nently in surveys of countries’ “quality competitiveness” (e.g., Aiginger 1998, Verma 2002,
Ianchovichina et al. 2003, and Fabrizio et al. 2007) and also are often used to distinguish
horizontal from vertical intra-industry trade ‡ows (e.g., Abed-el-Rahman 1991 and Aiginger
1997).

1
observed export prices into quality versus quality-adjusted-price components.

We de…ne quality to be any tangible or intangible attribute of a good that

increases all consumers’ valuation of it. Countries’ product quality relative to

a numeraire country is identi…ed by combining data on their observed export

prices with information about global demand for their products contained in

their trade balance vis a vis the world. The intuition behind our identi…cation

is straightforward and has been used extensively in the industrial organization

literature: because consumers are assumed to care about price relative to quality

in choosing among products, two countries with the same export prices but

di¤erent global trade balances must have products with di¤erent levels of quality.

Among countries with identical export prices, the country with the higher trade

balance is revealed to possess higher product quality.3

A major contribution of the paper is to generalize this intuition to a setting

where countries also are allowed to di¤er in the number of unobserved horizon-

tal varieties they export in each product category (e.g., red versus blue men’s

wool sweaters). Horizontal di¤erentiation is a standard aspect of recent trade

models, and allowing for it helps explain why many products are exported by a

wide range of countries. Incorporating it here is di¢ cult because it introduces

an additional factor besides quality that can increase consumer demand for a

country’s products. All else equal, consumer love of variety implies that coun-

tries producing a larger number of varieties in a product category export larger

quantities and therefore exhibit higher trade surpluses. Unless the number of

horizontal varieties that countries export is accounted for, this increase in net

trade will be interpreted, erroneously, as higher product quality. Our approach

assumes a negative relationship between quality-adjusted prices and the number

3. The use of market shares to infer unobserved consumer valuation is well-established

in the industrial organization and index number literatures (e.g., Berry 1994 and Bils 2004,
respectively). Here, countries’net trade with the rest of the world (conditional on trade costs)
is a natural expression of their “market share”.

2
of varieties countries export. We justify this assumption by appealing to theo-

retical …ndings in Romalis (2004) and Bernard et al. (2007) that demonstrate

that countries’comparative advantage sectors exhibit both relatively low prices

–due to relatively low factor costs –and a relatively high number of varieties –

due to disproportionate use of factor inputs.4

Using countries’ net trade with the rest of the world to identify consumer

demand imposes an important practical constraint on empirical implementation

of our method. Currently, the most reliable time-series data on countries’trade

balances are recorded according to comparatively coarse industries relative to

the much more disaggregated products (e.g., men’s wool sweaters) at which

some countries’export prices can be observed. To deal with this constraint we

derive a theoretically appropriate price index that aggregates countries’observed

product-level export prices up to the industry level. We refer to this index as the

“Impure Price Index”because it is based on prices that are “contaminated”by

quality. Our index has the useful property of being separable into quality versus

quality-adjusted-price components, but it is developed under the potentially

strong assumption that countries’ quality is constant across products within

industries. Thus, we are faced with an “aggregation trade-o¤”: while product

quality is more likely to be constant across products the more disaggregated

the industry, data on countries’global net trade becomes more scarce as well as

more susceptible to measurement error. Use of disaggregated industries may also

be problematic if countries’use of intermediate inputs straddles the industry at

which quality is being estimated; in this case, reported net trade in that industry

fails to account for all of its inputs. In a pilot examination of this issue in the

Data Appendix below, we …nd that apparel quality can be over-estimated for

4. Feenstra (1994) outlines a method for computing import price indexes that accounts for
the introduction of new product varieties. (See also Broda and Weinstein 2004). Given its
focus on changes in prices over time, that method requires no knowledge of cross-sectional
variation in the number of varieties countries export within product categories so long as that
number is constant over time for a subset of countries.

3
countries that import textiles to produce apparel.

Even though the Impure Price Index comparing two countries’export prices

is unobservable, we show that it is bounded by observable Paasche and Laspeyres

indexes de…ned over their common exports to a third country (i.e., the United

States). This result anchors a two-stage strategy for inferring countries’product

quality. In the …rst stage, we use the large set of bilateral Paasche and Laspeyres

bounds (e.g., Germany versus China, Switzerland versus Germany, France ver-

sus Thailand, etc.) to estimate an Impure Price Index for each country-industry-

year relative to a common numeraire. In the second stage, we use data on coun-

tries’global net trade in the industry to strip away variation in quality-adjusted

(or “pure”) prices from the estimated Impure Price Indexes. This procedure

yields estimates of quality that vary by country, industry and year.

We use our method to estimate manufacturing quality for the world’s 43

largest exporters over the period 1989 to 2003. The estimated Quality Indexes

reveal substantial variation in quality levels across countries in any given year

as well as across years. We …nd that relative quality for overall manufactur-

ing increases most dramatically for Ireland, Malaysia and Singapore over the

sample period, and falls most dramatically for Hong Kong and New Zealand.

Among countries that begin the sample period in the top tercile of quality, Aus-

tralia and Japan experience the largest relative declines. We also show that

our estimates of product quality and their evolution over time can deviate sub-

stantially from estimates of quality based on raw export prices. Indeed, changes

in estimated relative quality and raw export prices move in opposite directions

for one-third of the countries in our sample, including some of those with the

largest increases in our quality estimates. We also …nd greater narrowing in

estimated quality di¤erences than per capita GDP di¤erences over our sample

period. An interesting question for further research is the extent to which this

4
quality convergence reveals a catching up in terms of technological knowledge

by developing countries versus greater use of high-quality intermediate inputs

from developed economies.

This paper’s focus on cross-sectional variation in product quality di¤erenti-

ates it from a very large index number literature devoted to constructing quality-

adjusted cost-of-living indexes. Here, rather than measure quality changes in

bundles of products purchased over time, we identify quality variation over

simultaneously purchased bundles from di¤erent sources of supply. Since we

cannot observe products’underlying attributes, we are also unable to make use

of standard strategies –such as hedonic pricing –that link product attributes to

speci…c dimensions of quality.5 Our method complements such e¤orts, however,

because its use of publicly available trade data permits estimation of product

quality across a broad range of countries, industries and years for which sur-

veys of product characteristics may be unavailable or prohibitively expensive to

collect.6

Our analysis is more closely related to previous attempts in the international

trade literature to deal with potential variation in unit values not entirely due to

variation in product quality. Hallak (2006), for example, assumes a monotonic

relationship between per-capita income and “pure prices” at the sector level

while, in the closest precedent to this paper, Hummels and Klenow (2005) use

import prices and quantities to make inferences about the cross-sectional elastic-

ity of quality with respect to country income and size. Neither of these papers,

however, permits explicit estimation of product quality by country, sector, and

year, as is done in this paper.7 Our approach is also di¤erent from an earlier

5. Feenstra (1995), for example, demonstrates how information on product attributes can
be used to establish bounds on the exact hedonic price index.
6. The International Price Program of the U.S. Bureau of Labor Statistics constructs im-
port and export price indexes by combining survey data on …rms’prices with …rms’assessments
about changes in the quality of their products over time (Alterman et al. 1999).
7. More recently, Khandelwal (2008) has developed a method for estimating quality based
on the assumption of a nested logit demand system.

5
strand of literature primarily interested in analyzing the e¤ect of import quotas

on the quality composition of trade (e.g., Aw and Roberts 1986, Boorstein and

Feenstra 1987, and Feenstra 1988). In that literature, import quality increases

when the composition of imports shifts toward high-quality product categories.

Here, we take a within- rather than across-product view of quality variation.

Our results also relate well to recent e¤orts by Rodrik (2006), Hausmann,

Hwang and Rodrik (2007) and others to estimate the extent to which the export

quality of developing countries like China is equal to that of the world’s most

developed economies. Like Schott (2008) and Xu (2007), we …nd Chinese quality

to be relatively low compared to developed countries across all years of our

sample.

The paper is structured as follows. Section II outlines our assumptions about

consumer demand and introduces the Impure and Pure Price indexes that will

be the focus of our analysis. Section III shows that the unobservable Impure

Price Index is bounded by observable Paasche and Laspeyres indexes. Section

IV derives the relationship between the Pure Price Index and countries’sectoral

net trade. Sections VI through VII describe the application of our method to

identifying export quality trends for 43 large trading countries over the period

1989 to 2003. Section VIII concludes. Two appendixes attached to this paper

provide proofs of our main propositions and an examination of quality by manu-

facturing industry. A web-based technical appendix contains estimation details

and additional results.

II. Preferences and Price Indexes

This section describes the preference structure underlying our analysis and

formally introduces the price and quality indexes that are the focus of our

method.

6
II.A. Preferences

Varieties of goods are classi…ed into product categories (“products”for short),

which are in turn classi…ed into sectors. Sectors are indexed by subscript s =

1; :::; S, while products (within sector s) are indexed by subscript z = 1; :::; Zs .

Product categories are the level of aggregation at which prices are observed

while sectors are the level of aggregation at which countries’trade balances are

observed and hence quality is estimated. In our empirical investigation below,

products correspond to ten-digit U.S. Harmonized System (HS) categories while

sectors are de…ned alternatively as All Manufacturing, one-digit SITC man-

ufacturing industries or select two-digit SITC manufacturing industries. The

theoretical framework presented here focuses on sector s.

There are K countries, indexed by superscript k. Preferences are represented

by a two-tier utility function that incorporates consumer love of variety.8 The

upper tier is Cobb-Douglas while the lower tier is CES,

S
" Zs
K X
# s
Y X s 1 s 1
k k s
(1) U= ubss ; us = z s xz nkz ; s > 1;
s=1 k=1 z=1

where nkz is the number of horizontally di¤erentiated varieties of product z

produced by country k, xkz is the quantity consumed per variety, and s is the

elasticity of substitution between varieties. For compactness, we omit subindex-

ing z by s in the second summation of equation (1) and throughout the paper.

We note that by indexing products instead of varieties, we implicitly assume

symmetry across varieties of the same product.

k
The utility function includes two shifters, z and s. The …rst shifter, z,

varies across products but is constant across countries for a particular product.

It captures consumers’valuation of the essential characteristics common to the

8. Homothetic preferences, although standard in the international trade literature, are

potentially strong in this context as countries’demand for quality may vary with income.

7
heterogeneous varieties of a product. Consumers, for example, might have a
k
higher preference for varieties of tables than chairs. The second shifter, s,

varies across countries and sectors, but is constant across products within a

particular country and sector. It represents “quality”, which we de…ne as any

attribute of a good (other than price and those already captured by z) for which

all consumers are willing to pay more, and includes tangibles (e.g., durability) as

well as intangibles (e.g., product image due to advertising). These assumptions,

implicit in (1), are formalized as:

k
Assumption 1: z = z; 8k = 1; :::; K:

k k
Assumption 2: z = s; 8z = 1; :::; Zs :

The preference structure de…ned by equation (1) implies that product de-

mand depends on quality-adjusted or “pure” prices. Letting pkz be the export

price of a typical variety of product z produced in country k, we de…ne the

k
“pure” price of that variety by pekz = pkz =( z s ). The pure price is a quality-

adjusted price. It is also divided here by z for notational compactness, but

none of the results is a¤ected by this choice.

II.B. Price and Quality Indexes

In this section we introduce the price and quality indexes that are the focus

of our analysis. First de…ne an aggregator of observed product prices produced

8
in country k and sector s as9

" #1 1
X 1
s
1 X nk
1
Psk pkz Pz
s
nz z
s
; nz = 1 :
z
K
k Zs nkz
z

We can then de…ne the Impure Price Index (IPI) between countries k and k 0 as

0 0
(2) Pskk = Psk =Psk :

Psk is a weighted average of country k’s observed prices across products z in

sector s, where each z is weighted according to the “world average” number of

1
varieties (nz ) and the demand shifter ( z
s
) for that product.

The Impure Price Index is a summary measure of price variation between

goods produced by countries k and k 0 in sector s. It has three features worth

noting. First, because it is de…ned over observed prices it is “impure” in the

sense that its prices are “contaminated” by quality. Second, it is transitive:

0
choosing an arbitrary country, o, as numeraire, Pskk can always be recovered
0
from the ratio Psko =Psk o . Finally, though unobservable due to its inclusion

of unobserved variables such as the number of varieties countries export, this

index can be estimated. In the next section, we show that the unobservable

IPI is bounded by observable price indexes while in Section 5.1 we show how

those bounds can be used to estimate the IPI. An alternate index based on

nkz (rather than on nz ) would have the advantage of being a subaggregate of

the exact consumer price index and a more accurate predictor of countries’net

trade. However, as will become clear later, the fact that this alternate index

does not use weights that are common to all countries implies that it cannot be

9. To simplify notation, unless otherwise noted the subindexes under the summation sign
range over all elements of the relevant set, e.g., z = 1; :::; Zs and k = 1; :::K.

9
bounded by observable price indexes.
kk0 k k0
We de…ne a Quality Index, s = s= s , as the ratio of two countries’

quality levels in sector s, and de…ne a Pure Price Index (PPI), Peskk = Pesk =Pesk ,
0 0

1
P 1
1 s
as the ratio of pure price aggregators, Pesk nz pekz
s
. The Impure
z
Price Index can be decomposed into the Quality Index and the Pure Price Index:

0
kk0 e kk0
(3) Pskk = s Ps :

Estimating ko
s is our main objective. Although both ko
s and Pesko are un-

observable, we show in Section 5.2 how they can be identi…ed from estimates of

Psko and information on countries’net trade with the world in sector s.

III. Bounding the “Impure” Price Index

In this section we show that the unobserved Impure Price Index introduced

above is bounded by observable Paasche and Laspeyres indexes de…ned over the

prices (unit values) of country pairs’exports to a third country. This result is

the basis of the strategy for estimating the IPI as outlined in Section 5.1. Our

bounding of the Impure Price Index proceeds in two steps. First, we show that

observable Paasche and Laspeyres indexes bound unobserved “cost-of-utility”

indexes. Second, we show that these unobserved cost-of-utility indexes bound

the unobserved Impure Price Index.

III.A. Paasche and Laspeyres Bounds on Cost-of-Utility

Indexes

We de…ne unobserved cost-of-utility indexes and use revealed preference

to show that they are bounded by observed Paasche and Laspeyres indexes.

Though the bounding of cost-of-utility indexes by Paasche and Laspeyres in-

10
dexes is standard in the index number literature, our setup involves two compli-

cations. First, rather than concentrating on expenditures over the universe of

goods in two di¤erent time periods, we focus on contemporaneous expenditures

over subsets of the universe of goods purchased from a pair of exporting coun-

tries. Second, because we allow for horizontal di¤erentiation, our cost-of-utility

indexes need to deal with the number of varieties countries export –which need

not be the same in the two countries.

We focus on countries’ exports to a single “common importer”, which we

refer to as the United States given the data used in our empirical implementa-

tion. We note that the analysis would be identical were it to be applied to any

other common importer, or to a set of importers. For ease of exposition, we as-

sume in this section that all countries are “active”in (i.e., export to the United

States) the same set of products, deferring discussion of the more general case

of imperfect overlap to the Theory Appendix at the end of this paper. We sum-

marize the implications of imperfect overlap for Proposition 1 after introducing

the proposition below, and discuss the potential impact of imperfect overlap on

our empirical analysis in Section 5.

De…ne vectors pks and qsk to include, respectively, U.S. import prices and

quantities for all products in sector s coming from country k. Stack these

vectors across countries to form ps and qs . Stack the latter vectors across

sectors to form p and q. Analogously, de…ne vectors n, , and . A vector

of per-variety consumption x is implicitly de…ned by q and n. Finally, de…ne

qs k as the complement of qks with respect to q. Vector qs k includes import

quantities in sector s from all countries other than k, and also import quantities

in all other sectors from all countries (including k).

For country k of country pair kk 0 , we de…ne the constrained expenditure (or

11
 00
import) function ms;k (pks ; qs k ; n; ; ;u) as the solution to the problem

00
(4) min pks q
bks s:t: qks ; qs k ; n; ; ) = u;
U (b k 00 = 1; :::; K
bs
q k

where U is the representative consumer utility function.10 This function repre-

sents the minimum expenditure on varieties in sector s imported from country

k that the consumer would be required to make in order to attain utility level u
00
if import prices of those varieties were pks (rather than pks ), holding constant

the actual values of qs k ; n; ; .

To obtain an explicit functional form for ms;k , we use the preferences outlined
s
s 1
k
P k k k s
s 1
in equation (1). De…ne us nz z s xz only over varieties
z
exported by country k in sector s. The separability of the utility function in

(1) implies that U can be written as a function of uks and a function of arguments

held constant in problem (4), us k . Since U is strictly increasing in uks , there is

bks ; us k = u. Then, problem

bks , such that U u
a single value of this variable, u
P k k00 k
(4) reduces to choosing the per-variety quantities xkz that minimize nz p z xz
z
subject to uks = u
bks . The solution to this problem is the product of a CES

bks 11
aggregator measuring the unit cost of utility and the target level of utility, u

2 !1 31 1
s
X k00 s
00 00
(5) ms;k (pks ; qs k ; ; ;u) = 4 nkz pekz s
k
5 bks :
u
z s

By revealed preference, ms;k (pks ; qs k ; n; ; ;u) = pks qks . However, if prices

0
were pks instead of pks , the minimum import expenditure would be equal to
0 0
or lower than pks qks , because the amount pks qks is su¢ cient to attain utility u
0 0
but qks is not necessarily optimal given pks . Hence, ms;k (pks ; qs k ; n; ; ;u)

10. Neary and Roberts (1980) and Anderson and Neary (1992) use the constrained expen-
diture function to analyze consumption choices under rationing.
11. It is here where Assumptions 1 and 2 are critical. In equation (5) we use these assump-
00 00 k00 k00 k00
pk pk 00
tions to derive z
k k = z
k00 k00
z
k
z
k = pekz s
k .
z z z z z z s

12
 0
pks qks . Using these results, we obtain

kk 0 ms;k (pks ; qs k ; n; ; ;u) pks qks 0

(6) Ms;k = Hskk :
ms;k (pks 0 ; qs k ; n; ; ;u) pks 0 qks

Inequality (6) displays a standard result in index number theory stating that
0 0
kk
the cost-of-utility price index Ms;k is larger than a Paasche price index, Hskk ,

de…ned over the observed prices of the country pair’s exports to the U.S. in sector

s. We note that the Paasche index is de…ned here in a cross-sectional rather than
0
kk
a time-series context. Ms;k captures the change in minimum expenditure on

country k’s varieties (in sector s) that would be necessary to maintain utility u
0
if import prices of those varieties changed from pks to pks , holding constant their

number and characteristics (including quality), and the number, characteristics

and quantity consumed of all other goods.

We can combine equation (5) with inequality (6) to obtain

(7)
2 31 1
P ek
p
1 s s

6 z nkz e
z
Psk 7
0 0 0
kk0 kk0 6 7
ln Hskk kk
ln Ms;k = ln Pskk + ln s;k ; s;k 4P 0 1 s 5 :
ek
p
nkz z
e k0
P
z s

In a similar manner, we can focus alternatively on imports from country k 0

to obtain

0 0
kk0 ms;k0 (pks ; qs k ; n; ; ;U ) pks qks 0
(8) Ms;k 0 0 = Lkk
s ;
ms;k0 (pks 0 ; qs k ; n; ; ;U ) pks 0 qks 0

0
where Lkk
s is a Laspeyres price index. This is another standard result, which
0
kk
states that the cost-of-utility index Ms;k 0 is bounded from above by a Laspeyres

13
price index. Using the explicit functional form for ms;k0 , we obtain

(9)
2 31 1
P 0 ek
p
1 s s

6 z nkz ek
P
z
7
0 0 0
kk0 kk0 6 s
7
ln Lkk
s
kk
ln Ms;k kk
0 = ln Ps + ln s;k0 ; s;k0 4P 0 1 s 5 :
ek
p
nkz 0 z
e k0
P
z s

Equations (7) and (9) relate the implications of consumer cost minimization

to cross-sectional Paasche and Laspeyres price indexes, where each of the cost-of

utility indexes has observable bounds on one side.12 Although a standard result

in the index number literature shows that the cost-of-utility index for a consumer

with homothetic preferences is independent of the utility level – and bounded

both above and below – our allowance for horizontal di¤erentiation yields two
0 0
kk kk
cost-of-utility indexes because Ms;k and Ms;k 0 are de…ned over di¤erent numbers

0 0 0
of varieties, i.e., nkz and nkz , respectively. Ms;k
kk kk
and Ms;k 0 would be equal if, for

example, the number of varieties in countries k and k 0 were proportional to one

another for every product category.13

III.B. Paasche and Laspeyres Bounds on the Impure Price

Index

To bound the (unobservable) Impure Price Index by the observable Paasche

and Laspeyres indexes via the cost-of-utility indexes de…ned above, we must

12. Note that all prices (observed and pure) in this section are cif import prices, that is,
import prices inclusive of customs, insurance and freight charges. Under the assumption that
trade costs are constant across product categories within a sector (see Section 4), inequalities
kk0 ; M kk0 ; H kk0 ; Lkk0 are alternatively de…ned using free-on-board
(7) and (9) also hold if Ms;c s;d s s
(FOB) prices – i.e., exclusive of customs, insurance and freight charges – as all terms are
simply scaled by the relative trade costs between countries k and k0 and the United States.
As noted in Section 5, we use fob import unit values to measure U.S. trading partners’export
prices in our empirical analysis.
0 0
13. Note also that the indexes Hskk ; Lkk kk0 kk0
s ; Ms;k and Ms;k0 all weight prices in the numerator
0 0
and in the denominator with the same weights, respectively qks ; qks ; nkz ; and nkz . Our ability
0 0
to bound Pskk with those indexes in the next section depends crucially on Pskk also having
weights, nz , that are common in the numerator and denominator.

14
 kk0 kk0 0 0 0 0
show that ln s;k 0 and ln s;k0 0 so that Hskk kk
Ms;k Pskk kk
Ms;k 0

0
Lkk
s . In this section, we outline assumptions that are su¢ cient for these condi-

tions to hold.

Our …rst step is to decompose the number of varieties countries produce into
P k
three meaningful parts. Let nks = Z1s nz be country k’s average number of
z
0
0
1 nk nk
varieties across product categories in sector s. Let nkk
z = 2
z
nk
+ z
nk
0 be
s s

the (normalized) average number of varieties of product z in sector s across

members of the country pair. Then, the (normalized) number of varieties a

country produces can be expressed as the sum of three terms:

nkz 0 0
(10) ekk
= nz + n z +nek;kk
z :
nks

The …rst term is the world average for product z introduced in Section 2.14 The

second term is the “country-pair excess variety” in product z relative to the

0 0
ekk
world average, n z = nkk
z nz , which captures the extent to which the average

number of varieties in country pair kk 0 is above or below the world average.

The third term is country k’s “bilateral excess variety” for product z relative
0 nk 0 P kk0 P k;kk0
to kk 0 ’s average, n
ek;kk
z = nzk nkkz . We note that ez = 0,
n ez
n =0
s
z z
P k0 ;kk0
and nez = 0: that is, the pair kk 0 cannot have positive country-pair excess
z
variety in all z and neither country can have positive bilateral excess variety in
P 00 0
all z. Finally, ekz ;kk = 0: k and k 0 cannot both have positive bilateral
n
k00 =k;k0
excess variety in the same z.

Our second step is to de…ne the (normalized) bilateral di¤erence in countries’

14. On notation: recall that implicit in our use of the index z is the understanding that it
pertains to the z within sector s. Thus, all terms in equation (10) refer to a particular sector
s.

15
pure prices in product z as

!1 s 0
!1 s
0 pekz pekz 0 0
(11) pekk
z = ; ( pekz k = pekk
z ) :
Pesk Pesk0

0
A positive pekk
z indicates that country c has a lower pure price of z (relative

to the pure price aggregator) than country k 0 . A lower pure price may arise,

for example, due to comparative advantage, i.e., variation in exporters’relative

production e¢ ciency or factor costs.

Assumption 3 states that country k relative to country k 0 will tend to have

positive bilateral excess variety in those products in which it has a lower relative

pure price (the operator covs denotes sample covariances de…ned over all z in

sector s).

0 0 0 0 0
ek;kk
Assumption 3: covs n z ; pekk
z ekz ;kk ; pekz k
= covs n 0

This assumption is motivated by theoretical models of international trade

with product di¤erentiation that allow for trade costs and do not assume factor

price equalization (e.g., Romalis 2004, Bernard et al. 2007). These models …nd

that, across goods, the relative number of varieties between two countries is a

negative function of the countries’ relative prices. This …nding supports the

intuitive notion that countries should have a relatively higher (lower) number of

…rms in sectors or products in which they are relatively more (less) competitive,

i.e. those sectors with relatively lower (higher) prices. It is possible to refor-

mulate these models in terms of quality-adjusted variables. Thus reinterpreted,

these models predict that the relative number of varieties in a sector or product

is a negative function of relative pure (or quality-adjusted) prices.15

15. In a multi-country set up, the relative number of varieties between two countries is also
determined by the pure prices of third countries. Therefore, Assumption 3 implicitly imposes

16
 Assumption 4 imposes the restriction that there is no correlation between

country-pair excess variety and bilateral di¤erences in pure relative prices.

0 0
ekk
Assumption 4: covs n ekk
z ; p z =0

This assumption is not very strong, as there is no obvious relationship be-

tween the country pair’s excess variety relative to the world average and relative

comparative advantage among countries within the pair.

With assumptions 3 and 4 as well as our earlier assumptions about consumer

utility, we obtain the main result of this section:

Proposition 1. Under Assumptions 1 through 4, for any two countries k and

k 0 , the (unobservable) Impure Price Index is bounded by the (observable)

Paasche and Laspeyres indexes:

0 0 0
ln Hskk ln Pskk ln Lkk
s

Proof. See Theory Appendix.

This …nding provides the basis for our estimation of the Impure Price In-

dex in the …rst-stage of our empirical strategy. As noted above, it assumes all

countries are active in the same set of products. As discussed in the Theory

Appendix at the end of this paper, the more general case of imperfect overlap

may result in violations of Proposition 1. We show however that such viola-

tions are less likely when the number of mismatched products is low and when

mismatched products are more evenly distributed across countries in a pair. As

discussed further in Section 5, we attempt to mitigate the possibility of such

violations in our empirical analysis by excluding country pairs with few export

that bilateral price e¤ects dominate over price e¤ects with respect to third countries. We
thank a referee for making this point.

17
products in common and by considering subsets of our sample countries which

overlap in greater numbers of products.

IV. Net Trade as Indicator of Pure Price

Variation

This section derives the theoretical relationship between countries’net trade

and their Pure Price Indexes. Exporting goods from country k to country k 0
kk0 kk0
requires paying iceberg trade costs of s . Therefore, pkz s is the import price

of product z in country k 0 . Given the CES preferences over products in sector s

outlined above, it is easy to derive country k’s bilateral export and import ‡ows

in sector s with every other country. Summing export ‡ows over all partners

k 0 6= k, we obtain the value of country k’s exports,

2 1 s
3
0
X 6X nkz pekz kk s 7
k k0
(12) Exportss = 4 1 5 bs E
k0 6=k z (Gks 0 ) s

0 0 1 s XX 00 00 1 s
k00 k0
where Gks is a consumption-based price aggregator and Gks = nkz pekz s .
k00 z
0
E k is the expenditure of country k 0 and equals its income (Y k0 ) minus its trade

balance (T k0 ). The expression in brackets in equation (12) is country k’s share

in country k 0 ’s sectoral expenditure. Prices and quality levels a¤ect this share

only through their ratio, pekz .16

In a similar manner, we obtain the value of country k’s imports,

" 1
#
X nkz pekz s

(13) Importsks = 1 1 bs E k :
z (Gks ) s

16. We can associate an in…nite price pekz with a product z that is not produced in country
k. Since pure prices are elevated to a negative exponent, this product will have no e¤ect on
the volume of trade or the price aggregator.

18
Subtracting equation (13) from equation (12), we obtain country k’s net trade

with the world in sector s, Tsk , as a proportion of its expenditure in the sector,
!
Tsk X nk 1
z
pekz
s k
(14) = exp s 1
bs E k z
E k

!
X kk0
1 s
k k0
where s = ln E s
Gk 0 :
s
k0
k
The summary measure of trade costs, s, captures bilateral trade costs be-

tween all country pairs. First, it includes all outbound bilateral trade costs for
kk0 k
country k. Those costs, s , enter directly, so that s is smaller the higher are

those costs. Second, via Gks , k

s also includes all inbound bilateral trade costs
0
k k k
for country k, s , so that s is larger the larger are those costs. Finally, all

other bilateral trade costs enter indirectly through countries’consumption price

0
indexes, Gks , dampening the negative e¤ect of outbound bilateral trade costs.

As a result, net trade of country k is higher the higher are trade costs between

third countries.17

Equation (14) shows that a country’s net trade (per expenditure in the sec-

tor) is a function of its pure prices and numbers of varieties, its total expenditure,
k
and a summary measure of its bilateral trade costs, s. Our objective is to de-

rive a version of equation (14) that reduces the dimensionality of unobservables

and that can be related to the estimated Impure Price Index.

To achieve this objective, de…ne country k’s “multilateral excess variety” in

nk P k
ekz = nkz nz , where
product z as n ez = 0; 8k = 1; :::; K. The covariance be-
n
s
z
tween multilateral excess variety and (normalized) pure prices can be expressed

as the sum of a common component across countries ('s ) and a mean-zero,

17. See Anderson and van Wincoop (2003) for a detailed discussion of the e¤ects of trade
costs on trade ‡ows in a related setting.

19
country-speci…c idiosyncratic component18

1 s
(15) ekz ; pekz =Pesk
covs n = 's + k
s;

Based on the same theoretical results that motivate Assumption 3, we postulate

a negative relationship between the number of varieties and pure prices, de…ned

here across sectors rather than across products within sectors.19

s
Assumption 5: nks =Y k = Pesk ; 8k = 1; :::; K; s 0:

A particular case of this assumption is when s = 0, in which case the average

number of varieties in a sector is a constant proportion of income. Here, we allow

for a more general case where the number of varieties is allowed to decrease as

pure prices increase.20

The following Proposition describes the main result of this section.

Proposition 2. Under Assumption 5, country k’s net trade in sector s (above

the sector’s proportional share in total net trade) can be approximated as a

log-linear function of Pesk

Tsk bs T k
(16) = s + s ln Pesk + bs k
s + k
s
Ek

18. Note that this characterization does not impose any restriction on the covariance. For
estimation, we will assume that ks and the instrumental variable are uncorrelated.
19. In fact, in a coarse check of this assumption discussed further in the web-based technical
appendix, we …nd a negative relationship between our estimated pure price indexes and the
number of ten-digit HS products within manufacturing sectors that countries export to the
United States (normalized by GDP) over our sample period.
20. This relationship abstracts from home market e¤ects or “multilateral” e¤ects such as
being close to low- or high-pure price countries, which could a¤ect the number of varieties
that countries produce.

20
 where

k k
s = bs Z s ' s ; s = bs (1 s s) < 0; s = bs Zs s;

Proof. See Theory Appendix.

Proposition 2 provides a simple expression for the relationship between net

trade and pure prices. This proposition formalizes the key insight of the paper.

Price variation not accompanied with corresponding quality variation implies

variation in pure prices. Even though unobservable, pure prices are manifest

in sectoral trade balances. In particular, the surplus in a country’s sectoral net

trade –above the sector’s share in total net trade –should be larger the lower

are its pure prices.

In addition to pure prices, trade costs also in‡uence net trade. Proposition 2

characterizes this in‡uence. Since the proposition captures the impact of trade

costs on net trade conditional on pure prices, it does not provide a comparative

statics assessment of the e¤ect of trade costs on net trade. Changes in those

costs will typically a¤ect pure prices in general equilibrium, implying an indirect

e¤ect on net trade not captured in equation (16). Note that our method does

not require that we identify the economic forces that determine pure prices in

equilibrium. It only requires that we control for them. Variation in pure prices

can be driven by traditional sources of comparative advantage, or it can be the

result of macroeconomic conditions, such as over- or under-valued currencies.

Equation (16) can be interpreted as a demand function, where the sectoral

net trade with the world is the “quantity” variable, Pesk is the “price” variable,
k
and s is a demand shifter. The …rst term captures movements along the de-

mand curve: higher pure prices of country k in sector s are associated with a

worsening of this country’s net trade position in that sector. The second term

captures movements of the demand curve. Conditional on pure prices, higher

21
inbound trade costs relative to outbound trade costs shift this curve to the right.

We use countries’trade balances with the world as the “quantity”indicator

in our method to mitigate our inability to control for unobserved components of

bilateral trade costs, i.e., information costs, idiosyncratic transport costs, and

non-tari¤ barriers associated with commercial policy. By using trade balances

rather than either exports or imports alone, we cause unobserved components

of countries’trade costs that a¤ect both exports and imports in a country pair

to cancel out. By using countries’trade balances with the world, i.e. summing

a country’s trade ‡ows across all of its trading partners, we average out the

impact of unobserved, idiosyncratic components of bilateral trade costs.21 ;22

Still, unobserved components of trade costs that are neither canceled out by

using trade balances nor averaged out by using trade balances with the world

will inappropriately feed into our estimates of quality.

V. Estimation

In this section we demonstrate how our theoretical results can be used to

estimate U.S. trading partners’ relative manufacturing quality from 1989 to

2003. Estimation is accomplished in two stages. We discuss the strategy of each

stage, as well as their data requirements, separately. Throughout, we focus

on the key issues associated with implementing our method, deferring detailed

discussions of dataset creation to a separate, web-based technical appendix.23

21. Khandelwal (2008), by contrast, relies on “demand” information contained in the im-
ports of a single trading partner (the United States). An advantage of that approach is that
U.S. imports can be observed at a more disaggregate level than world trade. A disadvantage
is that, for the reasons noted above, one-way ‡ows to a single country are likely to be sub-
stantially more sensitive to mismeasurement of trade costs than countries trade balances with
the world.
22. We discuss results based on exports to the United States as an alternate measure of
“demand” in the web-based technical appendix.
23. Datasets and computer code developed to generate our results are also available with
this web-based technical appendix.

22
V.A. Estimation of First-Stage Impure Price Indexes

The …rst stage of the estimation uses Proposition 1 to estimate each country’s

Impure Price Index, Pbsko ; 8k 6= o, where country o is the numeraire country

(without loss of generality) and hats over variables denote estimates. For generic

country pair k and k 0 , the estimated indexes Pbsko and Pbsk o implicitly determine
0

the bilateral index Pbskk = Pbsko =Pbsk o . This index should satisfy the Paasche and
0 0

Laspeyres bounds for that country pair, as outlined in Proposition 1. Similarly,

for K trading partners, the K 1 estimated Impure Price Indexes Pbsko ; 8k 6= o,

implicitly determine K(K 1) bilateral indexes Pbskk ; 8(k; k 0 ); that should satisfy
0

the bilateral Paasche and Laspeyres bounds for all country pairs.

If export prices and quantities were observed without error, estimation would

entail searching for an interior solution to the set of observed Paasche and

Laspeyres bounds across country pairs. Given that import data may be mis-

recorded on customs documents, however, we allow for measurement error in

the bounds by assuming that Paasche and Laspeyres indexes are observed im-
0 0
precisely. Denote the “true”Paasche and Laspeyres indexes by Hs kk and Ls kk ;
0 0
respectively. We assume that the observed indexes, Hskk and Lkk
s , depart from
0 0 0
the true indexes by a multiplicative error: in logs, ln Hskk = ln Hs kk + %kk
h;s and
0 0 0
ln Lkk
s = ln Ls kk +%kk
l;s . We also assume that each error is distributed normally,

with mean zero and standard deviation s, and that the errors for each bound

are independent both of each other and of error terms for other bilateral pairs.24

Satisfying the inequality constraints of Proposition 1 for a given pair of

countries implies:

0 0 0 0 0
(17) ln Pskk ln Hs kk ) %kk
h;s ln Hskk ln Pskk
0 0 0 0 0
(18) ln Pskk ln Ls kk ) %kk
l;s ln Lkk
s ln Pskk :

24. This is a potentially strong assumption because the price (unit value) of a single product
might show up in many bounds, inducing correlated rather than independent errors.

23
Separately for each year t, we estimate a set of index numbers ln Pbsko ; 8k 6= o, and

the standard deviation of the error term b s by maximizing the joint likelihood

that the intervals de…ned by all “true” Paasche and Laspeyres bounds contain

the estimates, i.e. the likelihood that (17) and (18) are jointly satis…ed for each

country pair fk; k 0 g. This criterion implies maximizing the function

( " !# !)
X X ln Hskk
0
ln Pskk
0
ln Lkk
0
ln Pskk
0
s
ln L = ln 1 + ln
k kk0 >k s s

where is the cumulative normal.

Intuition for this estimator is provided in Figure I, which considers the

0
Paasche-Laspeyres interval for a single country pair k and k 0 , de…ned by ln Hskk
0
and ln Lkk
s . In the …gure, two cumulative normal distributions, each with stan-

dard deviation s, take values of one half at each end of the interval. Consider a

pair of Impure Price Index estimates relative to the numeraire and the location

of their (log) ratio ln Pbskk = ln Pbsko ln Pbsk o along the horizontal axis in the
0 0

…gure. According to equation (17), the height of the cumulative normal distrib-

ution to the left of ln Pbskk indicates the likelihood that the true Paasche index is
0

lower than the estimated bilateral index, that is, ln Hs kk < ln Pbskk . Likewise,
0 0

using equation (18), the height of the cumulative normal to the right of ln Pbskk
0

indicates the likelihood that the true Laspeyres index is greater than the esti-

mated bilateral index, that is, ln Ls kk > ln Pbskk . Choosing a particular value
0 0

for ln Pbskk inevitably involves increasing the value of one of these functions at
0

the expense of the other. If the objective were to maximize the likelihood that

ln Pbskk is within the true bilateral Paasche and Laspeyres bounds, only taking
0

into account the bounds of this particular country pair, then ln Pbskk would lie
0

in the middle of the interval and be equivalent to the well-known Fisher index.

However, because the choices of ln Pbsko and ln Pbsk o , which determine ln Pbskk for
0 0

this country pair, also in‡uence the …t of all other country pairs in which either

24
country k or k 0 are present, the estimates that maximize the joint likelihood for

all country pairs will not in general be located in the center of the interval for

countries k and k 0 . For that reason, ln Pbskk is drawn o¤-center in the interval
0

depicted in Figure I.

Our estimator has the advantage of penalizing estimates that lie inside the

interval only in relation to the likelihood that conformance to the theory is a

consequence of measurement error. Similarly, it penalizes estimates outside the

interval only in relation to the likelihood that violation of the bounds restric-

tion is not caused by measurement error. We note that this estimator is not a

conventional maximum likelihood estimator as it does not maximize the like-

lihood of observing the data (the bounds) given the parameters (the Impure

Price Indexes).25

V.A..1 First-Stage Data Requirements

Estimation of countries’ Impure Price Indexes requires data on countries’

export prices and quantities. Here, we rely on detailed U.S. import statistics

published by the U.S. Census Bureau. These data record the total customs value

and quantity of U.S. imports by year, source country and ten-digit Harmonized

System (HS) product classi…cation from 1989 to 2003. We focus on U.S. import

data given its level of detail and availability for such a long time horizon, but

note that our method can be generalized to include data from other countries,

which could be used to generate additional Paasche and Laspeyres bounds that

25. In the web-based technical appendix, we compare our estimator to three alternatives: a
quadratic penalty function centered at the midpoint of each country pair’s interval; a function
that only penalizes estimates outside the interval; and an index proposed by Hummels and
Klenow (2005) which compares countries’prices to those of the world over the set of goods they
have in common with the world. We …nd that the …rst two alternatives yield IPI and quality
estimates very similar to those reported below. Results using the third alternative vary more
substantially from those reported below. However, the goodness of …t of that alternative, i.e.,
the percent of …rst-stage Impure Price Index estimates that lie within the Paasche-Laspeyres
bounds, is considerably lower, thus supporting our choice of the estimator de…ned in the main
text.

25
could be incorporated into the estimation. Our use of U.S. trade data presumes

that U.S. import prices and quantities are representative of countries’exports

to other markets.26

We compute the unit value, or “price”, of export product z from source

country k, pkz , by dividing free-on-board import value (vzk ) by import quantity

(qzk ), pkz = vzk =qzk , where free-on-board refers to import values that are exclusive

of customs, insurance and freight charges.27 Examples of the units employed to

classify products include dozens of men’s cotton shirts in apparel, square meters

of wool carpeting in textiles and pounds of folic acid in chemicals. We focus

on manufacturing exports, where a product is classi…ed as manufacturing if it

belongs to SITC industries 5 through 8. Following standard practice, we ex-

clude SITC 68, non-ferrous metals, from manufacturing. We note that quantity

information is missing for approximately 20 percent of observations in the raw

data; these observations are dropped.

Unit values are noisy due to both aggregation and measurement error (GAO

1995). To mitigate the impact of these errors, we both restrict our analysis

to relatively large exporters and screen the raw data. First, we start with the

world’s top 50 exporters of manufactured goods by value. Second, we employ

two types of screens to eliminate suspect observations. “Primary” screening

drops observations where only a single unit is shipped in a year or where the U.S.

CPI-de‡ated annual import value is below $25,000 in 1989 dollars. “Secondary”

screening makes the primary quantity and value cuto¤s more stringent while

imposing four additional criteria. First, a (more stringent) Relevance Constraint

mandates that country-product-year observations must have quantity greater

than 25 and value (in 1989 dollars) greater than $50,000. Second, a Presence

26. This assumption may not be innocuous. In principle, it could be tested by comparing
the results of this section to results based on other countries’data
27. A sustained assumption in our framework is that the export unit values that we observe
are not systematically di¤erent from the prices charged to domestic consumers, which we do
not observe.

26
Constraint requires country-product observations to appear in more than two

years of the sample. Third, a Country-Pair Overlap Constraint insists that, for

a country-pair comparison to be included in the sample in any given year, the

two countries must export at least 25 products in common to the United States.

Finally, a Unit-Value Dispersion Constraint requires that country-product-year

observations be excluded if the country’s adjusted28 unit value is less than one-

…fth or more than …ve times the geometric mean of all prices for the product in

that year.

After secondary-screening the data, we impose a …nal constraint that data

required for both the …rst and second stage cannot be missing for more than

three years of the sample period. After all screens are implemented, we are left

with 43 countries, which constitute the sample we use in the remainder of the

paper.

The costs and bene…ts of screening the raw data can be discerned from Table

I. Each row of the table focuses on a di¤erent screen, while each column indicates

the a¤ect of the screen on a di¤erent aspect of the 2003 sample, though we note

that screening has a similar e¤ect across years. To promote comparability, all

rows in the table are restricted to the same set of 43 countries available after the

most stringent screening (that is, the screening in the …nal row of the table).

The …rst column of Table I demonstrates that secondary screening reduces

the value of imports captured in the sample by 11 percent vis a vis the primary-

screened sample. The next two columns of Table I show that secondary screening

also reduces country and country-product participation in the sample, lowering

28. The adjustment accounts for the likelihood that very high export prices are more likely
to be the result of misrecording if they come from countries with relatively low average export
prices, and vice versa. To implement this screen, we perform two iterations of the …rst-stage
estimation. In the …rst iteration, we estimate Impure Price Indexes after eliminating observa-
tions under the unit-value-dispersion constraint without making any adjustment to country’s
unit values. In the second iteration, we divide a country’s unit values by the estimated Impure
Price Index from the …rst iteration prior to implementing the unit-value-dispersion screen. We
note that omitting the second iteration has relatively little impact on our second-stage quality
estimates.

27
the number of country pairs for which data is available to 829 from 861 and

the median number of products country pairs export in common to the United

States from 347 to 228. As illustrated in the …nal column of the table, there are
0 0
very few incorrectly ordered Paasche and Laspeyres bounds (i.e., Lkk
s < Hskk )

in all three screens; for our preferred sample, just 0.6 percent of bounds are

ordered incorrectly. We exclude those bounds from our estimation.

The primary bene…t of screening is substantially tighter Paasche and Laspeyres

bounds. As indicated in the fourth column of the table, the median interval
0 0
length (ln Lkk
s ln Hskk ) under the preferred secondary screening is 0.74, less

than one-third the length under the primary screen, 2.51. The reduction in

interval length results in a substantial improvement in estimation precision.

Of the additional criteria imposed by secondary screening, the unit-value

dispersion constraint exerts the strongest a¤ect on median interval length. For

example, an “alternate”secondary screening (not shown) that omits the require-

ment that adjusted unit values be within one-…fth and …ve times the geometric

mean for the product-year results in a disproportionately large increase in me-

dian interval length (to 2.01 from 0.74) versus import value (to 97.8 from 88.8

percent).

The left-hand panel of Table II summarizes several dimensions of the pre-

ferred sample, by year. The …rst column of the panel illustrates that the sample

of countries is held constant at 43 for the entire sample period. The …nal column

of the panel shows that the median Paasche-Laspeyres interval across country

pairs measured in log points moves between 0.68 and 0.78 over the sample pe-

riod. The remaining columns of the panel demonstrate that the number of

country pairs, the total number of product-country-pairs, and the median num-

ber of common products across country pairs all rise over time. These increases

are driven by growth in the number of products countries export to the United

28
States over the sample period.

As highlighted in Section 3 and discussed further in the Theory Appendix

below, the imperfect overlap of export products between countries induces po-

tential violations to Proposition 1. Such violations might generate composition

bias in the estimates of the Impure Price Indexes and, as a result, in estimates

of the Quality Indexes. Further, growth in the product coverage of countries’

exports might change the extent of bias over time, also a¤ecting the estimated

time trends. As noted above in Section 5.1.1 we attempt to mitigate the in-

‡uence of composition bias via the use of the Country-Pair Overlap constraint

when screening the raw data.29 Below, we also compare the quality rankings of

the thirty largest exporters in our sample to alternate estimates derived from re-

stricting the analysis to just those thirty countries. Since these thirty countries

exhibit substantially higher export-product overlap than all countries in the

base sample, our …nding of similar relative quality in both estimations suggests

that composition bias, if present, is limited.

V.A..2 First-Stage Results

The right-hand panel of Table II summarizes the results of the …rst-stage

estimation by year. Column one of the panel shows that the log likelihood

declines in absolute value over time, while column two reports that the estimated

standard deviation, b s , is relatively constant at approximately 0.15 over the

sample period. The third column of the panel reports the estimation’s goodness

of …t in terms of the percent of …rst-stage Impure Price Index estimates that lie

within the Paasche-Laspeyres bounds. As indicated in the table, this share is

above 90 percent in all years and rises from 90.4 percent in 1989 to 93.8 percent

29. Data restrictions prevent implementation of other potential solutions to this problem.
We cannot restrict analysis to a set of continually exported country-products, for example,
due to numerous changes to Harmonized System product classi…cation codes over the sample
period.

29
in 2003.

Estimation of the …rst stage yields an Impure Price Index for each country

relative to the numeraire country. In Figure II, we report normalized log Impure

Price Indexes for all countries for the …rst and last years of the sample. This

normalization involves subtracting the mean log index across countries from

every country’s estimated log Impure Price Index, by year

1 X b k0 o
(19) ln Pbst
k;M ean
= ln Pbst
ko
ln Pst :
K 0
k

In particular, the normalized Impure Price Index for the numeraire country
X
(Switzerland), ln Pbst
o;M ean
, is equal to 1 K ln Pbst
k0 o
.
k
Estimated Impure Price Indexes generally accord with expectations. In the

…gure, countries nearer the lower left corner such as Pakistan (PAK) and China

(CHN) exhibit relatively low export prices in both years vis a vis the mean

while countries in the upper right corner like Ireland (IRL) and Switzerland

(CHE) exhibit consistently high relative export prices. Countries’ orientation

with respect to the grey forty-…ve degree line illustrates changes in relative

prices over time. Countries like Hungary (HUN) and Morocco (MAR) that

lie above the forty-…ve degree line exhibit rising relative export prices, while

those below the forty-…ve degree line like China (CHN) and Singapore (SGP)

experience declining relative prices. In both years, the ordering of countries

accords well with their level of development. Note that a mapping of country

codes to country names is provided in Table IV.

V.B. Estimation of Second-Stage Quality Indexes

The second stage of our estimation uses Proposition 2 to recover information

about countries’ relative quality from their …rst-stage estimated Impure Price

Indexes. First, we sum and subtract s ln Peso to the right-hand side of equation

30
(16) to express it as a function of the Pure Price Index (relative to numeraire

o) rather than of the price aggregator Pesk . Then, since we calculate bs from

data, we take the trade cost term to the left-hand-side. Finally, we use ln Pesko =
ko
ln Psko ln s to rewrite this equation as

(20) Test
k
= 0
st + s ln Pbst
ko
s ln ko
st + ko
s st + bs Zs k
st

where Test
k k
= Tst bs Ttk =Etk bs k
st , t indexes time periods, 0
st = st +

s ln Peso , and ko
st
ko
= ln Pst ln Pbst
ko
is the estimation error from the …rst stage.

The last three terms in equation (20) are unobservable and create a compound

error term that includes: countries’ product quality relative to the numeraire
ko ko
country ( st ); the estimation error in the …rst stage ( st ); and the idiosyncratic
k
component of the covariance between excess variety and pure prices ( st ) from

equation (15). Assuming that this compound error term is uncorrelated with
ko
the regressors is untenable. In particular, the quality component st may be

correlated with the estimated Impure Price Index: developed countries, which

tend to have higher export prices, are also likely to produce higher quality.

To deal with this endogeneity, we …rst specify a linear time path for the

evolution of product quality relative to the numeraire country,

ko ko ko
(21) ln st = 0s + 1s t + "ko
st

ko ko
where 0s and 1s are a country …xed e¤ect and the slope of a country-speci…c

time trend, respectively, and "ko

st represents deviations of quality from this trend.

As in the estimation of the …rst stage, results here do not depend upon the

choice of numeraire country, and we again choose Switzerland for this role.

Incorporating the country-speci…c linear time trend for quality into equation

31
(20), we obtain our second-stage estimating equation

(22) Test
k
= 0
st + s ln Pbst
ko ko
0s
ko
1s t + ko
st

ko ko ko ko
where 0s = s 0s and 1s = s 1s are country …xed e¤ects and time trends,
ko ko
respectively, and st = s ( st "ko
st ) + bs Zs
k
st is the error term. Note that the

term on the left-hand-side could be expressed relative to the numeraire country,

but that doing so would have an impact only on the year …xed e¤ects.

The inclusion of country …xed e¤ects in (22) eliminates the most obvious

source of endogeneity, i.e. the cross-sectional correlation between the time-

invariant components of countries’ prices and quality levels. The inclusion of

country-speci…c time trends further reduces the remaining correlation between

regressor and error term, as the latter term now includes only deviations of

quality from country-speci…c trends. However, correlation between "ko

st and

Pbst
ko
may still persist, as shocks to quality may be accompanied by increases

in (impure) prices.

To address this potential problem, we use the real exchange rate as an instru-

ment for the estimated Impure Price Index. As usual, the instrument needs to

satisfy two conditions. First, because the estimating equation includes country-

speci…c …xed e¤ects and time trends, the instrument has to be correlated with

ln Pbst
ko
, after controlling for the …xed e¤ ects and time trends. In other words, de-

viations of the real exchange from its own time trend have to be correlated with

similar deviations of Pbst

ko
. Macroeconomic conditions typically determine peri-

ods of over- and under-valuation of countries’real exchange rate around long-run

trends. These periods also determine changes in the international competitive-

ness of a countries’ exports, captured in our model by Pest

ko
. Since Pest
ko
is a

component of Pbst
ko
, periods of over- or under-valuation are also associated with

32
movements of Pbst
ko
, providing the necessary correlation. Second, the instrument

has to be uncorrelated with the error term "ko

st , which requires that shocks to

quality around the trend in sector s are not correlated with the real exchange

rate. While we cannot rule out that such a correlation exists, we judge it to be

relatively unimportant. Shocks to quality in sector s might be accompanied by

exactly o¤setting changes in prices, leaving pure prices – and hence net trade

in that sector –unchanged. Even if these shocks a¤ect pure prices, they might

have a negligible e¤ect on the real exchange rate. This is more likely to be true

if the shocks are temporary deviations around a trend, and if they are speci…c

to sector s, that is, uncorrelated with shocks to quality in other sectors. Finally,
ko k
we also assume that both st and st are uncorrelated with the real exchange

rate.

We estimate equation (22) using two-stage least squares (2SLS). Our esti-

mate of country k’s Quality Index relative to the numeraire is

!
ko bko + bko t
(23) ln bst = 0s 1s
;
bs

where t indexes the number of years since 1989 and the remaining right-hand

side variables are estimates from equation (22). Note that we identify only

the linear trend in quality. Deviations of quality from the linear trend are

confounded with the other two components of the error term and are therefore

not included in our estimated Quality Indexes.

Countries’estimated Pure Price Indexes are derived from equation (22) and
ko
the de…nition of ln bst in equation (23). They are equal to
!
b ko ko Test
k b 0st bko
(24) ln Pest = ln Pbst
ko
ln bst = st
:
bs

33
 We note that this estimate of the Pure Price Index inherits any estimation

error in both the Impure Price Index and the Quality Index. In particular

deviations of quality from the trend ("ko

st ) are mis-attributed to the Pure Price

Index.

V.B..1 Second-Stage Data Requirements

Second-stage data requirements are strong relative to data availability. Ob-

taining reliable information about countries’ trade balances, for example, is

challenging because countries vary greatly in how they report this information

to international agencies. Similarly, collection of countries’product-level trade

barriers did not begin in earnest until 1989 and has grown …tfully since then.

Here, we provide a brief description of how our datasets are constructed. Further

detail is available in our web-based technical appendix.

Trade balance data are drawn from the United Nations Commodity Trade

Statistics Database (COMTRADE). This dataset records bilateral import and

export ‡ows between countries by manufacturing industry and year. Our overall

approach to obtaining countries’ net trade is to subtract each country’s total

reported imports from its total reported exports by industry and year.30 We

measure countries’annual net trade in overall manufacturing as well as the in-

dustries within manufacturing discussed below. As required by equation (22),

we normalize trade balances by total expenditure (E k ). We compute E k by sub-

tracting total net trade from GDP. Both variables are drawn from the World

Bank’s World Development Indicators (WDI) database except from Taiwan’s

GDP, which comes from the Economist Intelligence Unit website. We also need

30. Unfortunately, country pairs’ reported trade ‡ows with each other are often mutually
inconsistent. Since our principal interest is the accuracy of countries’ overall net trade with
the world, we favor this approach, which maximizes reporting consistency within countries, to
the one taken by Feenstra et al. (1997, 2000), which generally relies on reporting countries’
import statistics to estimate bilateral trade ‡ows. Further details of our data re…nement
procedures are described in a web-based technical appendix.

34
to compute the share of manufacturing in total expenditure, bs = Esk =E k . To

compute the numerator, we subtract manufacturing net trade from manufac-

turing value added. The latter variable is drawn from the United Nations’

National Accounts O¢ cial Country Data. We obtain bs = 0:214 as the average

share across countries and years.

We measure trade barriers in terms of transport costs and tari¤s. We mea-

sure country pairs’bilateral transport costs using U.S. import data, which record

both the customs-insurance-freight (cif) and free-on-board (fob) value for most

import ‡ows. Restricting our analysis to the preferred screened sample de-

scribed above, we de…ne transport costs as akzt = cifzt

k
f obkzt =f obkzt and we

estimate ad valorem transport costs per mile across all z in industry s in year

t by regressing the relative value spent on customs, insurance and freight on

imports from country k on the distance the exports have travelled,

(25) ln ak;U
zt
S
= st ln Dk;U S + 0
st X
k;U S
+ 2kzt ;

where Dk;U S represents the great circle distance in kilometers between the

United States and country k and X k;U S represents additional controls, including

whether country k shares a common language or border with the United States
0
or was ever a colony of the United States. In the estimations below we set akk
st
0
equal to exp bst ln Dkk + b st X kk .
0 0

Tari¤ information is derived from the Trade Analysis and Information Sys-

tem (TRAINS) Database maintained by the United Nations Conference on

Trade and Development (UNCTAD). In principle, these data record countries’

most favored nation (MFN) tari¤s as well as any preferential (PRF) tari¤ rates

that might be available for a subset of trading partners at the eight-digit Harmo-

nized System level. In practice, product-country coverage in the dataset is very

35
sparse, hampering our ability to control properly for trade policy in equation

(22).
kk0
We compute bilateral trade costs s by adding the measures of bilateral

transport costs and tari¤s explained above. The aggregation of those measures
k
to construct the trade cost term st is more challenging because it requires

values for the unobserved consumption price indexes Gk0

s de…ned in Section 4.

Up to a factor of proportionality (captured by the constant in the regression),

X 00 00 1 s
the component nkz pekz in the indexes is the share of country k 00 in
z
world production of sector s in a world equilibrium with no trade costs. We

approximate this share by the share of country k 00 in “world” exports of that

sector, i.e., the total exports of all countries in the preferred estimation sam-

ple. While this approximation is imperfect, the theoretical and observed shares

should both largely be driven by country size. As a result, this approximate

measure should capture a substantial fraction of the relevant variation in the

unobserved shares.31

Finally, to compute countries’real exchange rates, we use the real e¤ective

exchange rate series reported by the Economist Intelligence Unit (EIU) on their

website. Though the EIU dataset is reasonably complete, we …ll in any holes in

it by using data from the World Bank and the International Monetary Fund.

V.B..2 Second-Stage Results

Table III reports second-stage estimates of s from the estimation of equa-

tion (22) by OLS and two-stage least squares (2SLS).32 Robust standard errors

31. The consumption indexes Gks also require an estimate of the elasticity of substitution
k
s . We compute st using s = 6 and note that alternative values of s ranging from 3 to 10
have almost no impact on our results.
32. Given our rejection of a unit root using the test developed by Levin et al. (2002),
we perform the estimation in levels rather than in di¤erences. The test is performed on
the dependent variable, each of the regressors, and the residual allowing alternatively for a
constant and for both a constant and a time trend. The null hypothesis that there is a unit
root is rejected at the 1% signi…cance level in all cases.

36
adjusted for clustering at the country level are reported below each coe¢ cient.33

As indicated in the table, the OLS estimate of s has the expected negative sign

but is statistically insigni…cant. The 2SLS estimate, on the other hand, is both

negative and statistically signi…cant as well as an order of magnitude lower than

the OLS estimate, -0.241 versus -0.028. The …nal row of the table reports an

F-statistic for the …rst stage of 2SLS of 37:7.

Log Quality Index intercepts and slopes, normalized by annual means as in

equation (19), are displayed in Figure III along with their 95 percent con…dence

bands.34 Estimated intercepts are equivalent to countries’relative log quality in

0:480
1989. As indicated in the …gure, China’s quality in 1989 is two-thirds (e )

that of the mean country in that year, while Germany’s is more than twice as

high (e0:768 ). Estimated slopes report how much relative quality rises or falls

vis a vis the mean country each year. Ireland has the highest slope while Hong

Kong has the lowest.

Figure III sorts countries according to their intercepts, from low to high.

Though these intercepts vary widely, they tend to be high for developed economies

like Switzerland and Sweden and low for developing countries like Malaysia and

the Philippines. Quality slopes also vary substantially across countries but ap-

pear to be inversely related to intercepts. Two noticeable outliers to this pattern

are Singapore and Ireland, both of which are estimated relatively imprecisely.

Normalized Quality Indexes across the sample period are displayed along

with 95 percent con…dence bands for a set of nine countries in Figure IV. As

indicated in the …gure, China’s relative quality is ‡at over time and generally

below those of Germany, Japan and Singapore. Relative quality rises for Hun-

33. An estimate and standard error for s that accounts for the fact that the IPIs are esti-
mated are computed using the bootstrap method described in the web-based technical appen-
dix. They are 0:254 and 0:091, respectively. Results are similarly close for our industry-level
estimates below.
34. Standard errors are computed using the delta method. Quality intercepts and slopes
are reported for each country in tabular form in the web-based technical appendix.

37
gary, Thailand, Malaysia and Singapore, though estimates for the latter two

countries are relatively imprecise.

Our results are robust to a number of sensitivity analyses. First, we obtain

similar point estimates after selectively removing each country from the esti-

mation, indicating that results are not overly dependent on the presence of any

particular country. Second, we …nd similar results using more or less stringent

secondary screens, though standard errors are generally larger the more lax is

the screening. Finally, we assess the potential impact of composition bias by

restricting the sample to the 30 largest exporters. This restriction doubles the

median number of products country pairs export in common across all years of

the sample period, substantially reducing the potential in‡uence of violations

to Proposition 1. We …nd that the rank correlations of relative quality rankings

across countries appearing in both samples is above 97 percent in all years.

VI. What Can We Learn from Quality

Estimates?

In this section we compare our estimates of countries’manufacturing quality

to raw export prices and examine links between quality and long-run growth. We

…nd that changes in raw relative export prices can be a poor approximation of

changes in quality, and that consideration of price and quality together provides

complementary information about variation in development strategies across

countries. Indeed, our results suggest two potential paths towards long-run

growth: via quality, as in the case of Malaysia and via price competitiveness, as

in the case of China.

One of the most important lessons to take from our estimates of manufac-

turing quality is that changes in quality inferred from our method can be quite

38
di¤erent from changes in countries’ relative export prices. Figure V compares

the change in countries’normalized log Quality Indexes versus their change in

normalized log Impure Price Indexes between 1989 and 2003. Though these

two changes are positively correlated (0.33), quality and prices move in oppo-

site directions for one-third of the sample. For Malaysia, Singapore, Thailand

and Indonesia, quality rises while raw export prices fall, while the opposite is

observed for many countries such as Argentina, Greece, Portugal, and Switzer-

land. These divergences between quality and impure prices are due to variation

in countries’ global net trade balances. Figure VI, for example, shows that

Malaysia’s rising trade balance combined with its relatively ‡at impure prices

results in rising estimated quality. For China, relatively moderate increases in

its manufacturing trade surplus combined with a falling Impure Price Index

imply falling pure prices and ‡at quality.

Quality levels across countries compress over time. The two panels of Table

IV display countries’quality rankings and normalized quality levels at four-year

intervals from 1989 to 2003. Countries are sorted according to their ranking

in 1989, with the …nal column in each panel noting the change in ranking or

level over the entire sample period. Countries whose rank changes by more than

ten places are noted with an asterisk. Singapore, Indonesia, Malaysia, and the

Philippines are the countries with the largest increases in quality ranking while

Australia, Hong Kong, New Zealand, and Poland are those with the largest

decreases. This reshu- ing of quality rankings is associated with strong quality

convergence. The mean quality level of countries below the overall average of all

countries rises from -0.48 in 1989 to -0.35 in 2003 while the mean for countries

above the overall average increases from 0.42 to just 0.44.35 A critical question

is whether this compression reveals catching up in terms of technological capa-

35. This narrowing is even more dramatic if Ireland is excluded.

39
bility by the developing world or quality upgrading via the use of higher-quality

intermediate inputs without corresponding gains in productivity.

The theory of economic growth has long established a connection between

quality and growth. In particular, the “quality ladder” models developed by

Grossman and Helpman (1991) and Aghion and Howitt (1992) postulate coun-

tries’ability to upgrade quality as a form of productivity gain that also implies

increases in GDP per capita. To gain further understanding of the compres-

sion in quality levels manifest in our estimates, we compare normalized quality

compression to the convergence in exporters’per-capita income over our sample

period. We divide our sample countries into two cohorts according to whether

per capita GDP is below or above the median in 1989. Interestingly, we …nd

that quality convergence has not been accompanied with convergence in GDP

per capita. While the di¤erence between the average quality of high-income

versus low-income countries decreased from 0.67 log points in 1989 to 0.38 log

points in 2003, this di¤erence remained almost constant for per capita GDP

(2.20 log points in 1989 versus 2.14 log points in 2003).

Quality and per capita GDP are strongly correlated in the cross section of

countries. The correlation between countries’normalized quality index and their

similarly normalized log per capita GDP is positive and statistically signi…cant

across all years of our sample, with an average correlation across years of 0.46.

However, consistent with the lack of correspondence between quality and per

capita income convergence described above, we …nd that this correlation declines

over time, from 0.54 in 1989 to 0.32 in 2003.36 The weakening association

between quality and per capita GDP also appears in the positive but smaller

correlation between changes in quality and changes in per capita income over

the sample period (0.30 and signi…cant at the 5 percent level). This correlation is

36. Hummels and Klenow’s (2005) estimates of the cross-sectional elasticity of quality with
respect to income in 1995 ranges from 0.09 to 0.23.

40
displayed visually in Figure VII. As indicated in the …gure, countries with above

average change in quality vary substantially in terms of their per capita GDP

growth. For several of these countries, including Ireland, Chile, and Hungary,

relatively high quality growth is accompanied by relatively high per capita GDP

growth, consistent with a process of quality upgrading based on gains in (quality-

adjusted) productivity. Other countries that display substantial quality growth,

however, do not exhibit strong growth in income. Malaysia, the Philippines,

Indonesia and Thailand, for example, are among the countries with the most

impressive increases in manufacturing quality, but this does not appear to have

translated in higher incomes. This outcome is consistent with quality growth

achieved via use of higher-quality inputs, in which case quality growth might

not require enhanced productive capabilities and hence need not raise income.

Figure VII also suggests alternative paths for attaining higher income that

are not associated with upgrading quality. Here, China – which combines ex-

traordinary per capita GDP growth with almost no change in quality – serves

as perhaps the most interesting and illustrative example of quality growth not

being a prerequisite for income growth. There are many potential explanations

for this outcome – e.g., the attractiveness of China’s large domestic market,

its transition from a command economy to a more market-based economy, the

e¤ect of its export promotion policies –all of which are worthy of further study.

Overall, the divergence in income and quality growth paths displayed in Fig-

ure VII suggests a variety of development strategies that countries might pursue

and the usefulness of understanding the economic mechanisms that each strat-

egy involves. Despite its importance, there has been relatively little empirical

investigation into the link between product quality in economic development,

mostly due to lack of measures of product quality.37 We hope the method and

37. Hausmann, Hwang and Rodrik (2006), for example, investigate the link between simi-
larity of developing countries’export baskets with those of developed countries and growth.

41
estimates proposed here help improve this situation.

VII. Conclusion

This paper attempts to …ll an important gap in the international trade and

development literature by proposing a method for identifying countries’product

quality over time. First, we show how an important but unobserved Impure

Price Index comparing countries’export prices is bounded by their observable

Paasche and Laspeyres indexes. Then, we develop a method for decomposing an

estimate of this Impure Price Index into Quality and Pure Price Indexes. This

method makes use of information on consumers’valuation of countries’products

contained in their net trade with the world and allows for both vertical and

horizontal product di¤erentiation. In contrast to a vast literature that associates

cross-country variation in export unit-values with variation in product quality

–implicitly assuming away cross-country variation in quality-adjusted prices –

our method allows for price variation induced by factors other than quality, e.g.,

comparative advantage or currency misalignment.

Implementation of our method reveals trends in product quality not evi-

dent in export prices alone. Indeed, for several countries, export prices and

quality evolve quite di¤erently. Our estimation also highlights the importance

of accounting for intermediate trade in estimating countries’ product quality.

Further theoretical and empirical e¤orts on this front will be quite useful.

Estimates of countries’product quality are obviously useful for testing mod-

els of international specialization and development. They may also bene…t other

…elds, such as productivity and growth, where, despite the existence of an in‡u-

ential theoretical literature linking the production of quality to economic growth

(e.g., Grossman and Helpman 1991, Aghion and Howitt 1992), empirical inves-

tigation is scarce. Quality-adjusted price indexes will likely also …nd use in the

42
labor literature. The distributional consequences of international trade implied

by the Stolper-Samuelson theorem, for example, cannot be properly identi…ed

if the import and export price changes used to empirically assess the theorem’s

relevance do not properly account for changes in countries’product quality.

Universidad de San Andres and NBER

Yale School of Management and NBER

43
 Appendix I –Theory

Proof of Proposition 1
0 0 kk0
Since we have already established that ln Hskk ln Pskk + ln s;c , to demon-
0 0 0
kk
strate that ln Hskk ln Pskk we only need to show that ln s;c 0, which is

equivalent to showing that

X nk 0
z
(26) pekk
z 0.
z
nks

nk 0 0
Since z
nk
ek;kk
=n z ekk
+n z + nz , the left-hand-side of (26) can be written as
s

X nk 0 X 0 0 X 0 0 X 0
z
(27) pekk
z = ek;kk
n z pekk
z + ekk
n z pekk
z + nz pekk
z :
z
nks z z z

0 P
P k;kk0 kk0 0 0
Using the property ek;kk
n z = 0, ez
n pez = Zs covs nek;kk
z ; pekk
z 0
z z
P kk0 P kk0 kk0 0 0
by Assumption 3. Using ez = 0;
n ez
n pez = Zs covs nekk
z ; p ekk
z =
z z
P
0 by Assumption 4. Finally, using the de…nition of Peskk ,
0 0
nz pekk z = 0.
z
Combining these results we obtain:

X nk 0 0 0 0 0
z
pekk
z ek;kk
= Zs covs n z ; pekk
z ekk
+ Zs covs n ekk
z ; pz 0;
z
nks

0 0
which demonstrates that ln Hskk ln Pskk . An analogous proof shows that
0 0
ln Pskk ln Lkk
s . Hence, the Paasche and Laspeyres indexes bound the Impure

Price Index,
0 0 0
ln Hskk ln Pskk ln Lkk
s :

44
Proof of Proposition 2

Substituting nkz = nks (nz + n

ekz ) into equation (14), we can rewrite the right-

hand side of that equation as

2 !1 !1 3
s s
nk 1 s X pekz X pekz
1 + sk Pesk exp k
s
4 nz + ekz
n 5:
E z Pesk z Pesk

X ek ek
Using the de…nition of Pesk and the fact that
p p
ekz ( Pezk )1
n s
ekz ; ( Pezk )1
= Zs covs n s
,
s s
z
the above expression can be rewritten as

2 0 !1 13
s
nk Y k 1 s pekz
1 + sk k Pesk exp k
s
41 + Zs covs @n
ekz ; A5 :
Y E Pek s

Yk Tk
Using Assumption 5, equation (15), and the fact that Ek
= 1+ Ek
, we can

substitute the latter expression for the right-hand side of (14). Rearranging

terms and taking natural logarithms, we obtain

(28)
Tsk Tk 1 s s
ln 1 + = ln 1+ Pesk exp k
s 1 + Zs 's + k
s :
bs E k Ek

Using ln(1 + x) ' x, and abstracting from the approximation error, we can

…nally express equation (28) as

Tsk bs T k
= s + s ln Pesk + bs k
s + k
s
Ek

where
k k
s = bs Zs 's ; s = bs (1 s s) < 0; s = bs Zs s

45
Imperfect Overlap in Active Products

Proposition 1 assumes that all countries export the same products. Here,

we outline the more general case in which the set of exported products varies

across countries. Let Is be the set of all product categories in sector s. De…ne a

country as “active” in product z if it reports positive exports in that category.

The set Is can be decomposed into two parts: Is(kk0 ) ; which includes categories

in which countries k and k 0 are both active (with Zs(kk0 ) denoting the number
00
of such categories), and its complement, Is(kk0 ) . Accordingly, de…ne pks(kk0 )
00
and qks(kk0 ) as the vector of prices and quantities, respectively, of imports from
k
country k 00 2 fk; k 0 g in product categories included in Is(kk0 ) . The vector qs(kk 0)

is the complement of qks(kk0 ) with respect to q.

In this setting, the constrained expenditure function should be de…ned over

k 00
product categories in Is(kk0 ) . As a result, ms;k(kk0 ) (pks(kk0 ) ; qs(kk 0 ) ; n; ; ;u)

solves the problem

(29)
00
k
min pks(kk0 ) q
bks(kk0 ) s:t: qks(kk0 ) ; qs(kk
U (b 0 ) ; n; ; ) = u; k 00 = 1; :::; K:
bk
q
s(kk0 )

The solution to problem (29) yields an expression analogous to (5) but with

summations ranging over elements of Is(kk0 ) . Similarly, inequalities (7) and (9)
kk0 kk0
continue to hold as long as s;k and s;k0 are rede…ned over products in Is(kk0 ) .

For the Impure Price Index to be bounded from below by the Paasche Index,
P nkz 0 nk 0 0
it must be the case that nk
pekk
z ek;kk
0 . Given that nzk = n z + nkk
z ;
s s
z2Is(kk0 )
and that Is(kk0 ) is the complement of Is(kk0 ) with respect to Is , we can write

(30)
X nkz 0 X 0 0 X 0 0 X 0 0
pekk
z = ek;kk
nz pekk
z + nkk
z pekk
z nkk
z pekk
z :
z2Is(kk0 )
nks z2Is(kk0 ) z2Is z2Is(kk0 )

46
The …rst term in (30) can be expressed as

X 0 0 0 0 X 0
(31) ek;kk
nz pekk
z ek;kk
= Zs(kk0 ) covIs(kk0 ) n z ; pekk
z + pekk
s
0
ek;kk
n z
z2Is(kk0 ) z2Is(kk0 )
X 0
pekk
s
0
ek;kk
nz
z2Is(kk0 )

where the second line uses Assumption 3 –now only including products in Is(kk0 )
P 0 P 0
in the covariance – and the property ek;kk
n z = ek;kk
n z , and
z2Is(kk0 ) z2Is(kk0 )
1
P 0
where kk0
pes Zs(kk0 ) pekk
z . The second term in (30) has already been
z2Is(kk0 )
0 0
shown to equal zero in the proof of Proposition 1 (note that nkk
z ekk
=n z + nz ).

Combining these results we obtain

X nkz 0 X 0 0 0
pekk
z ek;kk
n z pekk 0 kk
s + nz pekk
z ;
z2Is(kk0 )
nks
z2Is(kk0 )

which is greater than zero (i.e., Impure Price Index is bounded from below by

the Paasche Index) only if

X 0 0 0
(32) ec;kk
n z pekk 0 kk
s + nz pekk
z 0:
z2Is(kk0 )

Analogously, to guarantee that the Impure Price Index is bounded above by the

Laspeyres Index, we need

(33)
X 0 0 0 0 X 0 0 0
ekz ;kk
n pesk0 k + nkk
z pekz k = ek;kk
n z pekk
s
0
nkk
z pekk
z 0:
z2Is(kk0 ) z2Is(kk0 )

A su¢ cient condition for (32) and (33) to hold simultaneously is that the

two countries are active in the same set of goods – the assumption made in

Section 3 to simplify the exposition. In that case the set Is(kk0 ) is empty and

47
the left hand side of both inequalities are zero since they sum over elements of

an empty set. Unfortunately, since countries often are not “active”in the same

set of products, we need to analyze how imperfect overlap in active products

might violate the result of Proposition 1.

The summations in conditions (32) and (33) include two terms. The …rst
0
nc;kk
term (e z pekk 0
s ) is common to both conditions while the second term enters

each condition with the opposite sign. As a result, the larger the absolute value

of the second term, the more likely one of the two conditions is violated.

In considering the absolute magnitude of the second term, note that it is a

0
weighted sum of pekk
z (de…ned in equation (11)) across “mismatched” prod-

ucts, i.e. products in which one country is active but the other one is not. Also
0
note that pekk
z > 0 when k is active and k 0 is not, and vice versa. Thus, other

things equal, the absolute magnitude of this term increases with the number

of “mismatched” categories in the country pair and the asymmetry with which

these mismatched products are distributed across countries in the pair. Viola-

tions are less likely, for example, when countries k and k 0 are each active in ten

products not produced by the other than when c is active in twenty products

not produced by k 0 but every product produced by k 0 is also produced by k.

In the former, the elements of the sum will tend to cancel out as they have

opposite signs. In the latter, condition (32) is not satis…ed due to composition:
0
while the Impure Price Index Pskk is de…ned over Is , the Paasche index we
0
kk
observe, Hs(kk 0 ) , is computed over the subset Is(kk 0 ) . Therefore, even though

0 0 0 0
kk 38
Hskk Pskk is true, Hs(kk
kk
0 ) > Ps . Put another way, in this example, the

subset Is(kk0 ) is not a representative sample of the broader set Is over which
0
Pskk is de…ned. As a result, the Paasche index fails to include (mismatched)

0
38. Since pkz = 1 for any non-active z, including those products in the Paasche index would
0 0
result in ln Hskk = 1, as the corresponding elements pkz qzk in denominator of the index will
0 0 0
be in…nite. Thus, Hskk Ps will be (trivially) true but not informative to estimate Pskk .
kk

48
products that have a particularly low price in k relative to the (in…nite) price

in k 0 :39

The …rst term in conditions (32) and (33) is less problematic. We can write

0 1
1 BX 0 X 0C
(34) pekk
s
0
= @ pekk
z pekk
z A.
Zs(kk0 )
z2Is z2Is(kk0 )

P 0
It is not possible to sign pekk
z . However, as a benchmark we can use the
z2Is
P 0
fact that this sum is similar –except for the absence of weights –to nz pekk
z ,
z2Is
which we have shown equals 0. Thus, its departure from zero will depend on

the extent to which substituting a constant weight of 1 for nz a¤ects the sum.

Abstracting from this sum, the …rst term in (32) and (33) can be written as

0 10 1
1 B X 0C B X 0C
(35) @ ek;kk
n z A@ pekk
z A:
Zs(kk0 )
z2Is(kk0 ) z2Is(kk0 )

If mismatched products are asymmetrically active in country k, then both terms

in parentheses in equation (35) are positive and the expression overall is nega-

tive, as needed to satisfy both conditions simultaneously. If mismatched prod-

ucts are asymmetrically active in country k 0 , then both parentheses are negative

but expression (35) is still negative. In sum, even though it is not possible to

demonstrate the sign of the …rst term in (32) and (33), this analysis suggests

that it is likely to be negative and thus help satisfy both conditions.

In our empirical analysis, we attempt to mitigate violations of conditions

(32) and (33) by excluding country pairs with very few (i.e., less than 25) ex-

port products in common. As a robustness check, we have also re-estimated

quality for the 30 largest countries in the sample, whose export-product overlap

0
39. The Laspeyres index, by constrast, is una¤ected since, as qzk = 0 for non-active products
in k0 , including those products only adds zeros to its numerator and denominator.

49
is substantially higher than for the larger sample. As noted in the main text,

quality rankings across those 30 countries are very highly correlated with the

rankings reported in our baseline estimation.

Appendix II –Estimating Quality Within

Manufacturing

In this appendix we compute separate Quality Indexes for the four one-

digit SITC industries that constitute manufacturing. To explore the potential

in‡uence of countries’use of intermediate inputs outside of the sectors at which

quality is being estimated, we also estimate quality across the two, two-digit

SITC sectors that represent apparel and textiles.

One-digit SITC Sectors

Estimation of quality within manufacturing relies upon the same strate-

gies and datasets described above. To conserve space, we omit a discussion of

screening, but note that primary and secondary screens exert similar in‡uence.

The number of countries that can be included in the analysis varies by indus-

try because all countries do not participate equally in each industry. Of the

43 countries used for aggregate manufacturing, we are left with 27 countries in

Chemicals, 41 countries in Manufactured Materials, 37 countries in Machinery

and 41 countries in Miscellaneous Manufacturing.

Table A.1 reports estimation results as well as details of the …rst-stage es-

timation sample by industry and year. Across industries, the data are thicker

in terms of product-country-pair observations and median products in common

for Manufactured Materials and Miscellaneous Manufactures than for Machin-

ery and Chemicals. Goodness of …t in terms of the share of estimates falling

50
within bounds is highest in Machinery and lowest in Manufactured Materials

and Miscellaneous Manufactures, and this ordering generally remains consistent

with the ordering of their Paasche-Laspeyres intervals from high to low.

Figure A.1 reports estimates of countries’normalized …rst-stage Impure Price

Indexes by manufacturing industry for 2003 versus 1989. As indicated in the …g-

ure, prices are most tightly distributed in Chemicals (except for outlier Ireland)

and are most dispersed in Miscellaneous Manufactures. We …nd that countries’

Impure Price Indexes are positively correlated across industries. In 2003, this

correlation is highest for Manufactured Materials versus Miscellaneous Manu-

factures (0.82) and lowest for Chemicals versus Machinery (0.54).

Table A.2 reports the second-stage OLS (top panel) and 2SLS (bottom panel)

estimates of s by industry.40 The 2SLS estimates of s have the expected neg-

ative sign in all four industries, but the estimate for Chemicals is statistically

insigni…cant. The strength of the real exchange rate as an instrument for the

Impure Price Index varies across industries. The F-statistic for the …rst stage

of the 2SLS regression is high for both Machinery and Miscellaneous Manufac-

tures, low for Manufactured Materials, but especially low (0.01) for Chemicals.

A potential explanation for this result is that Chemical products are less hor-

izontally di¤erentiated than products in Machinery or Miscellaneous Manufac-

tures. If that were the case, export prices might not be responsive to movements

in countries’ real exchange rate and instead respond mostly to movements in

world prices. This hypothesis receives some support from the relatively low

price dispersion exhibited in the Chemical Impure Price Indexes (Figure A.1).

Normalized log Quality Index intercepts and slopes along with their standard

errors are displayed along with their 95 percent con…dence bands in Figure A.241

40. We compute bs for each one-digit sector using the procedure for overall manufacturing
described above. The values are 0.035, 0.043, 0.072, and 0.034 for SITC 5, 6, 7, and 8,
respectively. For more detail, see our technical appendix.
41. Standard errors are computed using the delta method. Quality intercepts and slopes are
reported for each country and industry in tabular form in the web-based technical appendix.

51
As with aggregate manufacturing, the ordering of quality intercepts generally

accords with expectations: developed economies have the highest intercepts in

Machinery, for example, while Italy’s intercepts are relatively high in Manufac-

tured Materials and Miscellaneous Manufactures, which include Textiles (SITC

65) and Apparel (SITC 84), respectively. Given the weak results for the Chem-

icals sector, we exclude if from further analysis.

Disaggregated quality estimates reveal substantial variation in quality inter-

cepts across industries within countries. China, for example, is at the low end

in both Manufactured Materials and Machinery but in the middle of the pack in

Miscellaneous Manufacturing. Hong Kong and Taiwan, on the other hand, have

relatively high intercepts for Miscellaneous Manufactures but are in the bottom

tercile of Machinery. Overall, we …nd a positive and statistically signi…cant cor-

relation of quality intercepts across countries for Manufactured Materials and

Machinery but little correlation between Miscellaneous Manufactures and the

other two sectors.

Quality Index slopes display similar variation: across countries the non-

Chemical industry slopes have the same sign in only 14 of the 43 countries

in the sample. These di¤erences are highlighted in Figure A.3 for the subset

of nine countries examined in Figure IV. For Singapore, relative quality in-

creases strongly in all three sectors. For Malaysia, quality increases strongly in

Machinery but is relatively ‡at in Manufactured Materials and Miscellaneous

Manufactures.

Quality convergence within manufacturing also varies across industries. Fig-

ure A.4 reports the evolution of mean quality for countries with initially high and

low per capita income. Trends are displayed for overall manufacturing as well

as for Manufactured Materials, Machinery and Miscellaneous Manufactures. As

indicated in the …gure, there is a substantial narrowing of quality in Machinery,

52
weaker convergence in Miscellaneous Manufactures and an unchanging quality

gap in Manufactured Materials.

Two-digit SITC Sectors: Apparel and Textiles

As noted in the introduction, our method for estimating product quality

involves an aggregation trade-o¤. For broad SITC categories such as all manu-

facturing, the assumption of constant quality across all products in the category

is strong, but data on countries’global net trade is more readily available and

more likely to be reliable. Another potential advantage of using broader SITC

sectors is their ability to dampen the e¤ect of countries’use of imported inter-

mediate inputs. Use of such inputs is an issue when they straddle the sectors

at which quality is being estimated. Countries with a strong comparative ad-

vantage in one sector, for example, might be large net exporters of that sector

but large net importers of a second sector which is an input to the …rst. All else

equal, this situation may lead quality in the …rst and second sectors to be over-

and underestimated, respectively.

In principle this problem can be solved by using either value added trade data

or input-output tables to map imported intermediates to …nal goods. Unfortu-

nately, data for pursing these solutions is generally unavailable across countries

and time. Here, we take advantage of the well-known linkage between Apparel

(SITC 84) and Textiles (SITC 65) to examine separately the estimated qual-

ity of apparel versus a hybrid sector, Apparel & Textiles (SITC 65&84), that

combines the goods from both two-digit SITC industries.

Table A.3 reports 2SLS estimates of s for SITC 65, SITC 84 and the hy-

brid Apparel & Textiles. As indicated in the table, estimates of s are negative

and signi…cant for all three groups of goods.42 Figure A.5 compares normal-

42. Quality intercepts and slopes for all three sets of goods are reported for each country in
tabular form in the web-based technical appendix.

53
ized Apparel versus Apparel & Textiles quality across countries in 2003. While

estimated quality for the two sets of goods is highly correlated, outliers are

apparent. Pakistan, for example, generally has a higher trade surplus in tex-

tiles than apparel. As a result, normalized quality for Apparel & Textiles is

substantially higher than it is for Apparel alone.

Such outliers suggest that controlling for intermediate inputs may have an

important in‡uence on estimated quality. As a result, it seems prudent to

include as much information about input-output linkages as possible when esti-

mating quality across disaggregate product categories. Over time, as collection

and dissemination of more detailed data on countries’ international trade be-

comes available, this task should become easier.

References

Abed-el-Rahman, K., “Firms’ Competitive and National Comparative Advan-

tages as Joint Determinants of Trade Composition”, Weltwirtschaftliches
Archiv, 127 (1991), 83-97.
Aghion, Phillippe and Peter Howitt, “A Model of Growth Through Creative
Destruction,” Econometrica, 60 (1992), 323-351.
Aiginger, Karl, “The Use of Unit Values to Discriminate Between Price and
Quality Competition,” Cambridge Journal of Economics, 21 (1997), 571-
592.
Aiginger, Karl, “Unit Values to Signal the Quality Position of CEECs. In The
Competitiveness of Transition Economies,” OECD Proceedings, 1998 (10),
1-234.
Alterman, William F., W. Erwin Diewert, and Robert C. Feenstra, “Interna-
tional Trade Price Indexes and Seasonal Commodities,” Washington DC:
U.S. Department of Labor, Bureau of Labor Statistics, 1999.
Anderson, James E., and J. Peter Neary, “Trade Reform with Quotas, Partial
Rent Retention, and Tari¤s,” Econometrica, 60 (1992), 57-76.
Anderson, James, and Eric van Wincoop, “Gravity with Gravitas: A Solution
to the Border Puzzle,” American Economic Review, 93 (2003), 170-192.
Aw, Bee Yan, and Mark J. Roberts, “Measuring Quality Change in Quota-
Constrained Import Markets,” Journal of International Economics, 21
(1986), 45-60.

54
Bernard, Andrew B., Stephen Redding, and Peter K. Schott, “Heterogenous
Firms and Comparative Advantage,”Review of Economic Studies, 74 (2007),
31-66.
Berry, Steven T., “Estimating Discrete Choice Models of Product Di¤erentia-
tion,” The RAND Journal of Economics, 25 (1994), 242-262.
Bils, Mark, “Measuring the Growth from Better and Better Goods,” NBER
Working Paper No. 10606, 2004.
Boorstein, Randi, and Robert C. Feenstra, “Quality Upgrading and its Welfare
Cost in U.S. Steel Imports, 1969-74,”NBER Working Paper No. 2452, 1987.
Broda, Christian, and David E. Weinstein, “Globalization and the Gains from
Variety,” Quarterly Journal of Economics, 121 (2006), 541-585.
Brooks, Eileen, “Why Don’t Firms Export More? Product Quality and Colom-
bian Plants,” Journal of Development Economics, 80 (2006), 160-178.
Fabrizio, Stefania, Deniz Igan, and Ahoka Mody, “The Dynamics of Product
Quality and International Competitiveness,” IMF mimeo, 2007.
Feenstra, Robert C., “Quality Change Under Trade Restraints in Japanese Au-
tos,” Quarterly Journal of Economics, 103 (1988), 131-146.
Feenstra, Robert C., “New Product Varieties and the Measurement of Interna-
tional Prices,” American Economic Review, 84 (1994), 157-177.
Feenstra, Robert C., “Exact Hedonic Price Indexes,”Review of Economics and
Statistics, 77 (1995), 634-653.
Feenstra, Robert C., “World Trade Flows, 1980-1997,”Center for International
Data, UC Davis, 2000.
Feenstra, Robert C., Robert E. Lipsey, and Harry P. Bowen, "World Trade
Flows, 1970-1992, with Production and Tari¤ Data," NBER Working Paper
No. 5910, 1997.
Feenstra, Robert C., Robert E. Lipsey, Haiyan Deng, Alyson C. Ma, and Hengy-
ong Mo, “World Trade Flows: 1962 to 2000”, NBER Working Paper No.
11040, 2005.
Feenstra, Robert C., John Romalis, and Peter K. Schott, “U.S. Imports, Ex-
ports, and Tari¤ Data, 1989-2001,” NBER Working Paper No. 9387, 2002.
Feenstra, Robert C, Alan Heston, Marcel Timmer and Haiyan Deng, “Esti-
mating Real Production and Expenditures Across Nations: A Proposal for
Improving Existing Practice,” UC Davis, mimeo, 2007.
Flam, H., E. Helpman, “Vertical Product Di¤erentiation and North-South
Trade,” American Economic Review, 77 (1987), 810-822.
General Accounting O¢ ce, “US Imports: Unit Values Vary Widely for Identi-
cally Classi…ed Commodities,” Report GAO/GGD-95-90, 1995.
Goldberg, Penny, and Nina Pavcnik, “Distributional E¤ects of Globalization in
Developing Countries,” Journal of Economic Literature, 45 (2007), 39-82.
Grossman, Gene and Elhanan Helpman, “Quality Ladders in the Theory of
Growth,” The Review of Economic Studies, 58 (1991), 43-61, 1991.
Hallak, Juan C., “Product Quality and the Direction of Trade,” Journal of
International Economics, 68 (2006), 238-265.
Hausmann, Ricardo, Jason Hwang, and Dani Rodrik, “What You Export Mat-
ters,” Journal of Economic Growth, 12 (2007), 1-25.

55
Hummels, David, and Peter Klenow, “The Variety and Quality of a Nation’s
Exports,” American Economic Review, 95 (2005), 704-723.
Ianchovichina, Elena, Sethaput Suthiwart-Narueput, and Min Zhao, “Regional
Impact of China’s WTO Accession,”in East Asia Integrates: A Trade Policy
Agenda for Shared Growth, Krumm, Kathie and Homi Kharas eds. (Wash-
ington, DC: The International Bank for Reconstruction and Development /
The World Bank, 2003).
Levin, Andrew, Chien-Fu Lin and Chia-Shang James Chu, “Unit root tests in
panel data: asymptotic and …nite-sample properties,”Journal of Economet-
rics, 108 (2002), 1-24.
Khandelwal, Amit, “The Long and Short of Quality Ladders,” Review of Eco-
nomic Studies, forthcoming, 2009.
Neary, J. Peter and K.W.S. Roberts, “The Theory of Household Behaviour
Under Rationing.” European Economic Review, 13 (1980), 25-42.
Rodrik, Dani, “What’s So Special About China’s Exports?”, Mimeo, Harvard
University, 2006.
Romalis, John, “Factor Proportions and the Structure of Commodity Trade,”
American Economic Review, 94 (2004), 67-97.
Schott, Peter K., “Across-Product versus Within-Product Specialization in In-
ternational Trade,” Quarterly Journal of Economics, 119 (2004), 647-678.
Schott, Peter K., “The Relative Sophistication of Chinese Exports,” Economic
Policy, 53 (2008), 5-49.
Verhoogen, Eric, “Trade, Quality Upgrading, and Wage Inequality in the Mex-
ican Manufacturing Sector: Theory and Evidence from an Exchange-Rate
Shock,” Quarterly Journal of Economics, 123 (2008), 489-530.
Verma, Samar, “Export Competitiveness of Indian Textile and Garment In-
dustry,” Indian Council for Research on International Economic Relations,
Working Paper No. 94, 2002.
Xu, Bin, “Measuring China’s Export Sophistication,” Mimeo, CEIBS, 2007.

56
 Table I
Sample Attributes for All Manufacturing in 2003, by Screening
Percent of
Unscreened Explicit Median Median Correctly
Sample's Country-Pair Common Interval Ordered
Import Value Comparisons Products Length Bounds
Unscreened Sample 100.0% 861 1213 4.46 99.9%
Primary Screened Sample 99.8% 861 347 2.51 99.8%
Preferred Secondary Screened Sample 88.8% 829 228 0.74 99.4%
Notes: Table displays several attributes of the estimation sample for all manufacturing in 2003 according to
three methods of screening the raw data. All samples contain the same set of 43 countries. Import value for
each sample is expressed as a percentage of the unscreened sample. Explicit country-pair comparisons is the
number of country pairs that appear in the sample; the maximum is 903 (i.e., 43*42/2). Median common
products is the median number of export products exported in common to the United States across country
pairs appearing in the sample. Median interval length is the median log difference between Paasche and
Laspeyres bounds. Correctly ordered bounds is the percent of bounds in the sample for which the Paasche
index is less than the Laspeyres index.
 Table II
Sample and First-Stage Estimation Attributes, All Manufacturing
Attributes of Estimation Sample Attributes of First-Stage Estimation
Median
Common Median Log First-Stage
Products Product- Paasche- Maximization Estimates
Country-Pair Across Country-Pair Laspeyres Objective Standard Within
Countries Observations Country Pairs Observations Interval Function Error Bounds
1989 43 811 133 208,108 0.74 -349 0.16 90.4%
1990 43 829 143 223,564 0.68 -334 0.14 90.8%
1991 43 814 144 219,596 0.69 -332 0.15 91.5%
1992 43 814 146 224,875 0.73 -322 0.14 91.2%
1993 43 823 156 239,190 0.74 -319 0.16 90.6%
1994 43 846 171 263,986 0.73 -320 0.16 91.8%
1995 43 858 185 292,615 0.76 -272 0.14 94.2%
1996 43 862 190 308,028 0.72 -251 0.13 93.5%
1997 43 866 206 328,629 0.69 -310 0.15 93.3%
1998 43 869 221 342,476 0.73 -291 0.15 93.5%
1999 43 873 226 350,882 0.76 -245 0.14 93.7%
2000 43 877 249 374,151 0.72 -300 0.16 93.0%
2001 43 875 238 358,160 0.78 -222 0.14 94.1%
2002 43 831 234 341,940 0.74 -239 0.15 94.5%
2003 43 829 228 330,968 0.74 -271 0.16 93.8%
Notes: First panel displays characteristics of the preferred first-stage estimation sample, by year. Second
panel displays attributes of the first-stage estimation.
 Table III
Second-Stage Regression Results for All Manufacturing
OLS 2SLS
Impure Price Index -0.028 -0.241 ***
0.023 0.084
Observations 640 640
R2 0.93 0.90
First-Stage F Statistic . 37.7
Notes : Table displays OLS and 2SLS regression results for estimation
of equation (22) on the preferred sample (see text). Coefficients for
country fixed effects and time trends are omitted. Heteroskedasticity-
robust standard errors adjusted for clustering at the country level are
reported below each coefficient. The instrument for the Impure Price
Index in the 2SLS results is the real exchange rate. *** denotes
statistical significance at the 1 percent level.
 Table IV
Quality Rankings, All Manufacturing
Rank Normalized Quality
Country 1989 1993 1998 2003 Change 1989 1993 1998 2003 Change
Switzerland (CHE) 1 2 2 4 -3 0.93 0.84 0.73 0.62 -0.31
Sweden (SWE) 2 3 5 6 -4 0.83 0.75 0.65 0.55 -0.28
Germany (DEU) 3 5 7 9 -6 0.77 0.66 0.54 0.41 -0.36
Finland (FIN) 4 4 3 3 1 0.67 0.67 0.67 0.67 0.00
Italy (ITA) 5 6 8 8 -3 0.66 0.59 0.51 0.42 -0.24
France (FRA) 6 8 9 10 -4 0.63 0.54 0.44 0.34 -0.29
Japan (JPN) 7 9 10 12 -5 0.57 0.47 0.33 0.20 -0.38
Belgium (BEL) 8 7 6 5 3 0.54 0.55 0.55 0.55 0.01
United Kingdom (GBR) 9 10 12 17 -8 0.47 0.38 0.26 0.15 -0.33
Denmark (DNK) 10 11 11 14 -4 0.45 0.37 0.27 0.17 -0.28
Ireland (IRL) 11 1 1 1 10 0.45 0.87 1.41 1.94 1.50
Austria (AUT) 12 12 13 16 -4 0.42 0.34 0.25 0.15 -0.27
Israel (ISR) 13 13 16 19 -6 0.38 0.27 0.14 0.02 -0.36
*Australia (AUS) 14 19 23 26 -12 0.27 0.11 -0.10 -0.31 -0.58
Taiwan (TWN) 15 17 18 22 -7 0.24 0.15 0.03 -0.09 -0.33
Spain (ESP) 16 18 19 24 -8 0.23 0.14 0.02 -0.10 -0.33
*Hong Kong (HKG) 17 21 25 31 -14 0.23 0.06 -0.15 -0.36 -0.59
Canada (CAN) 18 22 24 27 -9 0.21 0.06 -0.13 -0.31 -0.52
Norway (NOR) 19 20 20 21 -2 0.18 0.10 0.01 -0.08 -0.26
Netherlands (NLD) 20 16 15 13 7 0.17 0.17 0.17 0.18 0.01
Korea, Republic of (KOR) 21 15 14 11 10 0.17 0.19 0.21 0.23 0.06
*New Zealand (NZL) 22 25 30 38 -16 0.08 -0.08 -0.28 -0.48 -0.57
Portugal (PRT) 23 23 22 20 3 0.01 0.00 -0.01 -0.02 -0.04
Argentina (ARG) 24 26 26 25 -1 -0.11 -0.14 -0.18 -0.21 -0.10
Hungary (HUN) 25 24 17 15 10 -0.16 -0.07 0.04 0.16 0.31
Brazil (BRA) 26 27 29 28 -2 -0.19 -0.23 -0.28 -0.33 -0.14
*Singapore (SGP) 27 14 4 2 25 -0.19 0.19 0.66 1.13 1.31
Mexico (MEX) 28 28 31 33 -5 -0.33 -0.35 -0.38 -0.41 -0.08
Turkey (TUR) 29 29 33 34 -5 -0.33 -0.37 -0.41 -0.45 -0.11
Greece (GRC) 30 30 34 35 -5 -0.41 -0.42 -0.43 -0.45 -0.03
Romania (ROM) 31 32 38 40 -9 -0.42 -0.45 -0.48 -0.51 -0.09
*Poland (POL) 32 35 41 43 -11 -0.42 -0.47 -0.53 -0.59 -0.17
Colombia (COL) 33 36 40 41 -8 -0.45 -0.47 -0.50 -0.52 -0.07
South Africa (ZAF) 34 31 32 29 5 -0.46 -0.42 -0.38 -0.35 0.11
China (CHN) 35 37 37 37 -2 -0.48 -0.48 -0.48 -0.48 0.00
India (IND) 36 38 36 36 0 -0.52 -0.50 -0.48 -0.45 0.07
*Indonesia (IDN) 37 33 28 23 14 -0.59 -0.45 -0.27 -0.09 0.50
Morocco (MAR) 38 40 35 30 8 -0.60 -0.53 -0.44 -0.35 0.25
Thailand (THA) 39 41 39 32 7 -0.68 -0.59 -0.48 -0.37 0.31
Pakistan (PAK) 40 42 42 39 1 -0.69 -0.63 -0.56 -0.49 0.20
*Philippines (PHL) 41 39 27 18 23 -0.74 -0.52 -0.24 0.04 0.78
*Malaysia (MYS) 42 34 21 7 35 -0.83 -0.46 0.01 0.47 1.31
Chile (CHL) 43 43 43 42 1 -0.95 -0.84 -0.71 -0.58 0.37
Notes: Table records countries' quality ranking and normalized quality indexes by year. Countries are
sorted according to their 1989 ranking. Final column of each panel notes the change between 1989 and
2003. * denotes countries whose rank changes by more than ten places between 1989 and 2003.
 Table A.1
First-Stage Optimization Results, by Manufacturing Industry
Median Median
Median Common First- Median Common
First-Stage Log Products Stage Log Products
Estimates Maximization Paasche- Product- Across Estimates Maximization Paasche- Product- Across
Within Objective Standard Laspeyres Country-Pair Country Country Within Objective Standard Laspeyres Country Pair Country Country
Bounds Function Error Interval Obs Pair Obs Pairs Bounds Function Error Interval Obs Pair Obs Pairs

Chemicals (SITC 5) Manufactured Materials (SITC 6)

1989 94.7% -44 0.08 0.64 16,042 176 56 90.3% -189 0.13 0.58 59,398 499 66
1990 93.4% -40 0.08 0.69 18,085 198 59 90.2% -197 0.12 0.57 61,994 512 68
1991 95.0% -45 0.11 0.71 17,392 186 63 89.7% -211 0.13 0.55 59,284 510 66
1992 95.6% -54 0.11 0.73 20,035 212 62 88.9% -224 0.15 0.55 61,843 530 67
1993 92.8% -60 0.12 0.72 21,526 220 65 91.3% -238 0.15 0.59 67,827 558 72
1994 96.3% -54 0.11 0.71 23,017 226 66 90.1% -274 0.16 0.61 76,301 594 76
1995 94.9% -76 0.13 0.67 25,889 245 67 90.4% -277 0.17 0.66 84,583 614 82
1996 95.1% -47 0.07 0.76 27,998 262 70 91.4% -248 0.16 0.66 88,890 616 84
1997 93.7% -87 0.12 0.66 30,769 277 72 92.5% -238 0.15 0.65 94,855 628 88
1998 93.7% -88 0.13 0.75 32,375 285 71 91.3% -262 0.16 0.65 101,679 646 91
1999 94.0% -87 0.14 0.73 32,863 288 72 93.0% -233 0.14 0.66 104,242 658 94
2000 94.3% -85 0.12 0.65 34,919 295 74 91.7% -293 0.18 0.66 111,155 681 96
2001 92.2% -109 0.15 0.68 34,193 293 74 93.0% -234 0.15 0.72 103,982 658 94
2002 86.9% -127 0.21 0.65 33,684 278 73 94.7% -189 0.12 0.70 101,514 631 95
2003 84.1% -125 0.21 0.70 33,117 276 75 92.3% -204 0.15 0.69 93,619 589 94

Machinery (SITC 7) Miscellaneous Manufactures (SITC 8)

1989 95.2% -95 0.14 0.76 43,580 365 77 80.0% -302 0.12 0.45 76,610 651 71
1990 92.1% -75 0.11 0.77 44,778 374 78 78.0% -329 0.13 0.43 86,114 679 79
1991 91.1% -94 0.12 0.79 46,765 400 77 81.6% -291 0.13 0.43 82,742 643 76
1992 91.9% -97 0.12 0.85 44,618 384 76 82.6% -287 0.13 0.45 85,662 647 79
1993 91.4% -140 0.17 0.80 47,232 413 75 84.1% -290 0.14 0.52 89,817 651 79
1994 93.2% -90 0.10 0.77 54,816 437 84 85.8% -289 0.14 0.49 96,176 674 83
1995 94.7% -130 0.13 0.83 66,636 508 90 87.5% -279 0.13 0.46 101,637 682 91
1996 92.4% -154 0.13 0.70 74,858 525 96 88.5% -266 0.14 0.50 102,411 680 88
1997 90.9% -184 0.15 0.64 80,935 532 100 92.0% -231 0.11 0.51 107,525 699 94
1998 93.8% -150 0.12 0.72 82,866 559 97 90.4% -200 0.11 0.55 111,583 699 98
1999 92.5% -163 0.14 0.73 85,992 560 100 88.2% -266 0.14 0.56 113,628 695 102
2000 92.5% -137 0.12 0.72 93,946 578 107 88.4% -312 0.16 0.52 120,254 712 110
2001 95.7% -131 0.13 0.75 91,063 592 102 86.1% -276 0.14 0.54 113,308 694 107
2002 93.4% -145 0.15 0.75 86,000 566 98 89.2% -252 0.15 0.58 106,469 656 102
2003 93.7% -143 0.13 0.78 84,451 560 98 90.2% -219 0.12 0.57 105,211 662 102
Notes : Table displays characteristics of the first-stage estimation of Impure Price Indexes by manufacturing industry and year. The number of countries included in the
analysis varies by industry: there are 27 in Chemicals, 41 in Manufactured Materials, 37 in Machinery and 41 in Miscellaneous Manufacturing.
 Table A.2
Second-Stage Regression Results, by Manufacturing Industry
OLS
5 - Chemicals 6 - Manuf Mat 7 - Machinery 8 - Misc Manuf
Impure Price Index 0.007 -0.002 -0.015 *** 0.003
0.008 0.004 0.007 0.006
Observations 400 608 533 610
R2 0.97 0.97 0.92 0.96

2SLS
5 - Chemicals 6 - Manuf Mat 7 - Machinery 8 - Misc Manuf
Impure Price Index 0.089 -0.171 *** -0.090 *** -0.055 *
0.130 0.085 0.041 0.037
Observations 400 608 533 610
R2 0.91 0.77 0.89 0.94
First-Stage F Statistic 0.01 4.21 34.3 13.6
Notes : Table reports OLS and 2SLS regression results for equation (22). The instrument for the Impure Price
Index is the real exchange rate. Coefficients for country fixed effects and time trends are omitted.
Heteroskedasticity-robust standard errors adjusted for clustering at the country level are reported below each
coefficient. The instrument for the Impure Price Index in the 2SLS results is the real exchange rate. ** and ***
denote statistical significance at the 5 and 1 percent level, respectively.
 Table A.3
Second-Stage Regression Results for Apparel and Textiles
2SLS
SITC 65 SITC 84 SITC 65&84
Impure Price Index -0.022 -0.054 * -0.061 *
0.017 0.031 0.034
Observations 434 528 579
R2 0.97 0.91 0.95
First-Stage F Statistic 4.6 11.8 16.0
Notes : Table compares 2SLS regression results for estimation of equation (22) on noted
two digit industries and a hybrid industry that combines SITC 65 and SITC 84. The
instrument for the Impure Price Index is the real exchange rate. Coefficients for country
fixed effects and time trends are omitted. Heteroskedasticity-robust standard errors
adjusted for clustering at the country level are reported below each coefficient. * denotes
statistical significance at the 10 percent level.
 Hcd
s L cd
s
co
Ps
do
Ps

Figure I
Maximizing the Likelihood that the Observed Paasche and
Laspeyres Bounds Contain the Estimated Impure Price Index
 Normalized Impure Price Indexes, 1989 v 2003
All Manufacturing, Mean=0

1
IRL
CHE
NOR
.5

HUN SWE
BEL DNK
NLD ITAFINFRA
ESP GBRDEU
AUT
PRT NZL
MAR
2003

CAN
JPN AUS
0

GRC
ARG

POL TURCHL
PHL KOR
ZAF
ROM BRA
COL SGP
MEX
HKG
IDN MYS
TWN
-.5

INDTHA

PAK CHN
-1

-1 -.5 0 .5 1
1989
Note: Indexes are in logs and are normalized by the mean index across countries in each year

Figure II
First‐Stage Estimated Impure Price Indexes, 2003 Versus 1989
 Normalized Quality Index Intercept
All Manufacturing; With 95 Percent Confidence Interval

2
Log Intercept (Mean=0)
-1 0
-2 1

PAK

BRA

ESP

AUS

DNK

SWE
CHE
IDN
IND
CHN

HUN

CAN

ISR

JPN

DEU
CHL

PHL

COL
POL

NZL

IRL

BEL
ARG

HKG
ZAF

PRT

AUT
ROM
MYS

THA

MEX
SGP

FRA
ITA
MAR

GRC
TUR

KOR
NLD
NOR

TWN

GBR

FIN
Normalized Quality Index Slope
All Manufacturing; With 95 Percent Confidence Interval
.2
Log Slope (Mean=0)
0 -.1.1

PAK

BRA

ESP

AUS

DNK

SWE
CHE
IDN
IND
CHN

HUN

CAN

ISR

JPN

DEU
CHL

PHL

COL
POL

NZL

IRL

BEL
ARG

HKG
ZAF

PRT

AUT
ROM
MYS

THA

MEX
SGP

FRA
ITA
MAR

GRC
TUR

KOR
NLD
NOR

TWN

GBR

FIN

Note: Intercepts and slopes are in logs and are normalized by their means across countries.

Figure III
Normalized Log Quality Index Intercept and Slope, by Country
 Normalized Quality Indexes for Nine Countries, 1989-2003
All Manufacturing; With 95 Percent Confidence Intervals
China France Germany
-2 -1 0 1 2
Log Index (Mean=0)

Greece Hungary Japan

-2 -1 0 1 2

Malaysia Singapore Thailand

-2 -1 0 1 2

1990 1995 2000 1990 1995 2000 1990 1995 2000

Note: Index for each year is normalized by the mean across countries.

Figure IV
Normalized Log Quality Index for Select Countries, 1989 to 2003
 Change in Quality vs Raw Export Prices, 1989 to 2003
All Manufacturing

1.5
IRL

SGP MYS
Change in Normalized Quality
1

PHL
.5

IDN
CHL
THA HUN
MAR
PAK
IND KOR
CHN FIN NLD
BELPRTGRC
0

COL
MEX ARG ROM
BRA TUR
POL
AUTDNK ITA NOR
FRASWE CHE
TWNGBR
DEU ESP
JPN
-.5

CAN
AUS HKGNZL

-.5 0 .5 1
Change in Normalized Impure Price Index

Figure V
Normalized Quality Versus Change in Normalized Impure Price Index
 Normalized Impure, Quality and Pure Price Indexes
All Manufacturing
CHN MYS
.5
.25
Log Index (Mean=0)
0
-.25
-.75

1990 1995 2000 1990 1995 2000

Impure Prices Net Trade Quality Pure Prices

Note: All series are normalized by the mean across countries in each year.

Figure VI
Decomposition of China’s and Malaysia’s Impure Price Index
 Change in Manufacturing Quality vs Income
1989 to 2003

1.5
IRL

MYS SGP
Change in Normalized Quality
1

PHL
.5

IDN
CHL
THA HUN
MAR
PAK
ZAF IND KOR
FIN BEL NLD CHN
0

COL GRC PRT

ARGROM TUR MEX
BRA POL
ITA AUT NOR
SWE CHEFRA DNK
TWN
ISR GBR
ESP
JPN
-.5

CAN
AUS NZL HKG

-.5 0 .5 1
Change in Normalized PCGDP

Figure VII
Change in Normalized Quality Versus Change in Normalized Income
 Normalized Impure Price Indexes, 1989 v 2003
By Manufacturing Industry, Mean=0
5 - Chemicals 6 - Manuf Materials
1
CHE
IRL
IRL
.5

AUT CHE
JPN SWEFRA
FIN
FRA HUN NOR
SGP
JPN
AUTGBR
FINBEL
DEU
AUS
GBR
BEL SWE
DNKDEU NLD
ITA DNK
ESP
PRT
NOR CAN
0

ESP
HKG NLD
ITA CHL
BRAIND POL
MYS ARG
MEX
KOR
NZL
KOR
MEX AUS PHLTUR
BRA
CAN HKG
COL
-1 -.5

ARG
TWN THA TWN
INDIDN
CHN CHN
PAK
2003

7 - Machinery 8 - Misc Manufactures

CHE
1

NOR
IRL DNKFRA
ARG NLD DNK SWE
IRL
NLDITA AUT
.5

FIN
SWE DEU
ESP CHE PRT ESP
BEL
JPNGBR
BEL HUN FIN
DEU
ITAGBRAUT
AUS
FRA
CAN
GRC CAN AUS
MAR
POL
0

BRA
PHLKOR MEX JPN
PRT
SGP ROM TUR ARG SGP
POL CHL
PHL
COL KOR
HKG
-1 -.5

MYS IDNBRA
THA
MEX
TWN IND
IND
THA
HKG CHNTWNMYS
PAK ZAF
CHN

-1 0 1 2 -1 0 1 2
1989
Note: Indexes are in logs and normalized by the mean index across countries in each year

Figure A.1
Normalized Impure Price Indexes
 Log Slope (Mean=0) Log Intercept (Mean=0) Log Slope (Mean=0) Log Intercept (Mean=0)
-.2 0 .2 .4 -4 -2 0 2 4 -1 -.5 0 .5 -4 -2 0 2 4
MYS MYS MEX MEX
IDN IDN BRA BRA
THA THA
PHL PHL CAN CAN
CHN CHN THA THA
TUR TUR CHN CHN
HUN HUN
HKG HKG ARG ARG
ZAF ZAF NLD NLD
PRT PRT
IND IND ISR ISR
POL POL BEL BEL
TWN TWN TWN TWN
KOR KOR
BRA BRA CHE CHE
MEX MEX KOR KOR
NZL NZL
ARG ARG GBR GBR
NOR NOR HKG HKG
ESP ESP ESP ESP
CAN CAN
AUT AUT NOR NOR
AUS AUS IND IND
ITA ITA DEU DEU
NLD NLD
IRL IRL FIN FIN

7 - Machinery
5 - Chemicals

GBR GBR FRA FRA

FIN FIN
FRA FRA AUT AUT
ISR ISR DNK DNK
BEL BEL AUS AUS
JPN JPN
DNK DNK ITA ITA
CHE CHE JPN JPN
DEU DEU
SWE SWE SWE SWE
SGP SGP IRL IRL

Log Slope (Mean=0) Log Intercept (Mean=0) Log Slope (Mean=0) Log Intercept (Mean=0)
-.4 -.2 0 .2 .4 -5 0 5 10 -.05 0 .05 .1 -1.5 -1 -.5 0 .5 1

Figure A.2
SGP SGP HKG HKG
ZAF ZAF THA THA
POL POL CHN CHN
MEX MEX CHL CHL
ROM ROM SGP SGP
CHL CHL MYS MYS
IND IND PHL PHL
PAK PAK IND IND
MAR MAR POL POL
GRC GRC MEX MEX
BRA BRA IDN IDN
COL COL PAK PAK
TUR TUR TUR TUR
ARG ARG COL COL
IDN IDN ARG ARG
NLD NLD BRA BRA
CAN CAN HUN HUN
PHL PHL TWN TWN
BEL BEL GRC GRC
HUN HUN NZL NZL
SWE SWE CAN CAN
ESP ESP ZAF ZAF
AUS AUS KOR KOR
CHN CHN PRT PRT
GBR GBR NLD NLD
MYS MYS AUS AUS
JPN JPN ISR ISR

Note: Intercepts and slopes are in logs and are normalized by their means across countries.
ISR ISR ESP ESP
AUT AUT NOR NOR
8 - Misc Manuf

DEU DEU GBR GBR

DNK DNK JPN JPN
6 - Manf Materials

THA THA ITA ITA

FRA FRA DEU DEU
KOR KOR DNK DNK
FIN FIN FRA FRA
CHE CHE AUT AUT
IRL IRL SWE SWE
PRT PRT BEL BEL
ITA ITA IRL IRL
TWN TWN FIN FIN
HKG HKG CHE CHE
Normalized Quality Index Intercepts and Slopes, By Industry

Normalized Log Quality Index Intercepts and Slopes, by Manufacturing Industry

 Normalized Quality Indexes, 1989-2003
By Manufacturing Industry
China France Germany
-2 0 2
Log Index (Mean=0)

Greece Hungary Japan

-2 0 2

Malaysia Singapore Thailand

-2 0 2

1990 1995 2000 1990 1995 2000 1990 1995 2000

6 - Manuf Materials 7 - Machinery 8 - Misc Manuf

Note: Indexes are in logs and are normalized by the mean across countries in each year.

Figure A.3
Normalized Log Quality Indexes For Select Countries, by Manufacturing Industry
 Normalized Quality Indexes, 1989-2003
All Manufacturing, Average by 1989 Income Group
0 - All Manuf 6 - Manuf Materials
.5 1
Log Norm Index (Mean=0)
-1 -.5 0

7 - Machinery 8 - Misc Manufactures

.5 1
-1 -.5 0

1990 1995 2000 1990 1995 2000

High Income in 1989 Low Income in 1989

Note: Countries are divided into high- and low-income cohorts based on 1989 per capita gdp (PCGDP).

Figure A.4
Evolution of Mean Normalized Quality for Countries with High and Low Income in 1989
 Normalized Quality for SITC 84 versus SITC 65&84, 2003

1
CHE
BEL
MAR
AUT
ROM ITA
FRA
.5

HUN DEU PRT

JPN
84 - Apparel

HKG
TUR
ESP GBR SWE
AUS
IRL
CHN
POL
PHLGRC
0

CANTHA
MYS IDN
CHL

PAK
COL ISR
-.5

IND
KOR
MEX SGP TWN
BRA
ARG
ZAF
-1

-1 -.5 0 .5 1
65&84 - Apparel & Textiles
Note: Indexes are in logs and are normalized by the mean across countries in each year.

Figure A.5
Comparison of 2003 Quality Indexes for Apparel Versus Apparel & Textiles