i. introduction · sigma-convergence, referring to convergence in productivity levels, at least for...

UNCONDITIONAL CONVERGENCE IN MANUFACTURING*

Dani Rodrik

Unlike economies as a whole, manufacturing industries exhibit strong un-conditional convergence in labor productivity. The article documents this atvarious levels of disaggregation for a large sample covering more than 100countries over recent decades. The result is highly robust to changes in thesample and specification. The coefficient of unconditional convergence is esti-mated quite precisely and is large, at between 2–3% in most specifications and2.9% a year in the baseline specification covering 118 countries. The article alsofinds substantial sigma-convergence at the two-digit level for a smaller sampleof countries. Despite strong convergence within manufacturing, aggregate con-vergence fails due to the small share of manufacturing employment inlow-income countries and the slow pace of industrialization. Because of datacoverage, these findings should be as viewed as applying to the organized,formal parts of manufacturing. JEL Codes: O40, O14.

I. Introduction

Neoclassical growth theory establishes a presumption thatcountries with access to identical technologies should convergeto a common income level. Countries that are poorer and havehigher marginal productivity of capital should grow more rapidlyin the transition to the long-run steady state. In an open globaleconomy, access to foreign capital and foreign markets (whichremoves finance and market size as constraints) further strength-ens the presumption of convergence.

However, empirical work has not been kind to this propos-ition. Selected developing countries, such as those in East Asia,have grown quickly. But when poor countries are taken as awhole, there is no systematic tendency for them to grow fasterthan rich ones, over any reasonably long time horizon for whichwe have data.1 Whatever convergence one can find is conditional:It depends on policies, institutions, and other country-specific cir-cumstances. The only clear-cut exceptions to the rule seem to be

*This is a substantially revised version of Rodrik (2011a). I am grateful toUNIDO for making the INDSTAT2 and INDSTAT4 databases available. I alsothank Cynthia Balloch for research assistance; the Weatherhead Center forInternational Affairs and the Center for International Development at HarvardUniversity for financial assistance; and Alberto Abadie, Daron Acemoglu,Jonathan Temple, three referees, and the editors for useful suggestions.

1. On convergence in the decade before the global financial crisis of 2008–2009,see Subramanian (2011, chap. 4), and Rodrik (2011b).

! The Author(s) 2012. Published by Oxford University Press, on behalf of President andFellows of Harvard College. All rights reserved. For Permissions, please email: [email protected] Quarterly Journal of Economics (2013), 165–204. doi:10.1093/qje/qjs047.Advance Access publication on November 18, 2012.

165

states/regions within a unified economy such as the United States(Barro and Sala-i-Martin 1991).2

If growth rates are characterized by conditional instead ofunconditional convergence, economies will tend toward differentlevels of income in the long run. Lack of empirical support for(unconditional) convergence has led theory in the direction ofmodels with endogenous technological change, which do notnecessarily exhibit convergence, and to empirical work thatfocuses on identifying the conditioning variables that makes con-vergence feasible (see Acemoglu 2009 on theory, and Durlauf,Johnson, and Temple 2005 on empirical work).

In contrast to this large literature, I show in this article thatunconditional convergence does exist, but it occurs in the modernparts of the economy rather than the economy as a whole. Inparticular, I document a highly robust tendency toward conver-gence in labor productivity in manufacturing activities, regard-less of geography, policies, or other country-level influences. Thecoefficient of unconditional convergence (beta) is large—2.9% ayear in my baseline specification that covers 118 countries—andestimated quite precisely, with more disaggregated specificationsgenerally yielding somewhat higher estimates. A convergencerate of 2.9% implies that industries that are, say, a tenth of theway to the technology frontier (roughly the bottom 20% of theindustries in the sample) experience a convergence boost intheir labor productivity growth of 6.7 percentage points perannum (0.029� ln(10)).

Figures I, II, and III illustrate the central result of this art-icle and place it in proper perspective. Each figure shows the re-lationship between labor productivity in some base period (on thevertical axis) and its growth rate over the subsequent decade,controlling for period-specific influences. Figure I, where eachdot stands for a particular country in a specific decade, presentsa typical nonconvergence result for country-level productivity.There is no systematic tendency for countries that start withlower productivity (measured here by gross domestic product[GDP] per worker) to grow more rapidly.

2. Some studies also find unconditional convergence among the richerOrganisation for Economic Co-operation and Development (OECD) countries,but it is difficult to know what to make of this result in light of the obvioussample selection bias (Baumol 1986; DeLong 1988).

QUARTERLY JOURNAL OF ECONOMICS166

Figure II depicts the analogous relationship for individualmanufacturing industries. Each dot on this scatter plot repre-sents a two-digit industry in a specific country. (Illustrativeindustries: food and beverages, chemicals and chemical products,motor vehicles.) Each country enters Figure II with multipleindustries (but over a single time horizon, with the most recentdecade for which it has data). Period and industry-specific influ-ences are controlled using industry, decade, and industry�dec-ade fixed effects. Because there are no controls for country-leveldeterminants, Figure II represents a test of unconditional con-vergence similar in spirit to Figure I. (The need for period andindustry fixed effects are motivated subsequently.) The negativeand highly significant slope is unmistakable, illustrating the cen-tral conclusion: Manufacturing exhibits a strong tendency forunconditional convergence. Industries that start at lower levelsof labor productivity experience more rapid growth in labor

FIGURE I

Lack of Convergence in Economy-wide Labor Productivity

Variable on the vertical axis is growth of GDP per worker over four sepa-rate decades (1965–1975, 1975–1985, 1985–1995, 1995–2005), controlling fordecadal fixed effects. Source of data: PWT 7.0. Sample is restricted to countriesincluded in the manufacturing convergence regressions.

UNCONDITIONAL CONVERGENCE IN MANUFACTURING 167

productivity. As I show later, when country fixed effects areincluded the slope becomes even steeper. Conditional conver-gence is more rapid than unconditional convergence. But whatis striking in Figure II is the evident strength of convergence inthe data even in the absence of such controls. Figure III shows thesame convergence result for manufacturing in aggregate, withlabels identifying each country.3

My focus in this article is on what it is calledbeta-convergence, where beta refers to the slope of the relation-ship in Figures I–III. I also present evidence for significant

FIGURE II

Unconditional Unconditionce in 2-digit Manufacturing Sectors

Variable on the vertical axis is the growth of value added per workerin 2-digit manufacturing industries, controlling for period, industry, andperiod� industry fixed effects, where for each country the latest decade overwhich data are available is included. Source of data: INDSTAT2. For furtherdetails on data and methods, see text.

3. There are some apparent anomalies in Figure III arising from compositionaldifferences across countries. For example, Suriname (SUR) shows up as the countrywith the highest labor productivity. This is due to the fact that industry in thiscountry consists almost entirely of alumina and aluminum production, which ishighly capital- and energy-intensive. Such compositional issues provide an import-ant rationale for working with more disaggregated data.


sigma-convergence, referring to convergence in productivitylevels, at least for recent time periods. Even if beta-convergenceholds, countries may fail to converge in levels as long as randomshocks to the growth process are relatively large.4 In the presentsample of industries, beta- and sigma-convergence go together. Ifind that productivity dispersion was sharply reduced acrosscountries during 1995–2005 in the majority of two-digit manufac-turing industries. In manufacturing taken as a whole, sigmashrunk by 10 log-points during this period in the sample of 63countries for which comparison is possible—a reduction of theorder of 10%.

FIGURE III

Unconditional Convergence in Aggregate Manufacturing

Variable on the vertical axis is the growth of value added per worker inaggregate manufacturing controlling for period fixed effects, where for eachcountry the latest decade over which data are available is included. Source ofdata: INDSTAT2. For further details on data and methods, see text.

4. This is true especially if the shocks act in an offsetting manner, butsigma-convergence may fail even if the shocks are purely idiosyncratic (i.e., uncor-related with incomes per capita). Young, Higgins, and Levy (2008) discuss the dif-ference between beta- and sigma-convergence and document the absence ofsigma-convergence for U.S. counties.


I note that my data come from the United Nations IndustrialDevelopment Organization’s (UNIDO) industrial statistics data-base, which is derived largely from industrial surveys. Becausemicroenterprises and informal firms are often excluded from suchsurveys, we cannot be certain that the results are universallyvalid across all types of manufacturing activities. In the absenceof more complete coverage of manufacturing, these findings onconvergence should be as viewed as applying to the organized,formal parts of manufacturing.

To my knowledge, this is the first article to demonstrate un-conditional convergence in industry for a wide range of countriesand for detailed manufacturing industries. There does not seemto be any work that has looked at highly disaggregated data formanufacturing or at the manufacturing experience of countriesbeyond Organisation for Economic Co-operation and Develop-ment (OECD) and U.S. states (Bernard and Jones 1996a.1996b; see also Sørensen 2001). However, two recent related stu-dies deserve mention. In unpublished work, Hwang (2007) hasdocumented that there is a tendency for unconditional conver-gence in export unit values in highly disaggregated productlines. Once a country begins to export something, it travels upthe value chain in that product regardless of domestic policies orinstitutions.5 Hwang shows that the lower the average unitvalues of a country’s manufactured exports, the faster the coun-try’s subsequent growth, unconditionally. This article differsfrom Hwang in that it focuses on output rather than exportsand directly on productivity (rather than unit values). Conver-gence seems to kick in manufacturing regardless of whether pro-duction is exported.6 In addition, a recent paper by Benetrix,O’Rourke, and Williamson (2012) has documented convergencein industrial output since the late nineteenth century, with thestrongest results obtaining for the 1890–1972 period. Benetrixand colleagues focus on growth rates of aggregate industrial

5. Hwang demonstrates his result for both 10-digit U.S. HS import statisticsand 4-digit SITC world trade statistics. The first classification contains thousandsof separate product lines.

6. Also related is a paper by Levchenko and Zhang (2011), which estimatesmodel-based relative productivity trends for 19 manufacturing industries from the1960s through the 2000 and show that there has been steady convergence acrosscountries.


output rather than on disaggregated industrial productivity,which is the focus here.7

These results naturally raise the question of why conver-gence within manufacturing fails to aggregate up to the economyas a whole. Another contribution of this article is to reconcileunconditional convergence in manufacturing with its apparentabsence for GDP per worker. I analyze this question by combiningUNIDO data with Penn World Tables to derive implied laborproductivity estimates for the rest of the economy. With thesedata, I document three additional features of the data. First, non-manufacturing does not exhibit convergence. Second, manufac-turing’s impact on aggregate convergence is curtailed by its smallsize, especially in the poorer countries. Third, the growth boostfrom reallocation—the shift of labor to more productive manufac-turing—is not sufficiently and systematically greater in poorereconomies. Taken together, these three facts account for the ab-sence of aggregate convergence.

The analysis highlights the role of structural factors, in par-ticular the slow (and sometimes perverse) movement of resourcesacross economic activities with different convergence character-istics. The trouble from a convergence standpoint is thateconomic activities that are good at absorbing advanced technol-ogies are not necessarily good at absorbing labor. As a result, toomuch of an economy’s resources can get stuck in the ‘‘wrong’’sectors—those that are not on the escalator up. When firmsthat are part of international production networks or otherwisebenefit from globalization employ little labor, the gains remainlimited. Even worse, intersectoral labor flows can be perversewith the consequence that convergence within the ‘‘advanced’’sectors is accompanied by divergence on the part of the economyas a whole. McMillan and Rodrik (2011) illustrate some of theseperverse outcomes using the experience of specific countries.

The article proceeds as follows. Section II describes the dataand methods used for the estimation. Section III presents thebasic results and various robustness checks. Section IV considersthe reasons for convergence failing to aggregate up to the level ofthe entire economy. Section V provides concluding remarks.

7. The first version of this article (Rodrik 2011a) was completed before Ibecame aware of Benetrix, O’Rourke, and Williamson’s (2012) work on the subject.


II. Data and Methods

II.A. Date Source and Description

I use data from UNIDO’s INDSTAT2 and INDSTAT4 data-bases, which provide industrial statistics for a wide range of coun-tries at different levels of disaggregation (UNIDO 2011, 2012).My baseline sample is based on INDSTAT2, which has goodcoverage of countries at the ISIC two-digit level going back tothe 1960s. These statistics cover value added and employment,among others, for up to 23 manufacturing industries per country,allowing me to calculate labor productivity (value added per em-ployee) and its growth at that level of disaggregation. INDSTAT4provides more disaggregated data at the four-digit level for up to127 industries, but covers fewer countries and is spotty for earlieryears, making it impractical to work with it for periods thatextend before 1990. An earlier version of this article reportedsubstantially similar results using INDSTAT4 (Rodrik 2011a).Here, I rely mostly on INDSTAT2, which has the advantage ofallowing me to increase the country coverage as well as presentresults for before 1990.8 Results using INDSTAT4 serve to pro-vide robustness checks.

Even with INDSTAT2 I am hampered by the fact that fewcountries have long time series, with developing countries posinga particular challenge. The longer the time span we choose for theconvergence horizon, the smaller the number of countries thatcan be included. This makes any horizon longer than a decadeimpractical. Accordingly, the empirical analysis is based on dec-adal growth. Moreover, there is limited overlap across countriesfor any particular decade, so that choosing a fixed time span—say, 1995 to 2005—reduces the sample considerably, by half ormore.

In light of these limitations, my baseline consists of a pooledsample that combines the latest 10-year period for each countrywith data. The advantage of pooling is that it maximizes the num-ber of countries that can be included, yielding a sample of 118countries in all and allowing good coverage of poorer economies.

8. INDSTAT2 was released after Rodrik (2011a) was completed. I thank areferee for bringing it to my attention.


Because each country enters with around 20 industries, the totalnumber of observations in the baseline specification is greaterthan 2,000. I supplement this sample with two others to checkfor robustness and carry out further analyses:

(1) A panel, which stacks the four decades 1965–75,1975–85, 1985–95, and 1995–2005 for all countrieswith data for at least one of those decades. The panelsample covers a relatively large number of countries(99 in all). Its disadvantage is that it is highly unba-lanced, with developed countries having much bettercoverage than developing ones. For example, Japanenters the panel with 71 industries, whereas Algeriahas merely 12.

(2) A pure cross-section for 1995–2005. This sample hasthe fewest countries covered (58), for reasons alreadyexplained. I rely on the cross-section sample when Irequire a consistent comparison across countriesover a fixed time period (as when I analyze sigma-convergence or examine the failure of aggregation).

Some countries in INDSTAT2 have data for aggregate man-ufacturing, but lack disaggregated data. Thus the samplesbecome slightly larger than what was just noted when I analyzemanufacturing as a whole.

As mentioned in Section I, UNIDO’s data come from indus-trial surveys whose coverage differs across countries. Data fordeveloped countries refer for the most part to ‘‘all establish-ments.’’ But in developing countries, enterprises with fewerthan 5 or 10 employees are often not included. For this reason,the convergence results that follow should be read as applying tothe more formal, organized parts of manufacturing and not tomicro-enterprises or informal firms. An appendix (available onrequest) provides a summary of data coverage for each countryincluded in the regressions.

An important problem with the data is that INDSTAT pro-vides figures for value added in nominal U.S. dollars. What I needis a measure of growth in labor productivity in real terms. I canrecover the convergence parameter under certain assumptionsabout the process followed by U.S. dollar inflation rate for disag-gregated manufacturing industries. In what follows, I explain myapproach in greater detail.


II.B. Empirical Specification

Dividing nominal US$ value added by employment, I calcu-late nominal labor productivity for each industry vijt, where i de-notes the industry, j the country, and t the time period. The rate ofgrowth of labor productivity in real terms, yijt, is given in turn byyijt ¼ vijt � �ijt, where �ijt is the increase in the industry-level de-flator in dollar terms and a hat over a variable denotes percentchanges.

I assume (real) labor productivity growth in each industry isa function of both country-specific conditions and a convergenceeffect. The latter, in turn, is proportional to the gap between eachindustry’s initial productivity and its frontier technology, repre-sented by v�it. Hence:

yijt ¼ �ð ln v�it � ln vijtÞþDj,

where Dj is a dummy variable that stands in for all time- andindustry-invariant country-specific factors. The convergence co-efficient we are interested in estimating is �. Note that if ln vijt ismeasured with error, this specification potentially introduces abias toward overestimating the rate of convergence, since such anerror weakens the link between initial productivity and finalproductivity. This is a common problem in the empirical litera-ture on convergence (Temple 1998).

The last step is to specify a process for prices. I assume acommon global (U.S. dollar) inflation rate for each individual in-dustry up to an idiosyncratic (random) error term, such that�ijt ¼ �it þ "ijt. This is a reasonable assumption because manufac-tures are tradable and face common world prices. Of course, inpractice there are many reasons domestic prices may divergefrom world prices, even in tradables. Transport costs or tradepolicies such as import tariffs and export subsidies drivewedges between domestic and foreign prices. But these wedgesintroduce differences in price levels, not inflation rates.Equivalently, I assume that dollar inflation rates are not system-atically correlated with an industry’s distance from the techno-logical frontier. My results do not overstate convergence as longas dollar price inflation is not systematically higher in industriesthat are furthest away from the technological frontier. I postponefurther discussion of these issues to the next section, where Idiscuss potential complications arising from the absence ofprice information more extensively.


This allows me to express the growth of nominal labor prod-uctivity as follows:

vijt ¼ ��ðln vijt � ln v�itÞ þ �it þDj þ "ijt:ð1Þ

I assume "ijt is uncorrelated with other explanatory variablesand captures all other idiosyncratic influences on measured laborproductivity. Rearranging terms, I now have the final estimatingequation:

vijt ¼ �� ln vijt þ ð�it þ � ln v�itÞ þDj þ "ijt:ð2Þ

This can be expressed equivalently as

vijt ¼ �� ln vijt þDit þDj þ "ijt,ð3Þ

where Dit stands for ð�it þ � ln v�itÞ. Hence, I can regress thegrowth of labor productivity in nominal U.S. dollar terms onthe initial level of labor productivity, a set of industry� timeperiod fixed effects (Dit), and country fixed effects (Dj). In theregressions that follow, I include separate industry and perioddummies to soak up any additional confounding residual system-atic variation.

It is also possible to run this regression over a single timeperiod as a pure cross-section. In this case, the industry� timeperiod fixed effects are reduced to industry fixed effects:

vij ¼ �� ln vij þDi þDj þ "ij:ð30Þ

As specified, the estimate of � will be a measure of condi-tional convergence, since country-specific conditions are expli-citly controlled for by the country fixed effects. A test ofunconditional convergence consists of dropping these countrydummies and checking whether the estimated coefficient –� re-mains negative and statistically significant.

III. Empirical Results

III.A. Basic Results

Table I shows the results for the baseline specification alongwith its variants. The dependent variable in each case is the (com-pound annual) growth rate of labor productivity for two-digitmanufacturing industries. The regressors are the log of initial


TA

BL

EI

BA

SE

LIN

ES

PE

CIF

ICA

TIO

NA

ND

VA

RIA

TIO

NS

(DE

PE

ND

EN

TV

AR

IAB

LE:

GR

OW

TH

OF

LA

BO

RP

RO

DU

CT

IVIT

YIN

TW

O-D

IGIT

MA

NU

FA

CT

UR

ING

IND

US

TR

IES

OV

ER

AD

EC

AD

E)

Base

lin

eB

ase

lin

e,on

lyp

ost-

1990

data

Cro

ss-s

ecti

on,

1995–2005

Pan

el

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Log

init

ial

pro

du

ctiv

ity�

0.0

29**

*�

0.0

54**

*�

0.0

29**

*�

0.0

57**

*�

0.0

24**

*�

0.0

59**

*�

0.0

18**

*�

0.0

60**

*�

0.0

19**

*�

0.0

55**

*

(0.0

04)

(0.0

06)

(0.0

04)

(0.0

06)

(0.0

04)

(0.0

10)

(0.0

03)

(0.0

05)

(0.0

06)

(0.0

07)

Log

init

ial

pro

du

ctiv

ity,

1975–85

0.0

00

�0.0

09

(0.0

06)

(0.0

06)

Log

init

ial

pro

du

ctiv

ity,

1985–95

0.0

16*

0.0

08

(0.0

09)

(0.0

07)

Log

init

ial

pro

du

ctiv

ity,

1995–2005

�0.0

06

�0.0

11

(0.0

07)

(0.0

09)

Cou

ntr

yfi

xed

effe

cts

No

Yes

No

Yes

No

Yes

No

Yes

No

Yes

Ind

ust

ryfi

xed

effe

cts

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Per

iod

fixed

effe

cts

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Per

iod�

ind

ust

ry

fixed

effe

cts

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Nu

mber

ofco

un

trie

s118

118

104

104

58

58

99

99

99

99

Nu

mber

of

obse

rvati

ons

2,1

22

2,1

22

1,8

61

1,8

61

955

955

3,8

48

3,8

48

3,8

48

3,8

48

Not

es.

Th

ebase

lin

esp

ecifi

cati

onu

ses

all

cou

ntr

ies

wit

hth

ere

qu

isit

ed

ata

,u

sin

gon

lyth

ela

test

10-y

ear

per

iod

for

each

.T

he

pan

elsp

ecifi

cati

onp

ools

data

for

1965–75,

1975–85,

1985–95,

an

d1995–2005

for

all

cou

ntr

ies

wit

hre

qu

isit

ed

ata

.S

tan

dard

erro

rsare

clu

ster

edat

cou

ntr

yle

vel

.**

*p<

.01,

**p<

.05,

*p<

.10.


labor productivity and a host of fixed effects, depending on thespecification. Each regression is run first without and then withcountry dummies. As explained previously, these two specifica-tions yield the unconditional and conditional convergence coeffi-cients, respectively. Standard errors are clustered at the countrylevel in all specifications.

Column (1) is the baseline result, corresponding to the scat-ter plot displayed in Figure II. The estimated convergence coeffi-cient of 2.9% is highly significant, with a t-statistic of 6.95. Recallthat the baseline sample pools different decades for each country.There is little evidence of parameter heterogeneity across differ-ent time periods, however. I get quite similar estimates when Idrop pre-1990 observations from the baseline specification(column (3)) or when I run a pure cross-section for the morerecent period 1995–2005 (column (5)).9 I take note of the factthat the 1995–2005 results do not diverge much from the baselineas I have to focus on the cross-section for some of the subsequentanalysis.

The panel specification, which combines the four decades1965–75, 1975–85, 1985–95, and 1995–2005, yields a somewhatsmaller � of 1.8%, but this estimate is still highly significant witha t-statistic of 6.74 (column (7)). In column (9) I test explicitly fordifferences in the convergence coefficient across time periods byinteracting base labor productivity with decade dummies. Theexcluded decade is 1965–75. The results suggest a lower conver-gence estimate for 1985–95 but otherwise no strongly discernibletrends over time. In particular, there is no evidence of strongerconvergence in more recent decades. This is somewhat surprisingbecause one might have expected globalization and the spread ofglobal production networks to greatly facilitate technologicaldissemination and therefore catch-up. The result suggests con-vergence is an intrinsic property of manufacturing industries,and one that is not driven by the ups and downs of global eco-nomic integration.

Each specification in Table I is paired with its conditionalvariant, which includes country fixed effects. The estimated con-vergence coefficients always increase in size, typically doubling—or tripling, in the panel specification—when country dummies

9. The cross-section results for the latest decade for which I have data,1997–2007, are very similar to those for 1995–2005 (not shown). I present the1995–2005 results because of the somewhat larger country coverage.


are included. This is in line with the conditional convergence re-sults in the literature. As noted recently by Barro (2012), thereare reasons to think growth regressions with country fixed effectsyield upwardly biased estimates of the convergence rate when thetime horizon is short. This is due to the so-called Hurwicz-Nickellbias: the estimate of the coefficient on the lagged dependent vari-able is biased downwards in the presence of a fixed effect(Hurwicz 1950; Nickell 1981). The bias tends to zero as the timespan gets large, but can be large in short panels.10 For my pur-poses, because I am mainly interested in unconditional conver-gence (and hence the specifications without country fixed effects),it suffices to treat the conditional convergence estimates as upperbounds. They still provide a useful reference point for gauging themagnitude of the unconditional convergence rates.

It is to be expected that country-specific conditions that arecorrelated with initial productivity—policies, institutions, geog-raphy—play a role in determining the speed of catch-up. So evenin the absence of the Hurwicz-Nickell bias, it is natural that thebeta coefficient would become larger when fixed effects are intro-duced. What is surprising is the evidence for systematic and rapidproductivity convergence in individual manufacturing industrieswhen these country-specific conditions are not controlled forthrough country dummies. Once again, there is no evidence ofsignificant differences in convergence rates across time periods.In particular, the speed of (conditional) convergence does notappear to have increased in recent decades (column (10)).

In Table II, I compare the baseline with results at differentlevels of aggregation. The table shows specifications that are bothmore disaggregated (three- and four-digit, using INDSTAT4data) and more aggregated (manufacturing as a whole). The re-sults in column (7) correspond to Figure III. Each regression isrun in two versions, one with the largest possible sample andanother one that restricts the sample to a common list of coun-tries with the requisite data at all levels of aggregation. The latterfacilitates direct comparison across different levels of aggrega-tion. The estimated �s are statistically significant, typically at

10. In principle, the same bias exists when industry fixed effects are includedtoo, as in my baseline specification. But the dimensionality is much greater in thiscase since the de-meaning of the dependent variable takes place over both timeperiods and across countries. In any case, I also provide pure cross-section esti-mates for each industry separately, using no fixed effects.


TA

BL

EII

CO

MP

AR

ISO

NO

FR

ES

UL

TS

AT

DIF

FE

RE

NT

LE

VE

LS

OF

AG

GR

EG

AT

ION

(DE

PE

ND

EN

TV

AR

IAB

LE:

GR

OW

TH

OF

LA

BO

RP

RO

DU

CT

IVIT

YIN

MA

NU

FA

CT

UR

ING

IND

US

TR

IES

OV

ER

AD

EC

AD

E)

Base

lin

e(I

SIC

2-d

igit

)IS

IC3-d

igit

ISIC

4-d

igit

All

man

ufa

ctu

rin

g

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

All

Com

mon

sam

ple

All

Com

mon

sam

ple

All

Com

mon

sam

ple

All

Com

mon

sam

ple

Log

init

ial

pro

du

ctiv

ity

�0.0

29**

*�

0.0

28**

*�

0.0

29**

*�

0.0

30**

*�

0.0

31**

*�

0.0

31**

*�

0.0

30**

*�

0.0

23**

(0.0

04)

(0.0

08)

(0.0

06)

(0.0

06)

(0.0

04)

(0.0

04)

(0.0

06)

(0.0

09)

Cou

ntr

yfi

xed

effe

cts

No

No

No

No

No

No

No

—In

du

stry

fixed

effe

cts

Yes

Yes

Yes

Yes

Yes

Yes

No

—P

erio

dfi

xed

effe

cts

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Per

iod�

ind

ust

ryfi

xed

effe

cts

Yes

Yes

Yes

Yes

Yes

Yes

No

—N

um

ber

ofco

un

trie

s118

38

39

38

39

38

123

38

Nu

mber

ofob

serv

ati

ons

2,1

22

812

704

703

2,8

71

2,8

69

123

38

Not

es.

Sta

nd

ard

erro

rsare

clu

ster

edat

cou

ntr

yle

vel

,ex

cep

tfo

r‘‘a

llm

an

ufa

ctu

rin

g,’’

for

wh

ich

robu

stst

an

dard

erro

rsare

rep

orte

d.

Com

mon

sam

ple

refe

rsto

the

set

ofco

un

trie

sth

at

have

data

at

all

level

sof

aggre

gati

onsh

owed

inth

ista

ble

,fo

rth

esa

me

per

iod

.**

*p<

.01,

**p<

.05,

*p<

.10.


the 99% confidence level, at all levels of aggregation. The esti-mates at the three- and four-digit level are remarkably similar tothe baseline estimates, as are the estimates for manufacturing asa whole. Looking across the regressions with a common sample,one can see some indication that the estimated coefficient in-creases with the level of disaggregation—from 2.3% for manufac-turing as a whole to 3.1% at the four-digit level. But thesedifferences are not statistically significant. The similarity in themagnitude of the estimates suggests that the failure of uncondi-tional convergence to aggregate up to the economy as a whole isnot a result of compositional issues within manufacturing. I makeuse of this result later in the article.

III.B. Robustness

Table III presents a battery of additional robustness testsaround the baseline specification. Panel A prunes the baselinesample in a number of ways to ensure that countries/industrieswith specific characteristics are not driving the results. I excludein turn (1) countries with fewer than 10 industries; (2) observa-tions that correspond to the highest and lowest 10% values forgrowth; (3) observations in the top and bottom half, respectively,of the sample in terms of labor productivity; (4) former socialistcountries; and (5) OECD countries.

Remarkably, the convergence estimates remains highly sig-nificant across all these runs. They vary from a low of 1.1% whengrowth observations at the top and bottom deciles are excluded(column (3)) to a high of 6.0% for the low-productivity half of thesample (column (4)). Note that the result becomes, if anything,stronger when OECD countries are excluded (column (7)), withthe convergence coefficient rising from 2.9% to 3.8%. This is sig-nificant, because unconditional convergence has never been docu-mented outside the OECD.

Panel B of Table III carries out different types of robustnesstests, including weighting observations by value added, instru-menting initial labor productivity by lagged productivity (a checkagainst measurement error), recalculating growth rates by esti-mating a log-linear trend using all 10 annual observations(instead of just endpoints),11 and clustering standard errors by

11. This guards against introducing a spurious bias that may arise from laggedlabor productivity appearing directly on both sides of the regression equation, withopposite signs.


TA

BL

EII

I

RO

BU

ST

NE

SS

TE

ST

S(D

EP

EN

DE

NT

VA

RIA

BL

E:

GR

OW

TH

RA

TE

OF

PR

OD

UC

TIV

ITY

OV

ER

RE

LE

VA

NT

PE

RIO

D)

Pan

elA

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Base

lin

e

Excl

ud

ing

cou

ntr

ies

wit

hfe

wer

than

10

ind

ust

ries

incl

ud

ed

Excl

ud

ing

obse

rvati

ons

wit

hth

eh

igh

est

an

dlo

wes

t10%

valu

esfo

rgro

wth

Su

bsa

mp

lew

ith

labor

pro

du

ctiv

ity

inth

ebot

tom

half

ofth

esa

mp

le

Su

bsa

mp

lew

ith

labor

pro

du

ctiv

ity

inth

eto

ph

alf

ofth

esa

mp

le

Excl

ud

ing

form

erso

ciali

stco

un

trie

s

Excl

ud

ing

OE

CD

cou

ntr

ies

Log

init

ial

pro

du

ctiv

ity

�0.0

29**

*�

0.0

27**

*�

0.0

11**

*�

0.0

60**

*�

0.0

43**

*�

0.0

25**

*�

0.0

38**

*(0

.004)

(0.0

04)

(0.0

03)

(0.0

06)

(0.0

05)

(0.0

04)

(0.0

06)

Cou

ntr

yfi

xed

effe

cts

no

no

no

no

no

no

no

Ind

ust

ryfi

xed

effe

cts

yes

yes

yes

yes

yes

yes

yes

Per

iod

fixed

effe

cts

yes

yes

yes

yes

yes

yes

yes

Per

iod�

ind

ust

ryfi

xed

effe

cts

yes

yes

yes

yes

yes

yes

yes

Nu

mber

ofco

un

trie

s118

109

116

95

108

112

91

Nu

mber

ofob

serv

ati

ons

2,1

22

2,0

58

1,6

97

1,0

61

1,0

61

1,9

97

1,5

42


TA

BL

EII

I

(CO

NT

INU

ED)

Pan

elB

(8)

(9)

(10)

(11)

(12)

(13)

(14)

Wei

gh

ted

regre

ssio

nre

sult

s:ln

(valu

ead

ded

)w

eigh

ts

IVre

sult

sw

ith

5-y

ear

lagged

init

ial

pro

du

ctiv

ity

as

inst

rum

ent

Res

cale

dla

bor

pro

du

ctiv

ity

gro

wth

,ad

just

edfo

rre

al

exch

an

ge

rate

ap

pre

ciati

on

Sta

nd

ard

erro

rscl

ust

ered

by

ind

ust

ry

Cou

ntr

yfi

xed

effe

cts

wit

hn

oin

du

stry

con

trol

s

Gro

wth

rate

sca

lcu

late

dfr

omlo

g-l

inea

rtr

end

usi

ng

10

an

nu

al

obse

rvati

ons

Tes

tin

gfo

rn

onli

nea

rity

by

labor

pro

du

ctiv

ity

qu

art

iles

Log

init

ial

pro

du

ctiv

ity

�0.0

30**

*�

0.0

24**

*�

0.0

30**

*�

0.0

29**

*�

0.0

21**

*�

0.0

17**

*(0

.004)

(0.0

05)

(0.0

04)

(0.0

03)

(0.0

03)

(0.0

06)

Log

init

ial

pro

du

ctiv

ity,

low

est

qu

art

ile

�0.0

34**

*(0

.007)

Log

init

ial

pro

du

ctiv

ity,

seco

nd

qu

art

ile

�0.0

35**

*(0

.006)

Log

init

ial

pro

du

ctiv

ity,

thir

dqu

art

ile

�0.0

34**

*(0

.006)

Log

init

ial

pro

du

ctiv

ity,

fou

rth

qu

art

ile

�0.0

32**

*(0

.005)

Cou

ntr

yfi

xed

effe

cts

no

no

no

no

yes

no

no

Ind

ust

ryfi

xed

effe

cts

yes

yes

yes

yes

no

yes

yes

Per

iod

fixed

effe

cts

yes

yes

yes

yes

yes

yes

yes

Per

iod�

ind

ust

ryfi

xed

effe

cts

yes

yes

yes

yes

no

yes

yes

Nu

mber

ofco

un

trie

s118

104

102

23

118

107

118

Nu

mber

ofob

serv

ati

ons

2,1

22

1,6

15

1,7

54

2,1

22

2,1

22

1,7

74

2,1

22

Not

es.

Sta

nd

ard

erro

rsare

clu

ster

edat

cou

ntr

yle

vel

,u

nle

ssot

her

wis

en

oted

.**

*p<

.01,

**p<

.05,

*p<

.10.


industry rather than country. In all these runs, the resultsremain highly significant. Note in particular that the estimatedcoefficient remains virtually unchanged when I weight industriesby size (column (8)).12 Hence, the convergence result is not drivenby the experience of relatively small industries. Instrumentingfor initial productivity reduces the estimated � marginally to2.4% (column (9)).

The last column of Table III tests directly for nonlinearity of �by allowing the estimated coefficient to vary by labor productivityquartile. The results do not suggest any nonlinearity—at least atthe two-digit level (column (14)). There is greater evidence ofnonlinearity in the four-digit data (INDSTAT4), as I reported inthe earlier version of this article (Rodrik 2011a).13

Another way to examine the data is to scrutinize the conver-gence evidence on an industry-by-industry basis. This exercisenot only serves as a robustness test but is interesting in its ownright. I begin with a few scatter plots for individual industries.Figure IV shows convergence plots for the four largest industriesin the sample that also have good country coverage: ISIC 15 (foodand beverages), 24 (chemicals and chemical products), 29 (ma-chinery and equipment, n.e.c.), and 34 (motor vehicles). (In keep-ing with the baseline specification, each country enters thesescatterplots with the latest period for which it has data. I con-tinue to control for time-specific inflation trends by includingperiod dummies.) Note that these plots use data from just thespecified industries. As the negative slopes indicate, countriesthat started further behind tended to experience more rapid prod-uctivity growth in all four industries. The relationship is statis-tically significant at the 99% level in all four cases, withestimated coefficients ranging from 2.2% (ISIC 34) to 2.8%(ISIC 15).

The individual convergence coefficients estimated on anindustry-by-industry basis for each of our two-digit industriesare shown in Table IV. I regress, separately for each industry,the growth rate of an industry’s labor productivity against its

12. Using employment weights produces nearly identical results.13. In Rodrik (2011a) I also reported some evidence that industries differ in

their � coefficients. These differences are much less evident in the two-digit dataI use here. The only finding of note is that textiles and clothing show a mildly lowerconvergence coefficient. These results, not shown here, are available on request.


initial level across all countries in the sample that have therequisite data for the 1995–2005 period:

vij ¼ ��i ln vij þ "ij for i ¼ 1; ::: ; I:

This entails running as many regressions (23) as I have man-ufacturing industries. This is the direct analogue of runningcross-country growth regressions. Note that these specificationsdo not contain industry, period, or any other fixed effects. Many ofthese regressions cover 30–40 countries, so one should not be toodemanding in terms of statistical significance for industry-specific estimates. Nevertheless, the results are quite strong.Among the 23 industries, I find that 18 have statistically signifi-cant convergence coefficients, ranging from a low of 1.4% (tex-tiles) to a high of 5.5% (office, accounting, and computingmachinery). All but one of the industries exhibit unconditionalconvergence, with wearing apparel the sole exception.

FIGURE IV

Convergence in Specific Industries

Vertical axis represents relative growth rate of labor productivity, control-ling for period fixed effects. Each country enters with most recent decade forwhich data are available.


TA

BL

EIV

BE

TA-C

ON

VE

RG

EN

CE

CO

EF

FIC

IEN

TS

BY

IND

US

TR

Y,

1995–2005

RE

GR

ES

SIO

NS

ISIC

cod

eS

ecto

rN

um

ber

ofco

un

trie

sB

eta

coef

fici

ent

Sig

nifi

can

ce

DT

otal

man

ufa

ctu

rin

g63

�.0

20

***

15

Foo

dan

dbev

erages

43

�.0

26

***

16

Tob

acc

op

rod

uct

s31

�.0

20

**17

Tex

tile

s54

�.0

14

*18

Wea

rin

gap

pare

l,fu

r34

.005

n.s

.19

Lea

ther

,le

ath

erp

rod

uct

s,an

dfo

otw

ear

34

�.0

01

n.s

.20

Woo

dp

rod

uct

s(e

xcl

.fu

rnit

ure

)53

�.0

33

***

21

Pap

eran

dp

ap

erp

rod

uct

s52

�.0

24

***

22

Pri

nti

ng

an

dp

ubli

shin

g53

�.0

21

***

23

Cok

e,re

fin

edp

etro

leu

mp

rod

uct

s,n

ucl

ear

fuel

38

�.0

26

**24

Ch

emic

als

an

dch

emic

al

pro

du

cts

52

�.0

10

n.s

.25

Ru

bber

an

dp

last

ics

pro

du

cts

53

�.0

19

***

26

Non

met

all

icm

iner

al

pro

du

cts

56

�.0

32

***

27

Basi

cm

etals

50

�.0

46

***

28

Fabri

cate

dm

etal

pro

du

cts

53

�.0

22

**29

Mach

iner

yan

deq

uip

men

tn

.e.c

.37

�.0

11

n.s

.30

Offi

ce,

acc

oun

tin

g,

an

dco

mp

uti

ng

mach

iner

y23

�.0

55

***

31

Ele

ctri

cal

mach

iner

yan

dap

para

tus

32

�.0

23

***

32

Rad

io,

tele

vis

ion

,an

dco

mm

un

icati

oneq

uip

men

t26

�.0

43

***

33

Med

ical,

pre

cisi

on,

an

dop

tica

lin

stru

men

ts45

�.0

31

***

34

Mot

orveh

icle

s,tr

ail

ers,

sem

i-tr

ail

ers

35

�.0

19

*35

Oth

ertr

an

spor

teq

uip

men

t27

�.0

30

**36

Fu

rnit

ure

;m

an

ufa

ctu

rin

gn

.e.c

.52

�.0

31

**37

Rec

ycl

ing

22

�.0

19

n.s

.

Not

es.

Th

ese

coef

fici

ents

are

obta

ined

by

run

nin

gp

ure

cros

s-se

ctio

nre

gre

ssio

ns

for

each

ind

ust

ryse

para

tely

.**

*p<

.01,

**p<

.05,

*p<

.10,

n.s

.=n

otsi

gn

ifica

nt.


For an even more disaggregated analysis, I turn toINDSTAT4, which breaks down manufacturing into 127 separatesubsectors. This allows a more fine-tuned scrutiny of the conver-gence experience of individual industries, albeit for a reducedcountry sample. As I have shown in Rodrik (2011a, table 2),four-digit data yield highly significant global convergence coeffi-cients when all four-digit industries are pooled together, even forperiods that are as short as five years. Here I focus on the resultsof individual industry regressions for 2000–2005, because thismaximizes the number of countries (and hence observations).Once again, these are simple cross-section regressions for eachindustry separately, with no fixed effects.

Figure V summarizes the results by showing the distributionof estimated coefficients on initial productivity across the 127industries, normalized by their standard errors. The vast major-ity of the estimates are negative. Though not all of them are stat-istically significant, a surprising number of the negative ones are.Specifically, 79 (out of 127) of the industry regressions yield nega-tive and statistically significant coefficients (at the 95% level orhigher). By contrast, none of the (few) positive coefficients arestatistically significant.

There is a statistical sense in which the global estimate,which pools across industries and controls for industry fixedeffects, represents a weighted average of these industry-by-industry convergence estimates. Specifically, consider a givencross-section (such as 2000–2005 in the foregoing). Omittingtime subscripts, the relationship between estimates � and �i isgiven by:

� ¼XI

i¼1

�i

var ln vijjJ¼ i� �

PrðJ ¼ iÞPIl¼1 var ln vijjJ ¼ l

� �PrðJ ¼ lÞ

!,

where J denotes an industry-identifying variable. Thus, theweight that each industry estimate gets in the global estimatedepends on the variance of the independent variable ln vij

within the industry as well as the relative number of occurrencesof that industry in the global sample.14

14. The derivation of this expression, for which I am indebted to AlbertoAbadie, is available on request.


III.C. Complications Arising from Price Effects

One possible concern in interpreting these results is that myassumption of a common value added deflator (in dollars) for eachindustry, regardless of the country where it is located, may beintroducing a bias to the estimation. I justified this assumptionpreviously by arguing that the manufacturing industries in ques-tion are tradable, and hence face common world prices. Of course,I do not expect the law of one price to obtain perfectly, even forhomogeneous or standardized goods. Tariffs, subsidies, nontariffbarriers, and transport costs often drive a wedge between domesticand world prices. Such wedges may well be larger in the poorercountries. My estimation strategy can accommodate such devi-ations provided they do not vary over time in a way that is corre-lated systematically with distance from the technological frontier.

Suppose, for example, the ad valorem equivalent of tradecosts is �ijt. Then �ijt ¼ �

�it þ ð1þ �ijtÞ

�1d�ijt, where ��it is the

FIGURE V

Distribution of Convergence Coefficients, by 4-digit Industry (Normalized byStandard Error)

The figure is based on convergence coefficients generated from 127 industryconvergence regressions across 4-digit manufacturing industries. See text.


percent change in world prices. All I need for unbiased estimationis for ð1þ �ijtÞ

�1d�ijt to be uncorrelated with initial labor product-ivity ln vijt, conditional on a set of industry, period, and indus-try�period fixed effects. It is hard to think of strong reasons apriori as to why there should be such a pattern in trade costchanges.

One potential source of bias is systematic changes in realexchange rates. In principle, across-the-board increases in do-mestic costs such as wages should be offset, on average, by depre-ciation of the home currency, leaving dollar values generallyunchanged. But in countries that experience sustained move-ments in their real exchange rate, trends in value added ex-pressed in U.S. dollars will be misleading with regard toproductivity in individual manufacturing industries. The worstcase, from the perspective of the present article, would be if thelow-income countries that house a preponderant share oflow-productivity industries were the ones to experience real ex-change rate appreciations—a rise in domestic costs not compen-sated by currency depreciation. This would lead to an upwardbias in my convergence estimates.

To check against this possibility, Table III provides a versionof the convergence regression that explicitly ‘‘corrects’’ for realexchange rate changes. I rescaled the growth in value addedper worker by deflating it with (one plus) the rate of appreciationof the country’s real exchange rate.15 This reduces—across theboard—the measured productivity growth rates of industries incountries which have experienced real appreciations, while rais-ing them in countries with real depreciations. I rerun the baselinespecification using these adjusted values for the dependent vari-able (Table III, column (10)).

If observed convergence were due to real appreciation in thepoorer countries, the resulting estimates would be substantiallylower. In fact, the estimated � is actually slightly higher (.030)and statistically equally significant. (Note, however, that thesample size is somewhat reduced as the lack of price data forsome of the countries prevents me from computing real exchange

15. These are conventional bilateral real exchange rates vis-a-vis the UnitedStates. Domestic inflation rates have been calculated using producer-price indiceswhere possible, substituting the Consumer Price Index where the Producer PriceIndex is not available. The source for the data on exchange rates and price indices isthe International Monetary Fund’s International Financial Statistics. A few coun-tries had to be dropped because of lack of price data.


rate changes for them.) The bottom line is that there is no evi-dence that real exchange rate movements have distorted my basicfindings.

Another source of possible bias arises from compositionalchanges. Even four-digit industries are a mix of different activ-ities, and what appears as an increase in dollar values may be inreality a shift toward higher value-added activities within thesame industry. Equivalently, producers may be moving towardhigher-quality varieties, so that what looks like productivitygrowth is really quality upgrading (Schott 2004; Hwang 2007).It is possible that industries that start further away from thefrontier experience such shifts more rapidly. In the presence ofprice data, I would have been able to capture such changesthrough their effect on industry-specific prices.

Such biases are potentially of concern, but my results providesome comfort that they are not very important in practice. Asseen in Table II, the estimated speed of convergence changesvery little across two-, three-, and four-digit levels of aggregation.If anything, the estimate increases in value as I disaggregatefurther. If the observed productivity growth was due to a shifttoward higher quality products, the opposite result would hold.

Furthermore, even if the compositional biases were quanti-tatively significant, they would not detract greatly from the con-vergence result that is my focus. Moving into more sophisticated,higher value-added products and increasing physical productiv-ity are both ways of raising the available returns to labor. Theyare both channels for income convergence. To the extent thatquality upgrading takes place generally, and it does so more rap-idly in the poorer countries, it is simply another manifestation ofthe productive convergence I am interested in documenting.

A final interpretational difficulty relates to the possible roleof entry barriers in accounting for my results. The absence ofprices in the data forces me to conflate ‘‘revenue productivities’’with ‘‘quantity productivity’’ (Foster, Haltiwanger, and Syverson2008; Hsieh and Klenow 2009). Dollar values of productivity—‘‘revenue productivities’’—can change as a result of movements inprices as well as in physical productivities. The former, in turn,can be driven by shifts in entry barriers, independently of anychanges in physical productivity.

Consider, for example, an economy in which the failure ofmarginal value labor productivities to equalize across sectors isdue to barriers to labor mobility. As in Hsieh and Klenow (2009),


I could suppose these barriers are government-imposed restric-tions, such as preferential licensing or credit policies that advan-tage some firms or sectors over others. In this world, differencesin dollar labor productivities within countries would reflect themagnitude of these barriers, and my regressions would be captur-ing convergence and divergence in them.16 In particular, the ‘‘un-conditional convergence’’ result may reflect the fact thatindustries with high barriers have experienced a decline in bar-riers while industries with low barriers have experienced an in-crease. Interesting as this reading of the evidence may be in itsown right, it would be quite a different result than one aboutproductivity convergence.

However, in this case I am looking at industries across aswell as within countries. In fact, as the industry-by-industry re-gressions discussed previously (and the sigma-convergenceresults, discussed later) indicate, an essential part of the identi-fication for unconditional convergence comes from the variationwithin industries across different countries. Revenue productiv-ities are converging globally. It is not very plausible to attributedifferences in dollar labor productivities in, say, motor vehiclesacross countries to differences in entry barriers faced by motorvehicle producers in different countries, or to attribute conver-gence in revenue productivities to systematic changes in suchbarriers.

In addition, because low-productivity industries are in poorcountries, for changes in labor mobility barriers to account for myfindings, mobility barriers must have come down in poorer coun-tries and increased in the rich countries. If this were true, I wouldalso see much more rapid expansion in manufacturing in poorcountries. As I show later when I discuss aggregation issues,this doesn’t seem to be the case. The movement of labor into themanufacturing industries of the poorer countries is not signifi-cant or rapid enough to support such a conjecture. The sluggishpace of labor movement, in particular the relatively slow expan-sion of manufacturing in the poorest countries, does not suggestsystematic changes in barriers to mobility.

16. I am grateful to a referee for raising this possibility and for the relateddiscussion on revenue versus quantity productivity.


III.D. Sigma-Convergence

Beta-convergence does not guarantee sigma-convergence ina world where growth rates are driven not just by the forces ofconvergence but also by other determinants and shocks. The pre-sent data do not allow a very comprehensive analysis of sigma-convergence because I need to have a large sample of countrieswith data both at the beginning and end of the period to deter-mine whether the dispersion of productivity has diminished overany given time horizon. As discussed previously, the countrycoverage shrinks dramatically when I restrict the sample to across-section of any fixed time period.

The best I can do is to choose a time period that maximizesthe number of countries that can be included. Table V shows theresults for the 1995–2005 decade, which covers a variable numberof countries across different industries, and 63 countries for man-ufacturing as a whole. The overall pattern across two-digit man-ufacturing industries seems mixed, but a majority (16) showsdeclines in dispersion, some quite dramatically. Basic metals(ISIC 27); office, accounting, and computing machinery (ISIC30); and radio, television, and communication equipment (ISIC32) have all experienced substantial reductions in the standarddeviation of log productivity, of the order of 30 log-points or more,in a period as short as a decade. More remarkably, dispersion inmanufacturing as a whole has come down by 10 log-points in mysample of 63 countries. This amounts to a reduction of 10% indispersion.

Figures VI and VII provide a visual sense of these findings.The distribution of aggregate manufacturing productivity has aclear twin-peaked shape. Between 1995 and 2005, the two peakshave moved considerably closer to each other (Figure VI).

IV. Why Unconditional Convergence Does Not

Aggregate Up

The forces of convergence seem quite strong in manufactur-ing industries. It stands to reason that one ould uncover similarresults for certain other parts of the economy as well, perhapsmodern, tradable services such as financial or business services,among others. One might expect convergence at the sectoral levelto produce aggregate convergence as well, unless there are coun-tervailing forces pushing in the other direction. Yet the aggregate


TA

BL

EV

SIG

MA-C

ON

VE

RG

EN

CE,

1995–2005

ISIC

cod

eS

ecto

rN

um

ber

ofco

un

trie

sS

igm

ain

1995

Sig

ma

in2005

Dif

fere

nce

DT

otal

man

ufa

ctu

rin

g63

1.1

86

1.0

82

�0.1

04

15

Foo

dan

dbev

erages

43

1.0

86

0.9

42

�0.1

44

16

Tob

acc

op

rod

uct

s31

1.5

91

1.6

68

0.0

77

17

Tex

tile

s54

1.2

18

1.2

46

0.0

28

18

Wea

rin

gap

pare

l,fu

r34

1.1

66

1.3

01

0.1

35

19

Lea

ther

,le

ath

erp

rod

uct

s,an

dfo

otw

ear

34

1.1

44

1.2

69

0.1

25

20

Woo

dp

rod

uct

s(e

xcl

.fu

rnit

ure

)53

1.4

79

1.1

54

�0.3

25

21

Pap

eran

dp

ap

erp

rod

uct

s52

1.1

88

0.9

99

�0.1

89

22

Pri

nti

ng

an

dp

ubli

shin

g53

1.1

67

1.0

60

�0.1

07

23

Cok

e,re

fin

edp

etro

leu

mp

rod

uct

s,n

ucl

ear

fuel

38

1.4

15

1.8

35

0.4

20

24

Ch

emic

als

an

dch

emic

al

pro

du

cts

52

1.2

86

1.2

73

�0.0

13

25

Ru

bber

an

dp

last

ics

pro

du

cts

53

1.0

99

0.9

76

�0.1

23

26

Non

met

all

icm

iner

al

pro

du

cts

56

1.2

06

0.9

92

�0.2

14

27

Basi

cm

etals

50

1.2

91

0.8

93

�0.3

98

28

Fabri

cate

dm

etal

pro

du

cts

53

1.2

31

1.1

08

�0.1

23

29

Mach

iner

yan

deq

uip

men

tn

.e.c

.37

1.2

37

1.2

47

0.0

10

30

Offi

ce,

acc

oun

tin

g,

an

dco

mp

uti

ng

mach

iner

y23

0.9

94

0.6

88

�0.3

06

31

Ele

ctri

cal

mach

iner

yan

dap

para

tus

32

1.1

24

0.9

43

�0.1

81

32

Rad

io,

tele

vis

ion

,an

dco

mm

un

icati

oneq

uip

men

t26

1.2

29

0.9

14

�0.3

15

33

Med

ical,

pre

cisi

on,

an

dop

tica

lin

stru

men

ts45

1.3

80

1.2

02

�0.1

78

34

Mot

orveh

icle

s,tr

ail

ers,

sem

i-tr

ail

ers

35

0.9

45

1.0

12

0.0

67

35

Oth

ertr

an

spor

teq

uip

men

t27

1.0

92

0.9

44

�0.1

48

36

Fu

rnit

ure

;m

an

ufa

ctu

rin

gn

.e.c

.52

1.2

57

1.1

35

�0.1

22

37

Rec

ycl

ing

22

0.9

91

0.9

88

�0.0

03


FIGURE VI

Sigma-Convergence, All Manufacturing

FIGURE VII

Sigma-Convergence in Specific Industries


data do not support this conjecture. In this section I consider whyeconomies as a whole fail to exhibit unconditional converge des-pite the strong pull of convergence within manufacturingindustries.

Recall (from Table II) that aggregate manufacturing doesexhibit unconditional convergence. There is some evidence thatconvergence gets stronger the more one disaggregates withinmanufacturing. But it is clear that the bulk of the convergencefailure takes place as one goes from manufacturing (in aggregate)to the rest of the economy. So I focus here on the contrastingbehaviors of manufacturing and the economy (both as a whole),abstracting from aggregation issues within manufacturing.

To compare manufacturing to the rest of the economy, I needdata that go beyond what I have been using so far. I proceed bycombining INDSTAT2 with Penn World Tables 7.0 (Heston,Summers, and Aten 2011; PWT), which include data on aggregateGDP and (implicitly) total employment. Because data for GDPare in real purchasing power parity (PPP)-adjusted terms in thePWT, I first convert these to current dollars to render the econo-mywide data directly comparable to INDSTAT2.17 SubtractingINDSTAT2 manufacturing value added and employmentlevels from the aggregates in PWT yields presumptive valuesfor nonmanufacturing. We thereby obtain labor productivity fig-ures for nonmanufacturing and the entire economy that are con-sistent with both data sets. The employment shares ofmanufacturing and nonmanufacturing can be imputed in a simi-lar fashion.18

17. PWTs include data on real GDP per worker (rgdpwok), which isPPP-converted GDP chain per worker at 2005 constant prices. To undertake theconversion, first I recover the conversion factor between current and constantprices, by taking the ratio of current and constant price GDP data in PWT. Forcurrent GDP data, I use PPP-converted GDP per capita, G-K method, at currentprices (in I$) (cgdp). For constant price GDP data, I use PPP-converted GDP percapita (Laspeyres), derived from growth rates of c, g, i, at 2005 constant prices(rgdpl). The ratio of these two values ( cgdp

rgdpl) gives a price conversion factor betweencurrent and constant. Next I calculate a conversion factor between market pricesand PPP. For this I use the PWT values for PPP over GDP in national currency unitsper US$ (ppp) divided by the exchange rate to US$ in national currency units perUS$ (xrat). Using these two conversion factors, I convert PWT’s data on real GDPper worker into nominal U.S. dollars (ngdpwok = rgdpwok * ( cgdp

rgdpl) * ( ppxrat)). This gives

nominal GDP chain per worker at current prices (US$).18. To compute the employment share of manufacturing, I first compute the

total working population using data from PWTs. To do this, I divide PWT data on


I can now compare convergence behavior systematicallyacross different types of activities. I focus in this section on the1995–2005 cross-section, because these regressions do not mixdifferent time periods and are the easiest ones to interpret inthe present context. Column (2) of Table VI verifies the absenceof unconditional convergence in nonmanufacturing. The esti-mated convergence coefficient is very small and statistically in-distinguishable from zero. Column (3) replicates the exercise foreconomywide labor productivity, again confirming nonconver-gence. For comparison purposes, the bottom part of the tableshows the analogous results for the full panel.19

Other columns in Table VI highlight the importance of therelative size of manufacturing in driving convergence behavior.The employment share of manufacturing (a) appears to be a keyconditioning factor. As columns (5)–(7) show, a is a significantdeterminant of economywide growth. More important for my pur-poses, aggregate convergence seems to be conditional on a.Comparing columns (3) and (7), one can see that initial economy-wide productivity turns statistically significant in the aggregategrowth equation once I control for a. These are, of course, sparse

GDP per capita (rgdpch) by data on GDP per worker (rgdpwok). This gives thenumber of workers per capita. From this number and the total population figures(pop) in PWT, I calculate total employment. Total manufacturing employment isgiven by INDSTAT2 as the number of workers in manufacturing as a whole, in ISICcategory D. From this and the total employment number computed using PWT, theemployment share of manufacturing (a) can be calculated. To compute nonmanu-facturing labor productivity, I convert PWT data on total PPP converted GDP, G-Kmethod, at current prices in millions I$ (tcgdp) to nominal GDP, using the PPP–exchange rate conversion factor described in the previous note (ppp

xrat).Nonmanufacturing value added is calculated as the difference between this nom-inal value and the value added for manufacturing as a whole from INDSTAT2, inISIC category D. This number is then divided by nonmanufacturing employment(the difference between total employment and total manufacturing employment),to give nonmanufacturing labor productivity. From these numbers, growth ratescan be calculated to run convergence regressions for nonmanufacturing.

19. Interestingly, nonmanufacturing and the aggregate economy appear to ex-hibit some unconditional convergence in the baseline sample, at a rate of around1%, or one third that for manufacturing in the same sample. However, this is notrobust and seems to be a result of the correlation between the initial levels of prod-uctivity in manufacturing and nonmanufacturing. When initial manufacturingproductivity is also included in the regression, the coefficient on initial nonmanu-facturing productivity actually turns positive in both the aggregate economy andnonmanufacturing growth regressions (while manufacturing productivity is nega-tive and significant).


TA

BL

EV

I

EC

ON

OM

ICS

TR

UC

TU

RE

AN

DA

GG

RE

GA

TE

(NO

N)C

OV

ER

GE

NC

E

Dep

end

ent

vari

able

:d

ecad

al

gro

wth

ofla

bor

pro

du

ctiv

ity

in

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Man

ufa

ctu

rin

gN

onm

an

ufa

ctu

rin

gE

con

omy

Eco

nom

yE

con

omy

Eco

nom

yE

con

omy

Pa

nel

A:

1995–2

005

cros

s-se

ctio

nL

ogin

itia

lp

rod

uct

ivit

y,

man

ufa

ctu

rin

g�

0.0

20**

*�

0.0

26**

*�

0.0

15*

�0.0

13**

*(0

.006)

(0.0

09)

(0.0

08)

(0.0

04)

Log

init

ial

pro

du

ctiv

ity,

non

man

ufa

ctu

rin

g�

0.0

02

0.0

15**

0.0

02

(0.0

03)

(0.0

06)

(0.0

07)

Log

init

ial

pro

du

ctiv

ity,

econ

omy

(GD

P/w

ork

er)

�0.0

03

�0.0

10**

*(0

.003)

(0.0

03)

Alp

ha

(em

plo

ym

ent

share

ofm

an

ufa

ctu

rin

g)

0.2

90**

*0.3

11**

*0.3

86**

*(0

.093)

(0.0

71)

(0.0

93)

Per

iod

fixed

effe

cts

yes

yes

yes

yes

yes

yes

yes

Nu

mber

ofco

un

trie

s61

61

61

61

61

61

61

Nu

mber

ofob

serv

ati

ons

61

61

61

61

61

61

61

Pa

nel

B:

Pa

nel

spec

ifica

tion

Log

init

ial

pro

du

ctiv

ity,

man

ufa

ctu

rin

g�

0.0

10**

*�

0.0

23**

*�

0.0

10*

�0.0

11**

*(0

.003)

(0.0

05)

(0.0

05)

(0.0

03)

Log

init

ial

pro

du

ctiv

ity,

non

man

ufa

ctu

rin

g0.0

02

0.0

16**

*-0

.002

(0.0

02)

(0.0

04)

(0.0

05)

Log

init

ial

pro

du

ctiv

ity,

econ

omy

(GD

P/w

ork

er)

0.0

03

�0.0

09**

*(0

.002)

(0.0

03)

Alp

ha

(em

plo

ym

ent

share

ofm

an

ufa

ctu

rin

g)

0.3

09**

*0.2

97**

*0.3

44**

*(0

.049)

(0.0

34)

(0.0

43)

Per

iod

fixed

effe

cts

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Nu

mber

ofco

un

trie

s95

95

95

95

95

95

95

Nu

mber

ofob

serv

ati

ons

234

234

234

234

234

234

234

Not

es.

Cou

ntr

ysa

mp

leis

lim

ited

toth

ose

inth

e1995–2005

cros

s-se

ctio

nan

dp

an

elsa

mp

les

wit

hd

ata

for

all

colu

mn

s.P

an

elsp

ecifi

cati

onp

ools

data

for

1965–75,

1975–85,

1985–95,

an

d1995–2005.

Rob

ust

stan

dard

erro

rs(P

an

elA

)an

dst

an

dard

erro

rscl

ust

ered

at

cou

ntr

yle

vel

(Pan

elB

)are

rep

orte

d.

***p<

.01,

**p<

.05,

*p<

.10.


specifications that do not rule out alternative interpretations; acould be proxying for a wide range of other conditioning variableswith which it is correlated. The ‘‘conditional’’ convergence ratecontrolling for a is relatively low (1%) in light of estimates inBarro (2012), so this proxying is likely imperfect. In any case,these results are suggestive of the role played by the relativesize of manufacturing in explaining convergence behavioracross countries.

To investigate the issue more formally, I divide the economyinto manufacturing (m) and nonmanufacturing (n) activities.GDP per worker is the weighted average of labor productivityin these two activities: y ¼ �ym þ ð1� �Þyn, where the weight �is the share of the economy’s labor force employed in manufactur-ing. (I dropped country subscripts to avoid clutter.) Growth inGDP per worker is in turn expressed as

y ¼ ��m ym þ ð1� �Þ�nyn þ ð�m � �nÞd�,

where a ‘‘^’’ over a variable denotes proportional growth rates asbefore, and �m ¼

ym

y and �n ¼yn

y are the productivity premia/dis-counts for the two sectors.

I now impose some further structure on the growth decom-position, using the convergence results obtained thus far. Inparticular, I write the growth rates of manufacturing and non-manufacturing as:

yn ¼ g

ym ¼ gþ � ln y� � ln ym

� �,

where g is the underlying long-term balanced growth rate of theeconomy, y� is the productivity frontier in manufacturing, ym ismanufacturing labor productivity in the home economy, and �(>0) is the convergence coefficient in manufacturing. This formu-lation captures the basic asymmetry between manufacturing andnonmanufacturing, namely, that manufacturing is the benefi-ciary of a convergence ‘‘kick,’’ which peters out as the economygets closer to the frontier.

Substituting these in and noting that ��m þ ð1� �Þ�n ¼ 1,economywide growth becomes

y ¼ gþ ��m�ðln y� � ln ymÞ þ ð�m � �nÞ d�:ð4Þ


So growth equals an exogenous (or country-specific) compo-nent, a manufacturing convergence factor (that is decreasing inthe level of manufacturing productivity), and a reallocation term.The reallocation term captures the effect of changes in the com-position of employment across sectors when productivities differbetween manufacturing and nonmanufacturing. In particular,because �m >�n in the data, an increase in manufacturing employ-ment share (d�) raises growth. I denote the reallocation term as� � ð�m � �nÞ d�.

Equation (4) is helpful for understanding why manufactur-ing convergence does not translate into aggregate convergence.Most critically, it highlights the role of a, the share of manufac-turing employment. The economywide impact of manufacturinggrowth is mediated through this variable. One characteristic ofpoor countries is that they have very small formal manufacturingsectors. The average a for the poorest half of the baseline sampleis only around 5%. This means that even powerful convergenceeffects for manufacturing have only tiny consequences for theeconomy as a whole. The manufacturing premium �m tends tobe larger in poorer countries, but in practice this only partly com-pensates for the lower a (see later discussion).

In light of this, I would need the reallocation term � to notonly be large but to also vary systematically with incomes, withpoorer countries benefiting substantially more from reallocationtowards manufacturing. However, I fail to find such strong andsystematic reallocation effects in the data.

I illustrate these arguments with a numerical exercise,which quantifies equation (4). Table VII splits the 1995–2005sample into 10 deciles according to initial level of aggregateGDP per worker. I then compute the terms in equation (4) foreach decile. To perform the calculations, I choose values for a,�m, and ym that correspond to the averages for each decile. For�, I use .020, which is the value estimated for this sample ofobservations.20 For y* I use the average manufacturing product-ivity in the top decile of the sample (for manufacturing product-ivity). The value of g is set equal to 0 with no loss of generality.For each decile, the table shows the predicted manufacturinggrowth rate �ðln y� � ln ymÞ, the predicted aggregate growthrate due to manufacturing convergence ��m�ðln y� � ln ymÞ, and

20. This is different from the estimate noted in Table II because I am workingwith the 1995–2005 sample here (and not the baseline sample).


the total predicted growth rate including the reallocation term��m�ðln y� � ln ymÞ þ�.21

For example, for the poorest 10% of the sample, the averagevalues of the parameters are as follows: � ¼ 0:0278 , �m ¼ 7:7451,and ln y� � ln ym = 3.5160. As the first row of Table VII shows, thisyields a predicted manufacturing growth rate of 7.1%, an aggre-gate convergence term of 0.7%, and a predicted overall growthrate that is actually slightly lower at 0.6%.

Table VII shows a steep negative gradient in manufacturinggrowth as initial GDP per worker increases (column (4)). This isas expected from the results presented previously. The predictedmanufacturing growth rate goes down from 7.1% for the bottomdecile to 0.6% at the very top.22 The table also shows that verylittle of this gradient survives at the aggregate level once theeffect is scaled down by ��m, which is substantially smallerthan 1. The aggregate convergence term is a mere 0.7% for thebottom decile (an order of magnitude smaller). The key contribu-tor here is a, which not only is very small but increases as incomesrise, further moderating the forces of convergence (column (1)).This is only partly offset by the fall in �m (column (2)).

Furthermore, the reallocation term is tiny, often negative,and does not vary substantially across income levels to makemuch difference (column (6)). In principle, this term could havemade a substantial contribution to convergence if labor were tomove more rapidly to manufacturing in response to the largereturn differentials across sectors. The labor productivity differ-ential between manufacturing and nonmanufacturing ð�m � �nÞ isof the order of 300–700% for the lowest deciles. With a differentialof 500%, even if a were to increase by 0.5% annually (resulting inroughly a 5 percentage point increase in the employment share ofmanufacturing over a decade), the poorest economies in thesample would experience a growth boost of around 5�0.5 = 2.5percentage points. As it is, the actual change in a is a minutefraction of that and often negative (as can be observed from the

21. I use end-of-period �m and �n to compute � because I am dealing with dis-crete changes.

22. There is a small convergence effect even in the top decile because manufac-turing convergence depends on distance from the manufacturing frontier. Table VIorganizes deciles according to GDP per worker, not manufacturing productivity.


TA

BL

EV

II

PR

ED

ICT

ED

GR

OW

TH

RA

TE

SB

YIN

CO

ME

(GD

P/W

OR

KE

R)

DE

CIL

E

(1)

(2)

(3)

(4)

(5)

(6)

(7)

a� m

lny*

–ln

y m

Man

ufa

ctu

rin

ggro

wth

Aggre

gate

con

ver

gen

cete

rm�

Pre

dic

ted

aggre

gate

gro

wth

��

(3)

(1)�

(2)�

(4)

(5)+

(6)

Bot

tom

dec

ile

0.0

278

7.7

451

3.5

160

7.0

9%

0.6

7%

�0.0

016

0.5

1%

Dec

ile

90.0

739

3.4

913

3.3

287

6.7

1%

0.8

4%

�0.0

053

0.3

2%

Dec

ile

80.0

488

4.3

041

3.0

440

6.1

4%

0.4

2%

�0.0

004

0.3

8%

Dec

ile

70.0

731

3.7

142

2.5

166

5.0

7%

1.0

6%

�0.0

005

1.0

1%

Dec

ile

60.0

964

2.5

252

2.3

940

4.8

3%

0.6

9%

�0.0

003

0.6

6%

Dec

ile

50.0

794

2.8

562

1.8

194

3.6

7%

0.6

2%

�0.0

007

0.5

5%

Dec

ile

40.1

307

2.0

218

1.7

922

3.6

1%

0.6

2%

�0.0

003

0.6

0%

Dec

ile

30.1

415

1.3

768

1.1

459

2.3

1%

0.3

7%

�0.0

007

0.3

1%

Dec

ile

20.1

385

1.4

440

0.7

061

1.4

2%

0.1

7%

0.0

001

0.1

8%

Top

dec

ile

0.1

541

1.2

064

0.2

909

0.5

9%

0.0

9%

�0.0

006

0.0

3%

Not

es.

Sam

ple

:1995–2005,b

=.0

20,

lny*

=11.5

098;a,

y m,

lny*

an

dln

y mare

beg

inn

ing-o

f-p

erio

dvalu

es.

�is

calc

ula

ted

by

mu

ltip

lyin

gth

ech

an

ge

ina

over

the

dec

ad

ew

ith

the

end

-of-

per

iod

dif

fere

nce

sin

y’s,

an

dd

ivid

ing

the

resu

ltby

10

toan

nu

ali

zeit

.


negative entries in column (6)).23 Very similar results are ob-tained for the panel and baseline samples as well (not shown).

The net result is that the gradient for predicted economywidegrowth is barely distinguishable from zero (see column (7)). Thedifference between the gradients for manufactures and the ag-gregate economy is illustrated in Figure VI. In fact, the impliedconvergence factor for economywide growth is so small is that itwould be easily swamped by random measurement error or un-observed country-level determinants. For example, the smallnegative gradient in Table VII for predicted overall growth be-comes statistically indistinguishable from zero if I augment equa-tion (4) with a random error term distributed normally with mean0 and variance 0.0001 (a standard deviation in unexplainedgrowth rates across countries of just 1 percentage point).

In sum, aggregate nonconvergence appears to be explainedby the following combination of facts: (1) non-manufacturingactivities do not exhibit unconditional convergence; (2) poor coun-tries have little employment in manufacturing, depressing thecontribution of manufacturing to overall growth; (3) the shareof employment in manufacturing rises over the course of devel-opment, giving less-poor countries a growth boost; and (4) thereallocation effect is neither sizable enough nor systematicallylarger at lower income levels. In terms of quantitative magni-tudes, the first two factors play the dominant roles. From an eco-nomic standpoint, however, the last fact is perhaps mostinteresting, pointing to an important unexploited potential inpoor countries.

V. Concluding Remarks

I have provided evidence in this article that unconditionalconvergence is alive and well. One needs to look for it amongmanufacturing industries rather than entire economies. It is per-haps not surprising that manufacturing industries should exhibitunconditional convergence and, if the estimates here are to bebelieved, at quite a rapid pace, too. These industries produce trad-able goods and can be rapidly integrated into global production

23. Wong (2006) finds a very small reallocation effect in his study of 13 OECDcountries as well. Wong performs an accounting decomposition for OECD econo-mies, allocating convergence among seven different sectors and an interaction (re-allocation) term.


networks, facilitating technology transfer and absorption. Evenwhen they produce just for the home market, they operateunder competitive threat from efficient suppliers from abroad,requiring that they upgrade their operations and remain effi-cient. Traditional agriculture, many nontradable services, andespecially informal economic activities do not share thesecharacteristics.

The findings in this article offer new insight on the determin-ants of economic growth and convergence across countries. Theysuggest that lack of convergence is due not so much to economy-wide misgovernance or endogenous technological change but tospecific circumstances that influence the speed of structural re-allocation from nonconvergence to convergence activities. Thepolicies that matter are those that bear directly on this realloca-tion. As discussed in McMillan and Rodrik (2011) and Rodrik(2011b), what high-growth countries typically have in commonis their ability to deploy policies that compensate for themarket and government failures that block growth-enhancing

FIGURE VIII

Difference between Manufacturing and Economy-wide Growth Gradients.See text for explanation.


structural transformation. Countries that manage to affect therequisite structural change grow rapidly, and those that faildon’t.

Put differently, successful countries experience both product-ivity convergence in formal manufacturing and rapid industrial-ization. Unsuccessful countries make do with just the former.

Supplementary Material

An Online Appendix for this article can be found at QJEonline (qje.oxfordjournals.org).

John F. Kennedy School of Government, Harvard

University

REFERENCES

Acemoglu, Daron. Introduction to Modern Economic Growth (Princeton, NJ:Princeton University Press, 2009).

Barro, Robert J., ‘‘Convergence and Modernization Revisited,’’ NBER WorkingPaper No. 18295, 2012.

Barro, Robert J., and Xavier Sala-i-Martin, ‘‘Convergence across States andRegions,’’ Brookings Papers on Economic Activity, 1 (1991), 107–158.

Baumol, William, ‘‘Productivity Growth, Convergence, and Welfare: What theLong-Run Data Show,’’ American Economic Review, 76, no. 5 (1986),1072–1085.

Benetrix, Agustın S., Kevin H. O’Rourke, and Jeffrey G. Williamson, ‘‘The Spreadof Manufacturing to the Periphery 1870–2007: Eight Stylized Facts,’’ NBERWorking Paper No. 18221, 2012.

Bernard, Andrew, and Charles Jones, ‘‘Productivity and Convergence across U.S.States and Industries,’’ Empirical Economics, 21 (1996a), 113–135.

———, ‘‘Comparing Apples to Oranges: Productivity Convergence andMeasurement across Industries and Countries,’’ American EconomicReview, 86 (1996b), 1216–1238.

DeLong, J. Bradford, ‘‘Productivity Growth, Convergence, and Welfare:Comment,’’ American Economic Review, 78, no. 5 (1988), 1138–1154.

Durlauf, Steven, Paul Johnson, and Jonathan Temple, ‘‘Growth Econometrics,’’ inHandbook of Economic Growth, P. Aghion and S. Durlauf, eds. (Amsterdam:North-Holland, 2005).

Foster, Lucia, John Haltiwanger, and Chad Syverson, ‘‘Reallocation FirmTurnover, and Efficiency: Selection on Productivity or Profitability?’’American Economic Review, 98, no. 1 (2008), 394–425.

Heston, Alan, Robert Summers, and Bettina Aten, Penn World Table Version 7.0.Center for International Comparisons of Production, Income and Prices,University of Pennsylvania, May 2011.

Hsieh, Chang-Tai, and Peter J. Klenow, ‘‘Misallocation and Manufacturing TFPin China and India,’’ Quarterly Journal of Economics, 124 (2009), 1403–1448.

Hurwicz, Leonid, ‘‘Least-Squares Bias in Time Series,’’ in Statistical Inference inDynamic Economic Models, Tjalling C. Koopmans, ed. (New York: Wiley,1950).

Hwang, Jason J., ‘‘Patterns of Specialization and Economic Growth,’’ Ph.D disser-tation, Harvard University, 2007.


http://hwmaint.qje.oxfordjournals.org/lookup/suppl/doi:10.1093/QJEC/qjs047/-/DC1

qje.oxfordjournals.org

Levchenko, Andrei A., and Jing Zhang, ‘‘The Evolution of ComparativeAdvantage: Measurement and Welfare Implications,’’ NBER WorkingPaper No. 16806, 2011.

McMillan, Margaret, and Dani Rodrik, ‘‘Globalization, Structural Change, andProductivity Growth,’’ NBER Working Paper No. 17143, 2011.

Nickell, Stephen, ‘‘Biases in Dynamic Models with Fixed Effects,’’ Econometrica,49 (1981), 1417–1426.

Rodrik, Dani, ‘‘Unconditional Convergence,’’ NBER Working Paper No. 17546,2011a.

———, ‘‘The Future of Economic Convergence,’’ NBER Working Paper No. 17400,2011b.

Schott, Peter K., ‘‘Across-Product versus Within-Product Specialization inInternational Trade,’’ Quarterly Journal of Economics, 119 (2004), 647–678.

Sørensen, Anders, ‘‘Comparing Apples and Oranges: Productivity Convergenceand Measurement across Industries and Countries: Comment,’’ AmericanEconomic Review, 91, no. 4 (2001), 1160–1167.

Subramanian, Arvind. Eclipse: Living in the Shadow of China’s EconomicDominance (Washington, DC: Peterson Institute for InternationalEconomics, 2011).

Temple, Jonathan R. W., ‘‘Robustness Tests of the Augmented Solow Model,’’Journal of Applied Econometrics, 13, no. 4 (1998), 361–375.

United Nations Industrial Development Organization (UNIDO), INDSTAT4Industrial Statistics Database, 2011.

———, INDSTAT2 Industrial Statistics Database, 2012.Wong, Wei-Kang, ‘‘OECD Convergence: A Sectoral Decomposition Exercise,’’

Economics Letters, 93 (2006), 210–214.Young, Andrew T., Matthew J. Higgins, and Daniel Levy, ‘‘Sigma Convergence

versus Beta Convergence: Evidence from U.S. County-Level Data,’’ Journalof Money, Credit and Banking, 40, no. 5 (2008), 1083–1093.


i. introduction · sigma-convergence, referring to convergence in productivity levels, at least for...

Documents