Page 1

Do all countries grow alike?☆

J.W.B. Bosa, C. Economidoua, M. Koetterb,c,⁎, J.W. Kolarid

aUtrecht School of Economics, Utrecht University, Janskerkhof 12, 3512 BL, Utrecht, The Netherlands

bUniversity of Groningen, Faculty of Economics and Business & CIBIF, 9700 AV Groningen, The Netherlands

cResearch Center Deutsche Bundesbank, P.O. Box 10 06 02, G-60006 Frankfurt, Germany

dMays Business School, Texas A&M University, 4218 TAMU, College Station, Texas 77843-4218, USA

a b s t r a c ta r t i c l ei n f o

Article history:

Received 1 May 2007

Received in revised form 8 June 2009

Accepted 24 July 2009

JEL classification:

C33

O33

O47

Keywords:

Growth

Efficiency

Stochastic

Frontier analysis

Latent class

Regimes

This paper investigates the driving forces of output change in 77 countries during the period 1970–2000. A

flexible modeling strategy is adopted that accounts for (i) the inefficient use of resources, and (ii) different

production technologies across countries. The proposed model can identify technical, efficiency, and input

change for each of three endogenously determined regimes. Membership in these regimes is estimated, rather

than determined ex ante. This framework enables explorations into the determinants of output growth and

convergence issues in each regime.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

Over the past 30 years, significant effort has focused on providing

answers as to why some countries produce more than others. Yet,

growth differentials across countries still pose a puzzle to economists.

Standard economic models imply that the output level in an economy

depends entirely on the inputs used. For example, growth empirics

typically base cross-country regressions on a neoclassical production

function specification (Mankiw et al., 1992; Islam, 1995), often

expanded to include various sets of additional variables in an attempt

to explain economic growth.1However, considerable disagreement

remains regarding the explanatory variables to be included in the

analyses (see Temple, 1999, for a comprehensive survey). The per-

ceived failure of simple textbook models has stimulated a great deal

of interest in providing alternative theories of growth. Endogenous

growth theory emphasizes factors such as increasing returns to scale,

technology spillovers, learning-by-doing, and unobserved factors

(e.g., human capital), whereas the international economics literature

(Krueger, 1998; Dollar and Kraay, 2004) stresses the openness of

countries as an important conduit for growth.

This article develops a modeling strategy and presents empiri-

cal evidence that provides further insights into the determinants of

nations' growth. A structural methodology is adopted that allows for the

decomposition of output change into efficiency, technical, and input

change across a large panel of developed and developing countries. The

aim of the paper is to investigate whether all countries use the same

productionfunction,thesourcesofoutputgrowth,andifthereisevidence

of convergence. Policy implications of these results are discussed also.

Traditionally, cross-country growth empirics have assumed the

efficient use of inputs. The strong assumption that economic units

(countries) are always efficient (i.e., they always produce at the pro-

ductionpossibilityfrontier)impliesthatactualoutputisthemaximum

attainable output and that all countries are equally productive for a

given level of inputs. In reality, however, economic units may use the

best-practice (frontier) technology with varying degrees of efficiency.

Journal of Development Economics 91 (2010) 113–127

☆ We thank Bill Greene, Subal Kumbhakar, Clemens Kool, Bart Los, Luis Orea, Mark

Sanders, Spiro Stefanou, and Marcel Timmer, as well as seminar participants at Utrecht

School of Economics and participants at the European Workshop on Efficiency and

Productivity Analysis in Lille, for their valuable discussions. We are grateful for helpful

comments and advise received by an anonymous referee and the editor, Lant Pritchett.

Michael Koetter acknowledges financial support from the Netherlands Science

Foundation NWO (VENI grant 016.075.164). The opinions expressed are those of the

authors. The usual caveat applies.

⁎ Corresponding author. University of Groningen, Faculty of Economics and Business

& CIBIF, 9700 AV Groningen, The Netherlands.

E-mail addresses: j.w.b.bos@uu.nl(J.W.B.Bos),c.economidou@uu.nl(C.Economidou),

m.koetter@rug.nl (M. Koetter), j-kolari@tamu.edu (J.W. Kolari).

1See Barro (1991), Levine and Renelt (1992), Persson and Tabellini (1994) and

others.

0304-3878/$ – see front matter © 2009 Elsevier B.V. All rights reserved.

doi:10.1016/j.jdeveco.2009.07.006

Contents lists available at ScienceDirect

Journal of Development Economics

journal homepage: www.elsevier.com/locate/devec

Page 2

As a result, parameter estimates for the marginal effects of inputs are

biased in the presence of inefficiency. Efficient countries may increase

their output through technical change (i.e., shift of the frontier),

whereas inefficient countries may increase output by becoming more

efficient through the use of the best-practice technology.

Here, we account for inefficiency and estimate a stochastic pro-

duction frontier, which is the empirical analog of the theoretical

production possibility frontier. This modeling strategy therefore adds

structure to the unexplained residual. Under reasonable assumptions,

it disentangles the residual into inefficiency and measurement error.

Given this framework, we can decompose output changes into three

typesofchange: technical change(i.e.,shiftsof thefrontierovertime),

efficiency change (i.e., movements of a country toward or away from

the frontier), and input change (i.e., scale elasticity-adjusted changes

in factor use).

A growing body of recent empirical literature has conducted ef-

ficiency analyses along lines similar to that we propose in this paper,

but has used different modeling approaches. Previous studies have

decomposed output change into technical, efficiency, and input change.

For instance, Färe et al. (1994) use data envelopment analysis (DEA),

while Koop et al. (1999, 2000) and Limam and Miller (2004) apply

stochastic frontier analysis (SFA) to examine country-specific ineffi-

ciency in a number of developed and developing countries.2However,

not all countries necessarily share a single common frontier.3With

the exception of a handful of studies that allow for parameter hetero-

geneity when estimating frontier production functions, the frontier

literature has largely ignored this issue.

Similarly, conventional cross-country growth empirics mainly

examine the ‘average’ country (Temple, 1999) via a universal growth

model that governs the (per capita) output evolution in all countries.

However, if growth patterns diverge across countries, the ‘average’

country is not representative, parameter averages are less informative

about the factors that matter for a particular country (Solow, 1994),

and no country benefits from one-size-fits-all policy recommenda-

tions. More generally, the validity of treating all countries as a single

homogeneous group, for which the same variables have the same

effect on economic growth, seems increasingly questionable (see

Brock and Durlauf, 2001 for an extensive discussion of this issue).

In response to these concerns, a range of methodologies have been

proposed. For example, Durlauf and Johnson (1995) employ classifica-

tion and regression tree analysis, which identifies threshold values

in particular economic variables (e.g., output per capita, adult literacy

rates, etc.), to determine the appropriate grouping of countries.4Rather

than classifying countries ex ante into various groups on the basis of

geographic location or threshold values of particular economic var-

iables, Paap et al. (2005) and Davis et al. (2007) apply latent class

models that sort countries into different growth regimes according

to the similarities of their economic growth rates. They find that a

model with three and four groups of countries, respectively, is sta-

tisticallysuperiortoamodelthatassumeseconomiesarehomogeneous.

To improve the country classification, Davis et al. (2007) explore the

conditional (i.e., on institutions, openness, and macroeconomic policy)

distribution of countries' growth rates.

Some authors in the frontier literature have attempted to account for

heterogeneity in growth patterns. In exploring the sources of output

differentialsina panelofdevelopedanddevelopingcountries,Koopetal.

(2000) and Limam and Miller (2004) controlled for the quality of

productionfactorsusingeffectivelaborandcapital,insteadofactuallabor

and capital, and estimate regional frontiers.5The geographic division of

the sample is to a certain degree subjective, as some authors readily

admit,andmodelsmaybepoorlyidentifiedbecauseofthelackofdatafor

some regions such as Africa and Asia (see Koop et al., 2000, pages 286–

287). Tsionas and Kumbhakar (2004) instead proposed a stochastic

frontier production function augmented with a Markov switching

structure to account for different technology parameters across hetero-

geneous countries. Technology group membership depends on priors in

their Bayesian framework. Others, for example Koop et al. (2000),

critically view forming technology club memberships based on priors.

In this paper we allow for heterogeneous growth experiences.

Whereas most studies that classify countries apply an ex ante sorting

based on characteristics such as income and geography, we endogen-

ize the sorting of countries using a latent class model. The latent class

approachsupposes a simple parametric model and uses observed data

to estimate parameter values for each regime in the model. Among

the parameters estimated is the probability that a certain country in

a particular time period is a member of one of the regimes. These

probabilities result from a (multinomial logit) sorting equation and

depend on observable characteristics. In our case these characteristics

are conditioning variables common to the growth literature (Durlauf

and Johnson, 1995; Koop et al., 2000; Papageorgiou, 2002; Davis

et al., 2007) — namely, the level of human capital, openness to trade,

financial development andtheprimarysector share.We thusestimate

a regime-specific coefficient for each production factor. Each regime

exhibits ‘conditional independence’ because each variable is statisti-

cally independent of every other variable.

Hence, we advance methodology by introducing a structural and

flexible model that allows simultaneously for (i) the inefficient use

of resources, and (ii) different technologies across countries. We

augment the stochastic frontier production model with a latent class

structure, as proposed by Greene (2002a) and Orea and Kumbhakar

(2004). Using regime-specific production parameters, we identify

technical, efficiency, and input growth for endogenously determined

regimes. We introduce additional flexibility into the model by per-

mitting countries to switch between regimes over time. The efficiency

of countries in different regimes is estimated simultaneously but

relative to each regime's specific frontier. The latent class stochastic

frontier model enables us to avoid the routinely imposed assumption

of a common production function for all countries but yields results

that are comparable across countries at a given point in time.

Our work relates to and extends several important studies. Paap

et al. (2005) and Davis et al. (2007) also apply latent class models

to investigate growth experiences across a panel of countries. We

extend their work by accounting for the inefficient use of resources

2Various studies also investigate the role of efficiency in explaining growth

differentials for a panel of manufacturing industries in OECD countries. See, for

instance, Koop (2001) and Kneller and Stevens (2006).

3Theoretical contributions (Basu and Weil, 1998; Acemoglu and Zilibotti, 2001))

stress the ‘appropriateness’ of technology, suggesting that countries choose the best

technology available to them, given their input mix. On empirical grounds, a number

of works have emphasized that labor and capital cannot be equally productive in all

countries (Trefler, 1993; Tallman and Wang, 1994; Auerbach et al., 1994). Countries

are members of the same technology class if their marginal productivity of labor and

capital (the technology parameters that characterize the efficient production frontier)

are the same for a given level of inputs such that their input/output combinations can

be described by the same production frontier (Jones, 2005).

4A number of studies continue in this tradition. Papageorgiou (2002) extends the

work of Durlauf and Johnson (1995) by exploring whether trade can be used as a

threshold variable. Desdoigts (1999) proposes clusters based on culture, geographic

location, and OECD membership. Hobijn and Franses (2000) use a clustering method

as well, and find an abundance of convergence clusters. More recently, Bloom et al.

(2003), Canova (2004), and Sirimaneetham and Temple (2006) explore the existence

of multiple growth regimes. For instance, Sirimaneetham and Temple (2006) sort

economies into groups according to the value of an index of policy quality. Bloom et al.

(2003) argue that geographical variables determine the likelihood that a country will

be assigned to the two regimes they find. Canova (2004) takes a Bayesian approach to

examine income levels in Europe and, using initial income as a splitting variable, finds

four groups of countries.

5Koop et al. (2000) use the years of schooling embodied in the workforce to correct

for labor and agriculture and industry labor force participation to correct for physical

capital. They also allow for four different production frontiers: one for western

industrialized economies, one for East Asia, one for Latin America, and one for Africa.

Limam and Miller (2004) use the mean years of education and average age of physical

capital to account for quality of labor and physical capital, respectively. Like Koop et al.

(2000), they allow for heterogeneity by estimating regional frontiers based on five

geographic divisions: Africa, East Asia, South Asia, Latin America, and the West.

114

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 3

and allowing countries to change regimes. In relation to the frontier

literature, particularly studies by Koop et al. (1999, 2000), Koop

(2001), Limam and Miller (2004), Tsionas and Kumbhakar (2004),

and Kneller and Stevens (2006) which control for heterogeneity in

growth patterns across countries, we further account for heteroge-

neity in the growth patterns of countries without following ad hoc

or a priori clustering. Instead, we endogenize the regime allocation

by applying a latent sorting, conditioned on growth determinants

commonly used in prior literature.

Our empirical analysis is based on a sample of 77 countries during

the period 1970–2000, whereas similar studies generally examine

fewer countries over a shorter time span (Koop et al., 1999, 2000;

Koop, 2001; Limam and Miller Koopy, 2004; Tsionas and Kumbhakar,

2004;Paap et al., 2005; Kneller and Stevens 2006; Davis et al., 2007).

We proceed with our empirical analysis with three primary questions

in mind: (i) Do all countries follow the same growth experience?;

(ii) What are the sources of growth?; and (iii) What economic theory

provides a reasonable description of the growth processes?

Our results are easy to summarize. We find no evidence that

countries follow a common growth process, nor do we find that the

growth process of each country is entirely unique. Rather, we identify

three distinct growth processes or growth regimes. First, the mature

regimewithmostobservations,whichhostsmanyEuropeancountries

and the United States is characterized by high human capital ac-

cumulation. Second, the emerging regime, which contains many

Asian countries, is characterized by a relatively high level of financial

development. And third, the developing regime with many African

countries is characterized by a large primary sector share and high

degree of openness. Almost all regimes include countries from var-

ious geographical regions and/or income groups. Therefore, assigning

countries ex ante to certain groups based on their income or regional

criteria is not appropriate. We do find, however, that many countries

from the same region or with the same development level cluster in

the same regime.

Efficiency is economically and statistically important. Whereas the

neoclassical paradigm assumes that countries are fully efficient, we

demonstrate the fallacy of this assumption, since efficiency levels vary

across regimes. We find countries that belong in the emerging regime

exhibit the highest level of efficiency, whereas the least efficient

countries tend to be members of the developing regime.

The driving forces of growth also vary across and within regimes.

A consistent finding across all regimes shows that input accumulation

is an important source of growth, a finding we share with Koop

et al. (1999, 2000) and Limam and Miller (2004), among others. Our

findings suggest that countries' growth patterns do not necessarily

support just one growth explanation, such as the input accumulation

view or the productivity view. Instead, explaining countries' growth

performance requires a more pluralistic interpretation.

Countries can grow within a regime bycatchingup withthefrontier

(convergence) or migrating to a better regime (switching). Our results

showthatwiththeexceptionoffewcountries,regimeallocationisfairly

stable,andcountrieschangeonlyrarelyacrossregimes.Mostmigrations

pertain to countries switching between the mature and the emerging

regime. With regard to convergence, we find strong evidence of ‘con-

vergence clubs’ as different regimes converge at different rates to their

regime's steady state.

Overall, many of our findings could not have been obtained using

traditional approaches, such as imposing constant returns to scale,

ignoring inefficiency, assuming a single, common production func-

tion. By adopting a flexible modeling approach, we gain additional

insightsinto policiesthatshouldfostergrowth.We findno support for

the one-size-fits-all policy. More education alone, for instance, may

put countries in more advanced regimes, but for education to be

effective it needs the support of other development measures, such as

enhanced factor allocation, financial development, and trade policies.

The presence of significant inefficiency in one of the regimes fur-

ther suggests that development policies geared toward a better ex-

ploitation of existing technologies, rather than promoting technical

advances to push the production possibility frontier, might be bene-

ficial for some countries.

The remainder of this paper proceeds as follows. Section 2 pre-

sents the methodology and the econometric specification for esti-

mation. Section 3 discusses the data. Empirical results are presented

in Section 4. Section 5 concludes.

2. Methodology

We first introduce a model of production that accounts for in-

efficiency. We then augment the model with a latent class structure to

allow for more than one type of production. Finally, we decompose

the output change for each regime into technical, efficiency, and input

change.

2.1. A Stochastic frontier model of production

We model the performance of countries using a stochastic frontier

production model.6A frontier production function defines the maxi-

mum output attainable, given the current production technology and

available inputs.

If all industries produce at the boundary of a common production

set that consists of an input vector with two arguments, physical

capital (K) and labor (L), output can be described as:

YitT= f Xit;t;β

ð Þexp vit

f g;

ð1Þ

where Yit⁎ is the frontier (optimum) level of output in country i at

time t; Xitis the vector of inputs, namely, physical capital, Kit, and

labor, Lit; f and the parameter vector β characterize the production

technology; t is a time trend variable that captures neutral technical

change (Solow, 1957); and vitis an i.i.d. error term distributed as N(0,

σv

TwoaspectsofEq.(1)areworthnoting.First,thefrontierrepresents

a set of maximum outputs fora range of inputvectors. Therefore,at any

moment in time, it is defined by observations from multiple countries,

not just one. This definition differentiates our modeling approach from

conventional empirical growth approaches in which the leader country

achievingthe highestleveloftotalfactor productivity (TFP), constitutes

the frontier (Bloom et al., 2002; Cameron et al., 2005). An implicit but

nontrivial assumption in this literature suggests that technical progress

is described by the observations of a single country over time. Second,

ourmodelingapproach treats thefrontierasstochastic byincluding the

error term vit, which accommodates noise in the data and therefore

allows for statistical inference. In this respect, it fundamentally differs

from other (non-parametric) frontier industry-level analyses (Färe

etal.,1994;KumarandRussell,2002;LosandTimmer,2005)thatdonot

allow for random shocks around the frontier.7

Some countries, however, may lack the ability to employ existing

technologies efficiently and subsequently produce less than the fron-

tier output. If the difference between optimum and actual (observ-

able) output is represented by an exponential factor, exp{−uit}, then

the actual output, Yit, produced in each country i at time t can be

expressed as a function of the stochastic frontier output, Yit=Yit⁎exp

{−uit}, or equivalently:

2), which reflects the stochastic character of the frontier.

Yit= f Xit;t;β

ð

where uit≥0 is assumed to be i.i.d., with a half-normal distribution

truncated at zero, |N(0, σu

Þexp vit

fgexp −uit

fg;

ð2Þ

2)|, and independent from the noise term,

6Stochastic frontier analysis (SFA) was introduced by Aigner et al. (1977), (Battese

and Corra (1977), and Meeusen and van den Broeck (1977).

7For comprehensive reviews of frontier methodologies, see Kumbhakar and Lovell

(2000) and Coelli et al. (2005).

115

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 4

vit.8Efficiency, exp{−uit} is measured as the ratio of actual over

maximum output, exp −uit

f

{−uit}=1 implies full efficiency.9

A country is inefficient if it fails to absorb the best-practice tech-

nology. In this way, our approach is comparable to conventional, non-

frontier studies (Bernard and Jones, 1996a,b; Cameron et al., 2005)

that measure impediments to the absorptive capacity using TFP

changes. However, in these frameworks TFP changes cannot be sep-

arated into technical change and efficiency change (Kumbhakar and

Lovell, 2000). In addition, most studies assume a single technology

attainable by all countries in the sample. Instead, we explicitly model

the possibility of different technology regimes.

g =

Yit

YitT, where 0≤exp{−uit}≤1, and exp

2.2. Latent classes

Existing literature has proposed a range of methods to tackle

heterogeneity in countries' growth experiences. A common approach

is to include country fixed effects and dynamic panel data analyses

(Islam, 1995). Although this approach controls for differences in

average growth rates, it fails to control for differences in the marginal

effects of the regressors. An alternative approach identifies groups of

countries with similar growth behavior — for instance, similar income

(Auerbach et al., 1994) or human capital levels (Durlauf and Johnson,

1995) or degrees of openness (Papageorgiou, 2002), or the same

geographiclocation (Koopet al.,2000;LimamandMiller, 2004) — and

then estimates the production functions for each cluster of countries

separately.

Our approach diverges from these studies in that we endogenize

the classification of countries into different classes (regimes) using

a latent class model. The latent class approach employs a simple

parametric model to estimate regime-specific parameters of the

model.10The probability that a country belongs to a particular regime

can be calculated from a (multinomial logit) sorting equation and

depends on observable characteristics. In line with the economic

growth literature, we distinguish between four conditioning variables

that may sort countries into different groups, which we specify in

vector V. Human capital, openness to trade, financial development,

and the share of primary sector are growth determinants that may

affect factor accumulation, efficiency change, or technical change that

are parameters in the frontier production model.

Human capital affects output through various channels.11Human

capital contributes to factor augmentation. Barro (1991), for instance,

argues that a significant part of the effect of human capital on growth

is channeled through an increase in the investment rate for physical

capital.12Human capital also enhances the effectiveness of the work-

force, as it enhances the ability of the latter to learn, absorb, and work

with new technologies created by innovation efforts, thus contribut-

ing to the absorptive capacity of the economy (Abramovitz, 1986;

Benhabib and Spiegel, 1994). Furthermore, it accounts for aspects of

innovation not captured by the innovation sector (e.g., R&D), in-

cluding ‘learning-by-doing’ and ‘on-the-job-training’ (Romer, 1989;

Redding, 1996). Therefore, human capital can affect the inputs of

production, physical capital, and labor, as well as efficiency (through

absorption of existed advanced technologies) and technical change

(through innovation), which in turn influence the economic perfor-

mance of a country.

Another important conduit of growth is international trade.13

Openness to trade promotes the efficient allocation of human and

capital resources through comparative advantage and increases their

productivity. It further facilitates the dissemination of knowledge

and technological progress.14In particular, exporting may involve

some learning effects due to exposures to international contacts with

buyers and customers. These effects likely foster knowledge and

technology spillovers, such as access to technical expertise including

new product designs and new production methods.15Imports of qual-

ity foreign capital goods also serve as a means to acquire foreign

technology through reverse engineering.16Therefore, we include

opennessto tradein our analysisas a latentregimemembershipprob-

ability determinant.

Financial intermediaries shape the economic performance of a

countrybychoosingwhichfirmsgettousethesociety'ssavings.Awell-

developed financial sector can increase the marginal productivity of

capital by allocating funds to the projects for which the marginal

product of capital is highest by collecting information to evaluate

alternative investment projects (Greenwood and Jovanovic, 1990) and

by inducing investors to invest in riskier but more productive tech-

nologies via risk sharing (Schumpeter, 1934; Pagano, 1993). In the

absence of banks, households can guard against idiosyncratic liquidity

shocks only by investing in productive assets that can be promptly

liquidated, which causes them to forgo investments that are more

productive but also more illiquid. This inefficiency can be considerably

reduced by banks, which pool the liquidity risk of depositors and invest

most of their funds in more illiquid and more productive projects.17

Because it affects the productivity of inputs, efficiency, and technical

change in an economy, we also include financial development as a

latent regime membership probability determinant.

Finally, inefficient factor markets may affect the growth perfor-

mance of a country. To understand why, consider dual economy

effects. The marginal product of similar factors may not be equal

within a country due to reallocation impediments, such as labor in

an agricultural sector, which typically is less than perfectly mobile.

Vollrath (forthcoming) and Temple (2005) argue that the primary

sector share can affect growth in (at least) two ways. First, a large

primary sector can negatively affect growth if labor productivity is

low in this sector. The same influence may hold for its effect on (labor-

augmenting) technical change. Second, a high primary sector share

increases the effect of reallocation impediments and thereby reduces

the efficiency with which countries produce. Therefore, we let a

country's group membership be co-determined by sectoral structures

when estimating factor shares and, more important, group-specific

technical inefficiency levels.18

8We decompose the residual in Eq. (2), exp{vit}exp{− µit}, and identify its

components, exp{vit} and exp{− uit}, by re-parameterizing λ in the maximum

likelihood procedure, where λ(= σu/σv) is the ratio of the standard deviation of

efficiency over the standard deviation of the noise term, and σ (= (σu

composite standard deviation. The frontier can be identified by the λ for which the log

likelihood is maximized (see Kumbhakar and Lovell, 2000).

9Countries also may be inefficient if they use an input mix for which the prices of

inputs are not equal to the marginal returns to these inputs. Measuring this ‘allocative’

efficiency requires accurate input price data, which are particularly difficult to measure.

Therefore, we do not consider allocative efficiency and use the term efficiency only to

refer to technical efficiency.

10Throughout this paper, we use the terms ‘class’ and ‘regime’ interchangeably.

11Ontheeffectofhumancapitalongrowth,seeNelsonandPhelps(1966),Abramovitz

(1986), Lucas (1988), Romer (1989), Benhabib and Spiegel (1994), and (Cameron et al.,

2005).

12This evidence receives further support from Krueger and Lindahl (2001), and

Cannon (2000).

2+σv2)1/2) is the

13Classical references include Ben-David and Loewy (1998), Edwards (1998), and

Frankel and Romer (1999).

14These arguments are illustrated in the endogenous growth models offered by

Young (1991), Grossman and Helpman (1991), and Eicher (1999).

15For instance, the purchase of an input requires some degree of customization or

extended coordination between the seller and the buyer. Pack and Saggi (2001)

develop a model in which the sellers have an incentive to provide technology to

buyers, even if that technology may spill over to other sellers and buyers.

16When countries successfully imitate high-quality imported goods, they gain more

insight into how these goods are engineered and how to improve them. Connolly

(1998) discusses this ‘learning-to-learn’ effect.

17For empirical evidence, see King and Levine (1993), Easterly (1999), and Beck et al.

(2000), among others.

18We thank the editor for suggesting the primary sector share as a latent regime

membership probability determinant.

116

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 5

In our empirical specification, human capital (H), openness to

trade (T), financial development (F), and the primary sector share (P)

condition the allocation of countries to a specific regime. Within each

regime, countries share the same set of parameters, as in Eq. (2).

However, each regime also has its own set of parameters. Note that

conditioning group membership on this vector V affects all parts of

the growth decomposition for three reasons. First, countries are now

compared to the frontier of ‘their’ estimated peer group and thus

likely to be more efficient compared to a single frontier approach.

Second, factor elasticities of both capital and labor will be different

since the slope of regime-specific production frontiers will differ in V

too. Third, each frontier can now shift at their own pace, thereby

allowing for different technical change per regime. To estimate Eq. (2)

we must specify the functional form of the production frontier.

Specificationtestsfavoratranslogspecificationproductionfunction.19

In turn, for a translog specification with a general index of technical

change specified by means of time dummies Dt (see Baltagi and

Griffin, 1988) and regimes z(=1,..,Z), we can write a latent class

stochastic frontier as:

lnYit= βz+ β1jzln Kit+ β2jzln Lit+1

+1

2β22jzlnL2

+ δktjzln KitDt+ δltjzlnLitDt+ vitjz− uitjz′

2β11jzln K2

it

it+ β12jzln KitlnLit+ γtjzDt

ð3Þ

To operationalize Eq. (3), we must allocate each observation it

to a regime z. This is done by first making the contribution of each

observation to the likelihood function conditional on its regime mem-

bership. The unconditional likelihood then can be averaged over the

latent classes using the prior probability of membership in a class

(regime) as weights of the membership in class z.

In our conditional latent class frontier model, regime membership

probability conditional on the vector V (consisting of the four con-

ditioning variables, H, F, T, and P) determines regime membership.

Greene (2005) shows these conditional probabilities can be estimated

using a multinomial logit model:

θit=

exp Vitθz

ð

z = 1exp Vitθz

Þ

PZ

ðÞ;

ð4Þ

where θ measures the odds of belonging to regime z, conditional on

the values of the set of conditional variables Vit.

The resulting system of Eqs. (3) and (4)) is estimated by maxi-

mizing iteratively back and forth between posterior group probabil-

ities from Eq. (4) and the (weighted) log-likelihood function used to

estimate Eq.(3).20The likelihoodmaximization in Eq. (3) dependsnot

only on inputs and outputs per industry but also on efficiency (λ and

σ). Therefore, in contrast to a priori clustering on the basis of some

individual proxy, both the technology parameters β and efficiency u

can be determined endogenously through latent sorting into Z classes.

In summary, we redefine the production frontier as a latent class

frontier characterized by a system of equations: Z stochastic production

frontiers and a multinomial logit model with conditioning variables

(human capital, openness to trade, financial development, and the pri-

mary sectorshare)that accounts for the sorting (of countries) into each

of the Z regimes.

An important feature that distinguishes our modeling approach

from previous latent class studies (Greene, 2002a,b, 2005; Orea and

Kumbhakar, 2004) is that we allow countries to switch regimes over

time. For our sample of 77 countries observed over a maximum of

31 years, we define six different time periods: 1970–1974, 1975–

1979, 1980–1984, 1985–1989, 1990–1994, and 1995–2000. Eqs. (3)

and (4)) are estimated on annual data, and observations for some

years may be missing. Within each period, observations per country

are not independent because the country must fall within one of

the regimes during that period, and the probability of being in a

regime depends on the average of the variables used to estimate

regime membership.21However, across periods, observations on a

single country are treated as independent. For example, n moving

from t=5 (the last year of the period 1970–1974) to t=6 (the first

year of the period 1975–1979), a country is treated as a different

i in our panel dimension it and can switch regimes.

The advantage of this approach is that a country can be in one

regime in one period and in another regime next period.22As a result,

the regime allocation of a country is not restricted, and a country's

allocation in a given period is independent of its allocation in other

periods.This flexibilityadds an importantdimension toour analysisof

the components of countries' growth in that we can study regime

migrations. We turn next to decomposing output growth for different

regimes.

2.3. Decomposing output growth

A key aim of this paper is to relate our results to some of the major

macroeconomic debates about why and how some countries grow

faster than others. We therefore decompose output growth for each

regime into three components: input growth, represented by move-

ments along the frontier; technical growth, reflected by shifts of the

production frontier; and efficiency growth, captured by movements

toward (or away from) the production frontier as countries absorb

and implement best-practice technologies and reduce (or increase)

technical inefficiencies.

We take logs and totally differentiate Eq. (2) with respect to time,

which yields a convenient expression of output growth for every

regime, z:

gr Yit

ðÞ =

:Yit

Yit

=Aln fit

At

−Auit

At

+ eK

it

:

Kit

Kit

+ eL

it

:

Lit

Lit

;

ð5Þ

where εit

output with respect to the inputs, physical capital and labor,

respectively, and the dotted variables refer to time derivatives.23

The first term,Alnfit

where TCit>0 indicates an upward shift of the production frontier

(technical progress). Technical change can be attributed to capital

change (TCit

inputs in the form of pure technical change (TCit

−Auit

represents a reduction in inefficiency. Because we allow inefficiency

to vary freely over time, the time evolution of our efficiency term is

Kand εit

Ldenote the partial elasticity of the stochastic frontier

At, corresponds to technical growth, TCit=

Aln fit

At,

K) or labor change (TCit

L), or it may be independent of the

P). The second term,

Auit

At, corresponds to efficiency change, ECit= −

At, where ECit>0

19We test whether a translog is preferable to a Cobb–Douglas specification, which

appears in most prior literature. Our tests (see Section 4.1) support a translog

specification. Estimations of specifications with more flexible functional forms (Fourier

flexible) suffered from multicollinearity problems.

20The likelihood function is LF i;tjz

1

σzϕ

σj

and ϕ and Φ are the probability density and cumulative distribution functions of

standard normal distribution, respectively (see Greene, 2005).

ðÞ = f YitjKit;Lit;t;βz;δz;σz;λz

ðÞ =Φ λjeitjz

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

??

Φ 0

ð Þ

eitjz

??

, where ɛit|z=Yit|z − f(Kit|z, Lit|z, t; βz), λz=σujj

σvjz, σz=

σ2

ujz+ σ2

vjz

q

,

21In this modeling approach the allocation of a country in a regime in a specific

period depends on the period averages of the conditioning variables. We consider this

to be in line with theory, as we expect the allocation of a country in, for example, the

period 1990–1994 to depend on the average level of human capital (and the other

conditioning variables) in that period, rather than the initial level in 1970.

22Orea and Kumbhakar (2004) also propose a latent class frontier model. The subtle

difference between the models of Greene (2002a,b, 2005) and that of Orea and

Kumbhakar (2004) is described by Orea and Kumbhakar (2004, p.172). In the latter,

the log density (likelihood function) for an individual (or a country, here) is defined as

the same over all time periods in the model. In contrast, it defined for each individual

at each time t in the model of Greene (2002a,b). To allow countries to switch regimes,

we use the latent class model specified by Greene (2002a,b, 2005).

23For clarity, we delete the latent regime subscript z in this section.

117

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 6

not captured by a specific functional form (see Jondrow et al., 1982).

We approximateAuit

last two terms, ek

it it

ICL

itit

factor accumulation or changes in input factor elasticities. For example,

if a country exhibits constant returns to scale, changes in the level of

input factors do not influence the rate of change of output growth. If

laborexhibits,forexample,increasingreturnstoscale

anincreaseinthelaborforce

Lit

of output growth.

Table 1 summarizes the output growth decomposition for every

regime, z, based on the production function specified in Eq. (3).

Atby the growth rate of uitover time (uit− uit − 1

:

Kit

Kit+ el

:

Lit

Lit. The input change can vary for two reasons: pure

uit − 1

). The

it+

:

Lit

Lit, capture the input change, ICit= ICK

it= ek

:

Kit

Kit+ el

Aln f K;L;t;β

ð

AlnLit

Þ

??

= 1,

:

Lit

??

0 furtherincreasestherateofchange

3. Data

Our sample consists of 77 countries over the period 1970–2000.

The countries included in our sample are listed in Table A.3 in the

Appendix A. Annual data are retrieved from various sources. Output

(Y), measured as real gross domestic product (GDP), is constructed

from the Penn World Tables, version 6.1 (PWT 6.1), by taking the

product of the real per capita GDP, measured in 1996 international

purchasing power parity (PPP) dollars (chain index) and the national

population numbers. Labor force (L), measured in millions, is also

taken from the Penn World Tables. The computation of capital stock

(K) series, in 1996 international PPP dollars, follows the perpetual

inventory method in Hall and Jones (1999).24

To estimate the number of regimes and respective membership

probabilities, we rely on four conditioning variables commonly used

in the economic growth literature. Data on human capital (H), mea-

sured as the average years of education of the population that is

at least 25 years old, are retrieved from Barro and Lee (2001).25

Openness to trade (T), measured as the sum of exports and imports

relative to GDP, is obtained from the World Bank World Development

Indicators (2006). From the samedata source,we retrieve the primary

sector share (P) relative to total GDP. Finally, as a proxy for financial

development (F), we use the amount of deposits held in the financial

systemas a percentage of GDPprovided in Beck et al. (2000).26Table 2

contains the descriptive statistics.

4. Results

First, we examine whether there is a single production function.

That is, we test whether there is one universal model that can

adequately describe the growth experience of all countries. Second,

we present a tripartite output growth decomposition for countries

with similar growth experiences, according to their identified regime.

We relate these results to macroeconomic debates about whether

input accumulation or productivity drives output growth and to the

convergence hypothesis.

4.1. Is there a universal production function?

We start by investigating whether countries in our sample can be

described by a common production function. In estimating the latent

class frontier model defined by Eqs. (3) and (4), we first must

determine the number of regimes, Z. Multiple regime endogenous

growth models (Azariadis and Drazen, 1990; Kejak, 2003) merely

suggest the possibility of multiple steady states or growth regimes,

without being explicit about the exact number of regimes. Without

theoretical guidance into the ‘optimal’ number of regimes, we must

rely on statistical methodologies. We determine the number of re-

gimes in our preferred specification by following the suggestions

provided by Orea and Kumbhakar (2004) and Greene (2005).

We formally test for the optimal number of regimes, Z, using log-

likelihood ratio tests and the Akaike and Schwartz Bayesian

information criteria (AIC and SBIC, respectively), as we outline in

Table A.1 in the Appendix A. The preferred specification has the

highest log-likelihood value and the lowest AIC or BIC values. The test

results in Table A.1 favor a specification with three regimes over those

specifications with two.27Hence, our conditional latent class model

defined by Eqs. (3) and (4) supports the existence of three regimes.28

Table 3 below contains the estimated parameters for the translog

production function with a time trend (top panel), efficiency param-

eters (middle panel) and membership probability parameters (bot-

tom panel) for each of the three regimes we identify: emerging (A),

mature (B) and developing (C). Before explaining in subsequent sec-

tions in greater detail the growth process and further characteristics

that give rise to this taxonomy, we test whether the parameter esti-

mates differ significantly across regimes using Wald tests for joint

equality across regimes (A, B and C) (see the top panel of Table A.2

Table 1

Decomposition of output growth.

:

Yit

Yit= TCit+ ECit+ SCit

TCit= TCP

TCK

TCL

TCP

ECit= −

ICit= SEK

ICK

it

ICL

it

it+ TCK

it+ TCL

it

it= δktlnKit

it= δltlnLit

it= γt

uit− uit −1

uit −1

it+ SEL

it= ek

it= el

it

K̇it

Kit;where eK

L̇it

Lit; where eL

it= β1+ β11lnKit+ β12lnLit+ δktDt

it= β2+ β22lnLit+ β12lnKit+ δltDt

24We use a depreciation rate of 6% and the average growth over the first 10 years to

determine a country-specific average growth rate. For robustness we also calculate a

backward-looking capital stock using data from 1960 onwards. The results are

qualitatively similar. Our capital stock series has wider coverage than the PWT 6.1

variable for capital stock per worker, which is only available for 62 countries after

1965. When the two series overlap, the correlation coefficient between their log levels

is 0.97.

25Workers in different countries have different levels of skills. Typically, these skills

develop through education and experience. The lack of data about the latter prompts

us to measure education according to the years of schooling embodied in the labor

force. Given missing annual data, we use a linear interpolation per year. Assuming that

human capital is constant per five-year period does not change the results

qualitatively.

26See Beck et al. (2000) on February 21, 2006.

Table 2

Descriptive statistics.

MeanStd. dev.

Output (Y)

Capital (K)

Labor (L)

Human capital (H)

Openness to trade (T)

Financial development (F)

Primary sector share (P)

250.332

69.429

16.474

5.444

62.730

38.072

13.768

737.888

221.566

42.174

2.834

38.574

27.610

11.972

There is 1913 observations for 77 countries between 1970 and 2000.

27Three is a maximum number of regimes at which neither multicollinearity nor

over specification prohibits convergence of the maximum likelihood estimator. An

unconditional three-regime specification without further group determinants Z is

rejected, as are the Cobb–Douglas and translog specifications with a linear time trend.

These results are available on request.

28In response to comments of an anonymous referee, we also specified a latent class

model with the conditioning vector, V, as part of the deterministic kernel of the latent

class model, similar to some cross-country growth regression literature in which

explanatory variables such as human capital, openness to trade, and financial

development enter directly into the production function. Just as the additional

interaction of time and squared terms discussed before, this specification suffers from

over identification, multicollinearity, and convergence problems and therefore cannot

be estimated.

118

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 7

in the Appendix A). Low p-values less than 1% demonstrate that

parameters are jointly significantly different across the three regimes.

However, statistically significant differences for the parameters

in the production function are insufficient to assess whether the

specification of multiple regimes with their own production frontiers

is important for analyzing the output growth of countries. The middle

panel of Table 3 shows that inefficiency matters too. For regime C, the

efficiency parameter, λ(=σu/σv, the ratio of the standard deviation of

efficiency over the standard deviation of the noise term), is 2.053 and

significant at the 1% level. As such, inefficiency is approximately twice

as great as noise in this regime. The same result holds for regime B,

where inefficiency is one-and-a-half times times the size of noise (λ

is 1.525) and significant at the 1% level. In regime A, however, the

production process is fully efficient, as exemplified by the insignifi-

cance of λ.

Finally, the bottom panel of Table 3 demonstrates the importance

of the conditioning variables. The use of a multinomial logit spe-

cification implies an estimation of membership likelihood relative to

the reference group, orregime C here. Financialdevelopment, primary

sector share, and openness to trade have significant effects on the

probability of belonging to regime A. For regime B human capital,

primary sector share, and openness to trade are significant. For

example, an increase in financial development (human capital) of 1%

increases the probability of belonging to regime A (B) by 1.04%

(1.53%).29Wald tests (see the bottom panel of Table A.2 in the

Appendix A) show that the joint effect of a change in all four regime

determinants on the probability of belonging to regime A differs from

its effect on the probability of belonging to regime B. However, the

coefficients for human capital and the primary sector share are not

significantly different between regimes A and B. Financial develop-

ment, in turn, is critical distinguishing between regimes A and B. The

effect of financial development on regime A membership is both

higher and more significant. These results confirm the importance

of the mix of regime determinants. Individual determinants (e.g.,

financial development) may be important, but the interaction be-

tween regime determinants is especially crucial to identify relevant

peer groups of countries.

To explore differences among the three regimes further, we

present each regime's factor elasticities, εit

estimates, uit, and the marginal rate of technical substitution, MRTS, in

Table 4.

As Table 4 clearly shows, there are differences between regimes A

and B on the one hand and regime C on the other. The latter regime,

which is the smallest in terms of number of observations, has all the

characteristics of a less developed regime. Average inefficiency is

more than 10%, and the productivity of capital, as measured by the

capital elasticity (εit

with a very low marginal rate of technical substitution. Regime A

exhibits the highest labor elasticity and is almost 100% efficient.

Capital elasticity is highest in regime B, which also has the highest

marginal rate of technical substitution. The differences in factor

elasticities across regimes are also statistically significant, as shown in

the middle panel of Table A.2 in the Appendix A.

The relatively lower values of elasticities of capital and labor in

regime C can be explained in conjunction with Table 3. Regime A has

the highest level of financial development, whereas regime B exhibits

the highest level of human capital, as the bottom panel of Table 3

shows. More human capital and a better developed financial system

should contribute to the productivity of capital and labor and there-

fore increase the probability of that country belonging to a regime

with higher capital or labor elasticity. The average values for these

conditioning variables confirm the status of the lesser developed

regime C, which is characterized by a low level of human capital and

financial development.

More pronounced differences exist between the primary sector

share of regimes A and B compared to regime C. The share of the

relatively unproductive and often inefficient agricultural sector

(Vollrath, forthcoming) is the highest in regime C. A high share of

this relatively unproductive sector increases the probability that a

country in such a regime exhibits low productivity of capital and labor

due to existing inefficiencies. Regime C is also characterized by a high

openness to trade and the association of this trait with lower elas-

ticities of labor and capital supports some existing concerns about the

benefits of openness to trade for developing countries. If market or

Kand εit

L, technical efficiency

K) is less than half that of the other regimes, in line

29We calculate probabilities by taking the exponent of the coefficients from the

bottom panel of Table 3.

Table 4

Factor elasticities, efficiency, and marginal rate of technical substitution.

Regime ARegime BRegime C

Capital elasticity (εit

Labor elasticity (εit

Technical efficiency (uit)

MRTS

Observations

K)0.570

0.396

0.999

2.198

689

(0.086)

(0.160)

(0.000)

(3.182)

0.658

0.299

0.953

2.500

919

(0.116)

(0.114)

(0.024)

(0.874)

0.255

0.556

0.898

0.699

305

(0.119)

(0.326)

(0.058)

(10.060)

L)

There is 1913 observations for 77 countries over the period 1970–2000. All calculations

are based on latent class-specific parameter estimations evaluated at the mean.

Standard deviations in parentheses. MRTS is the marginal rate of technical substitution,

calculated as the ratio of scale elasticity-adjusted capital to labor change.

Table 3

Latent class frontier estimation results.

Regime

ABC

VariableCoeff.Coeff.Coeff.

Latent technology regime parameters

Constant

−0.669

ln K

ln L

ln K2

−0.050

ln L2

−0.124

ln K×ln L

D2

D3

D4

D5

D6

ln K×D2

ln K×D3

ln K×D4

ln K×D5

ln K×D6

ln L×D2

−0.159

ln L×D3

−0.030

ln L×D4

−0.325

ln L×D5

−0.172

ln L×D6

−0.295

(0.106)⁎⁎⁎

(0.012)⁎⁎⁎

(0.017)⁎⁎⁎

(0.005)⁎⁎⁎

(0.007)⁎⁎⁎

(0.006)⁎⁎⁎

(0.034)⁎⁎⁎

(0.030)⁎⁎⁎

(0.031)⁎⁎⁎

(0.031)⁎⁎⁎

(0.034)⁎⁎⁎

(0.012)⁎⁎⁎

(0.017)⁎⁎

(0.016)⁎⁎⁎

(0.014)⁎⁎⁎

(0.015)⁎⁎⁎

(0.018)⁎⁎⁎

(0.024)

(0.022)⁎⁎⁎

(0.019)⁎⁎⁎

(0.020)⁎⁎⁎

−0.334

0.394

0.579

0.036

0.000

0.018

0.098

0.338

0.194

0.243

0.207

0.103

0.133

0.169

0.185

0.147

−0.222

−0.350

−0.372

−0.442

−0.369

(0.019)⁎⁎⁎

(0.011)⁎⁎⁎

(0.017)⁎⁎⁎

(0.005)⁎⁎⁎

(0.008)

(0.006)⁎⁎⁎

(0.022)⁎⁎⁎

(0.023)⁎⁎⁎

(0.023)⁎⁎⁎

(0.024)⁎⁎⁎

(0.025)⁎⁎⁎

(0.014)⁎⁎⁎

(0.015)⁎⁎⁎

(0.014)⁎⁎⁎

(0.016)⁎⁎⁎

(0.016)⁎⁎⁎

(0.020)⁎⁎⁎

(0.021)⁎⁎⁎

(0.020)⁎⁎⁎

(0.023)⁎⁎⁎

(0.020)⁎⁎⁎

−1.015

0.403

0.204

0.024

0.279

−0.065

0.373

0.195

−0.141

0.163

0.093

0.081

−0.011

−0.120

−0.210

−0.190

−0.088

0.118

0.238

0.031

0.213

(0.150)⁎⁎⁎

(0.056)⁎⁎⁎

(0.075)⁎⁎⁎

(0.015)⁎⁎⁎

(0.016)⁎⁎⁎

(0.012)⁎⁎⁎

(0.145)⁎⁎⁎

(0.151)

(0.159)

(0.154)

(0.148)

(0.043)⁎⁎

(0.046)

(0.055)⁎⁎⁎

(0.061)⁎⁎⁎

(0.064)⁎⁎⁎

(0.075)

(0.075)

(0.082)⁎⁎⁎

(0.077)

(0.075)⁎⁎⁎

0.471

0.659

0.060

0.681

0.439

0.599

0.446

0.679

0.031

0.032

0.214

0.140

0.168

Efficiency parameters

σ

λ

0.115

0.011

(0.004)⁎⁎⁎

(1.119)

0.139

1.525

(0.008)⁎⁎⁎

(0.282)⁎⁎⁎

0.275

2.053

(0.020)⁎⁎⁎

(0.469)⁎⁎⁎

Regime membership probability parameters

Constant3.276

Human capital0.176

Financial

development

Primary sector

share

Openness

to trade

Observations689

(1.085)⁎⁎⁎

(0.158)

(0.016)⁎⁎⁎

2.528

0.425

0.022

(1.079)⁎⁎⁎

(0.157)⁎⁎⁎

(0.016)

Reference group

Reference group

Reference group0.039

−0.095(0.025)⁎⁎⁎

−0.094(0.025)⁎⁎⁎

Reference group

−0.031(0.006)⁎⁎⁎

−0.027(0.006)⁎⁎⁎

Reference group

919305

Standard errors in parentheses; the data refer to 1913 observations on 77 countries over

the period 1970–2000; Dk, k=2, 3, 4, 5, and 6 are time dummies for the periods 1970–74,

1975–79, 1980–84, 1985–89, 1990–94, and 1995–2000, respectively; σ[=(σu

and λ[=σu/σv] are efficiency parameters; the log-likelihood value is 1119.54; significance

at the 10/5/1% levels (⁎/⁎⁎/⁎⁎⁎).

2+σv

2)1/2]

119

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 8

institutional imperfections exist, such openness actually can lead to

under-utilization of human and capital resources, concentration in

extractive economic activities, and specialization away from techno-

logically advanced increasing-returns sectors (Grossman and Help-

man, 1991; Sachs and Werner, 1999; Rodriguez and Rodrik, 2001).

Therefore, it is not surprising that the (least) developed economies in

regime C exhibit low productivities of capital and labor.

Estimates in Table 4 also reveal significant heterogeneity in the

elasticities of inputs across regimes. We find almost constant returns

to scale in all regimes, as most of cross-country regressions assume

when examining growth differentials (Mankiw et al., 1992). However,

in regime C capital elasticity is lower than labor elasticity, which

contrasts with results from the growth empirics literature that reports

a marginal product of capital as high as 0.60 (Mankiw et al., 1992).

Furthermore, efficiency levels are statistically significant and different

for every regime. Overall, we conclude that there is no average or

representative country. Also, all countries do not operate at the fully

efficient production frontier. Instead, we find different regimes of

countries, each with its own specific characteristics. Rather than

merely assigning countries to groups, we find support for using the

four conditioning variables jointly to determine the allocation of a

country to regimes A, B, or C.

Another distinctive feature of our model is the possibility that a

country may change regimes over time. This raises the question if and

how often countries do change membership over time. We depict

regime migrations in Table 5 including the frequency and absolute

numberofregimeallocationchangesbetweenanytwotimeperiods.30

While the migration pattern observed may appear rather dynamic,

imposingregimestickinessbynotpermittingregimeswitchesimplied

counter-intuitive country groupings in an earlier version of this paper,

presumably reflecting changing technology and group determinants.

We prefer here to permit countries to follow different growth

processes due to regime switches similar in spirit to Jerzmanowski

(2006), rather than imposing the rigid assumption of unique

equilibrium growth per country.

Most observations are located on the diagonal of Table 5, which

indicates that overall countries appear to change relatively rarely in

terms of their production structure. At the same time, especially for

countries in regimes A and B a stable group allocation seems difficult.

Different technology regimes appear to be relevant for some

countries' growth processes at different times. We checked if some

countries are ‘borderline’ cases in the sense that our model allocated

them to regimes A or B with a conditional probability that is close to

50%, the conventional cut-off level in the multinomial logit model of

Eq. (4). However, the conditional probability of group membership is

very high in almost all cases. In fact, it is above 90% for more than 90%

of the sample. Hence, the relatively high frequency of regime switches

between regimes A and B is not the result of our model's flexibility.

This active migration pattern across technology regimes is to some

extent in line with Jerzmanowski (2006). Using a Markow-regime

switchingmodel, henotes thatalmostall countries in hissample ‘visit’

each of the growth regimes identified there on the basis of output-

per-worker growth dynamics alone. In this sense, the frequency of

migrations we observe does not seem excessive. It is also important

to note that most migrations between the two fairly developed

technology regimes A and B involve countries that switch back and

forth (see also Table A.3). Of the 79 (42+37) switches between

regimes A and B, approximately 60% pertain to countries that change

back and forth.31Put differently, 36 out of 42 moves from A to B (27

out of 37 from B to A) are accounted for by the same countries, which

may simply be hard to classify. In line with Jerzmanowski (2006), we

also find that regime migrations are rarely a viable strategy to escape

poverty traps, since exiting the worst performing regime C for good is

rare. Most extreme upgrades from regime C to A are followed either

by gradual (for example Rwanda) or straight (for example Togo)

‘downgrades’ back to the regime C. Finally, as shown in the last

column of Table 5, few countries move from regimes A and B to

regime C.

A natural question that arises at this point pertains to which

countries belong to which regime and how plausible the allocations

are. Table A.3 in the Appendix A provides a list of the countries and

shows regime memberships for each of the six periods across which

they are permitted to switch regimes. Each country in our sample in

each period joins the regime for which it has the highest conditional

probability. Most countries belong to either one or two regimes over

the entire sample period. Regime B is the most populated followed by

regime A.

To confirm the plausibility of our classification, we provide Table 6,

which shows the difference between the subjective and objective

probabilitiesofbeingaregimemember.Thesubjectiveprobabilityisthe

ratio of the number of years a country has been a member of a regime

to the total number of years it appears in the sample. The objective

measure is calculated as the product of the number of years a country

appears in the sample and the relative size of the regimes (e.g., 689/

(689+919+305) for regime A). We calculate both probabilities

after determining the regimes. A positive number in Table 6 indicates

that countries from a certain region are more likely to be members of a

regime than if the regions were randomly distributed across regimes.32

The evidence in Table 6 shows that geography matters. Our

classification justifies, to a certain extent, the regional classification

argument (countries in the same geographical region may have

similar endowments, such as natural resources). Asian countries are

most likely to be members of regime A, whereas regime B is very

likely to contain European countries. All regions are underrepresented

in the labor-intensive and inefficient regime C, with the notable

exception of Africa (mainly sub-Saharan countries). A few Asian and

Latin American countries, such as Pakistan, Indonesia, and Honduras,

Table 5

Migration matrix.

To regime

ABCTotal

From regime

A

54.37

(56)

26.24

(37)

17.02

(8)

34.71

(101)

40.78

(42)

72.34

(102)

12.77

(6)

51.55

(150)

4.85

(5)

1.42

(2)

70.21

(33)

13.75

(40)

100

(103)

100

(141)

100

47

100

(291)

B

C

Total

Numbers denote the percentage probability of moving from regime to another. The

number of countries per cell appears in parentheses.

30Since we distinguish six episodes, we have five migration matrices but display only

the aggregate, unconditional migration probabilities.

31Finland and Trinidad and Tobago migrate in and out of these two regimes a total of

three times.

32We exclude Oceania from the table due to few observations.

Table 6

Do geography and income matter?

RegionRegime ARegime BRegime C

Africa

Americas

Asia

Europe

−0.102

0.044

0.239⁎⁎

−0.073

−0.283⁎⁎⁎

0.060

−0.162⁎

0.215⁎⁎⁎

0.385⁎⁎⁎

−0.104⁎⁎

−0.077⁎

−0.142⁎⁎⁎

Income level

Output per capita15.749 (12.480)17.977 (13.885)3.048 (5.156)

Numbersrefertothedifferencesbetweentheconditionalandunconditionalprobabilitiesof

being a regime member. Differences between conditional and unconditional probabilities

are tested using a two-sided t-test. Significance at the 10/5/1% levels (⁎/⁎⁎/⁎⁎⁎).

120

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 9

are also allocated to regime C, but they eventually exit this worst

performing regime and enter either regime A or B. In addition, we find

some evidence in favor of the importance of similar production

structures (countries with similar level of development, such as

income per capita, may have similar structures of production). Table 6

shows that high-income countries (e.g., European, and the United

States) tend to be members of regime B, medium-income countries

(e.g., Asian nations) often appear in regime A, and low-income

countries (e.g., African nations) are mostly in regime C.

Our classification can therefore, to some extent, be compared with

studies that apply similar methodologies (e.g., stochastic frontier

analysis) but base their country allocation ex ante on geography

(Koop et al., 2000; Limam and Miller, 2004). However, in every region

some countries may behave very differently from other countries in

the same geographical region. For example, as Davis et al. (2007)

argue, purely based on geography, the Philippines would be placed

with other countries in East Asia, although its development is much

more akin to that of many Latin American countries. This observation

is supported by our classification. Other studies, for example, Paap

et al. (2005) and Davis et al. (2007), use latent class analysis and

allocate countries based on multiple conditioning variables, in line

with our approach. However, these studies do not consider the pos-

sibility that countries can move over time to a more (less) advanced

regime, which is a realistic scenario.

In summary, three main findings emerge from our analysis so far.

First, all countries do not follow a common growth process, nor is the

growth process of any country entirely unique. Instead, we find three

distinct growth processes or, equivalently, three growth regimes. In

addition,wefindthatsomecountriesovertimeimprove(ordeteriorate)

their production and move to a more (less) advanced regime.

Second, (in)efficiency, which has been widely ignored by con-

ventional growth empirics, is statistically important in our study and

quantitatively different across regimes. We find that countries from

regime A are fully efficient, whereas the least efficient countries

appear in regime C. This division implies that development policies

for some countries geared toward a better exploitation of existing

technologies, rather than augmenting new ones, may be beneficial.

Third, membership in a certain regime depends on the joint effect

of multiple factors — namely, human capital, openness to trade,

financial development, and primary sector share. Our results clearly

show that no single factor can explain the allocation of countries to a

certain regime.

To shed more light on the growth experience of the three regimes,

especially the quite ‘similar’ regimes A and B, we turn next to the

decomposition of output growth for every regime.

4.2. How do countries grow?

Having identified the number of regimes and their characteristics,

we next consider how countries in each regime grow. Therefore, we

decompose output growth per regime into three components: input

growth, technical growth, and efficiency growth, as in Eq. (5). To

allow for potential heterogeneity in growth patterns within regimes,

we identify high-, medium-, and low-growth countries according to

the 33rd and 66th percentiles of the overall growth distribution as

cut-off points in each regime. In Table 7 we present the break down

of countries in each regime according to their growth performance.

Fig. 1 graphically presents a more detailed decomposition of technical

change and factor accumulation.

In the comparison of regime B with regime A in Section 4.1, we

noted that the differences between their conditioning variables and

theircoefficients appeared marginal.However,we cautionedagainsta

comparison of individual variables and coefficients and emphasized

the multivariate effect of the conditioning variables. We confirm this

latter point by realizing that output growth decomposition in Table 7

is markedly different for each regime.

We identify regime A as the emerging regime, which contains

many Asian countries and has the most productive labor among all

regimes as well as a relatively high level of financial development

and human capital that is somewhat lower than that for countries

in regime B. The output growth decomposition in Table 7 also re-

veals that countries in regime A grow primarily as a result of factor

accumulation and technical change.33As Fig. 1 shows, the former

effect consists predominantly of capital accumulation, whereas the

latter influence consists mostly of pure technical change (especially in

the high-growth countries in this regime). These results confirm the

importance of financial development in facilitating capital accumu-

lation, through both domestic savings and foreign capital (Pagano,

1993) as well as in reallocating capital to firms that generate the

greatest technical change (Schumpeter, 1934). Efficiency change is

positive but very small in this highly efficient regime.

We can relate these findings about regime A to the ongoing debate

aboutthesourcesoftheimpressiveoutputgrowthperformanceofsome

East Asian countries. Some studies (Young, 1994) argue that East Asian

countries grow primarily through factor accumulation, whereas others

(PackandPage,1994)pointtotheroleoftheproductivityofinputs.Our

results suggest that factor accumulation is important but technical

change is a key factor overall, especially for high-growth countries.

Regime B is the mature regime, which contains many European

countries and the United States, has the highest output per capita,

indicates an important role for human capital, and achieves the

highest capital elasticity of all regimes. Output growth for countries in

this regime is driven by factor accumulation, especially through

increases in the capital stock. Such increases may result from a high

human capital stock, which increases the rate of investment (Barro,

1991; Krueger and Lindahl, 2001; Sianesi and Reenen, 2003). The

regime is also characterized by negative technical change, largely due

to labor-augmenting technical regress, as we show in Fig. 1. The main

explanation for the negative technical change in the low-growth

Table 7

Output growth decomposition.

Regime A

High MediumLowTotal

gr(Yit)

TCit

ECit

SCit

N

0.309

0.217

0.003

0.067

325

(0.193)

(0.194)

(0.012)

(0.029)

0.075

0.012

0.002

0.061

161

(0.038)

(0.039)

(0.011)

(0.023)

−0.108

−0.170

0.004

0.057

203

(0.085)

(0.091)

(0.020)

(0.023)

0.123

0.055

0.003

0.063

689

(0.227)

(0.220)

(0.015)

(0.026)

Regime B

HighMediumLowTotal

gr(Yit)

TCit

ECit

SCit

N

0.210

0.123

−0.001

0.073

211

(0.069)

(0.086)

(0.013)

(0.023)

0.058

−0.003

0.000

0.061

371

(0.036)

(0.040)

(0.015)

(0.018)

−0.069

−0.124

−0.004

0.059

337

(0.077)

(0.090)

(0.022)

(0.024)

0.042

−0.018

−0.002

0.063

919

(0.120)

(0.119)

(0.018)

(0.022)

Regime C

HighMediumLowTotal

gr(Yit)

TCit

ECit

SCit

N

0.302

0.225

0.008

0.047

165

(0.126)

(0.153)

(0.042)

(0.031)

0.069

0.017

0.004

0.048

74

(0.036)

(0.050)

(0.030)

(0.026)

−0.157

−0.191

−0.003

0.037

66

(0.158)

(0.158)

(0.051)

(0.032)

0.136

0.085

0.004

0.045

305

(0.222)

(0.217)

(0.042)

(0.030)

Standard errors in parentheses. High growth >66th percentile of the total growth

distribution. Low growth <33rd percentile of the total growth distribution. N is the

number of observations in each regime. Variables are as defined in Table 1.

33Some low-growth countries in this regime exhibit technical regress. For example,

Brazil exhibits technical regress in the early 1990s. Since 1989 was the first

presidential election after 29years of military rule, technical regress may reflect the

burden of setting these decades of economic mismanagement straight.

121

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 10

subsample of regime B is the (temporary) migration of countries from

regimes A and C (e.g., South Africa and Indonesia) to regime B. The

same is true for the small negative efficiency change.

Finally, we characterize regime C as the developing regime. It

contains many African countries, with very low output per capita, a

high degree of openness to trade, and a large primary sector share.

Output growth for these countries results from technical change,

efficiency change, and factor accumulation. The latter is fairly modest

and involves both capital and labor accumulation. Technical and

efficiency change are important, but as Table 7 shows, they have the

greatest effect on high-growth countries. From Fig. 1, we observe that

technical change in high-growth countries in this regime is both pure

and labor-augmenting. In contrast, low-growth countries in the

developing regime experience both pure and labor-augmenting

technical regress, which implies an inward shift of their production

possibility frontier. The latter may initially seem odd. However, the 66

observations in this low-growth subsample include primarily obser-

vations of Uganda and Rwanda in the early 1990s (which were

marked by various conflicts in The Great Lakes region), Congo at the

end of the same decade (the time of the Second Congo War), Zambia

(during 1986–1994 when it suffered one of the highest debt burdens

due to the collapse in the price of copper), and other sub-Saharan

countries in Africa that have suffered from various political, health,

and/or natural disasters during these years. Both the group allocation

of our model and the implication of the actual destruction of some

economies' production possibilities thus seems logical.

Recall that the share of the relatively unproductive and often

inefficient agricultural sector (Vollrath, forthcoming) is the highest in

regime C. The negative relationship between output per capita and

primary sector share across our regimes may reflect a dual economy

reallocation problem (Vollrath, forthcoming; Caselli, 2005). Labor in

the agricultural sector, which tends to be less than perfectly mobile,

may lead to factor market misallocations that seriously hamper

growth if labor productivity in this sector falls too low.

RegimeC is alsocharacterized bya highdegree of openness totrade,

which should contribute positively to growth such as through the

knowledge spillovers from importing and exporting (Edwards, 1998;

Ben-David and Loewy, 1998; Frankel and Romer, 1999). However, the

effects of trade on growth depend on the composition of exports in

particular. Hausmann et al. (2007) develop an index that measures the

‘quality’ of countries' export baskets and provide evidence that only

countries that produce and export high-productivity goods perform

better in terms of growth. Imports only enhance growth when they

include high-quality foreign capital goods, which embody advanced

foreign technology, and when anadequate level of humancapital exists

to perform reverse engineering and possibly improve on the imported

technology (Connolly, 1998). These two conditions rarely can be met in

low-income, developing countries. Furthermore, with market or

institutional imperfections, openness can lead to the under-utilization

of resources, concentration in extractive economic activities, or spe-

cialization away from technologically advanced, increasing-returns

sectors (Grossman and Helpman, 1991; Sachs and Werner, 1999;

Rodriguez and Rodrik, 2001).34Our results therefore support the

positiveeffectsofopennessongrowththroughtechnologyspilloversfor

high-growth countries, as well as the composition and market/

institutional imperfections arguments for low-growth countries in

regime C. We do not find unequivocal support for the export-led-

growth hypothesis in our sample. Cross-sectional studies that rely on

the‘average’countrythusappeartoobscuretheimportanceofdifferent

growth regimes and the potential heterogeneity within a regime.

The output growth decomposition thus leads to several important

conclusions. Though the forces that drive growth differ across regimes,

capital accumulation is consistently an important component of growth

across all the regimes. This finding is consistent with the results of

several prior studies (Koop et al., 1999, 2000; Koop, 2001; Limam and

Miller, 2004; Davis et al., 2007). However, it differs from Solow (1956)

who found that capital accumulation accounts for between one-eighth

and one-quarter of total output growth in the United States, whereas

34According to Grossman and Helpman (1991), a country may specialize in a non-

dynamic sector as a result of its openness, which causes it to lose out on the long-run

benefits of increasing returns. The underlying imperfection involves the contracts or

financial markets that induce people to follow a myopic notion of static comparative

advantages. Sachs and Werner (1999) also develop a model in which specialization in

extractive, natural-resource sectors diverts the economy from achieving technological

progress. In this case the underlying imperfection is the institutional weakness that

encourages natural-resource depletion for quick gains among only certain societal

groups. Finally, Rodriguez and Rodrik (2001) review the theoretical arguments

regarding why openness might be detrimental to developing countries.

Fig. 1. Mean output growth components of technical change and factor accumulation.

122

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 11

productivityaccounts for more thanhalf ofoutputgrowth inmost other

countries.35We find instead that growth in high- and medium-growth

countriesacrossallregimesresultsfromtechnicalchange,andgrowthin

low-growth countries is mainly driven by factor accumulation. There-

fore, we find support for the technical change-driven growth explana-

tion.Overall,ourresultsdonotsupportauniformexplanationofgrowth.

Theset-upofourmodelallowscountriestoswitchregimes.Asimple

waytoshedmorelightonthefactorsthatdetermineregimemigrations

is to test if the covariates we used to predict group membership differ

significantly between those countries that move and those that stay

within regimes. In Table 8 weindicate whether, for those countries that

change regimes, each of the conditioning variables is significantly

different than the values for the rest of the countries in the regime from

which the country has departed. We use two tests: a parametric t-test

and a non-parametric Kruskal–Wallis rank test. Table 8 reports the p-

values from these tests for each transition depicted in Table 5. The

numberofobservationsequalsthenumberofcountriesmoving(e.g.,42

moving from A to B) multiplied by the number of years in the period

prior to moving (e.g., 5 such that N=42×5=210). A positive

(negative) sign indicates that the conditioning variable is higher

(lower) than that of the other countries in the regime. For example, in

the second column of the first section of Table 8, the p-values of 0.0104

and 0.0042 and significance levels of 5% and 1% indicate that countries

that moved from regime A to regime C had significantly lower human

capital levels than did countries that remained in regime A.

From the last two columns of Table 8, we observe that increases in

financial development (for migrations to regime B) and human capital

(for migrations to regime A) are important strategies with which

countries can ‘escape’ from regime C. In addition, as countries in that

regime develop, their primary sector share decreases, and they may

become less dependent on exports (resulting in a lower openness to

tradeasitismeasuredhere).ThistoomakescountriesinregimeCmore

likely to move to migrate to regimes A and B. Countries that migrate

from regimeB toregimesA have a relatively low level of human capital.

In addition to human capital, a poor financial development and a

relatively large primary sector share contribute to migrations from

regime A and B to regime C.

Yet regime migrations are not the only way in which countries can

improve their performance. Within each regime, less technologically

advanced countries can grow faster than advanced ones because they

only need to copy the technology of the latter. This notion underlies

muchoftheconvergenceliterature.Intheirreviewoftheconvergence

literature, Durlauf et al. (2005) find that the estimation of conver-

gence rates can be improved by augmenting the Solow–Swan model

with human capital (Mankiw et al., 1992), by considering regional

convergence clusters (Mankiw, 1995; Quah, 1996), and by employing

econometric advancements such as panel data analysis (Islam, 1995).

Still, many studies continue to find either no or considerably different

convergence rates.

Apossibleexplanationforthispersistentconvergencepuzzlemaybe

the neglect of different regimes. Countries may converge (at different

rates) within but not necessarily across regimes, as we argue here. To

test for convergence, we replicate the modeling and estimation

approach of the seminal work of Mankiw et al. (1992). We only differ

fromtheoriginalanalysisofMankiwetal.(1992)inonerespect.Wenot

only test for convergence for the whole sample, but also for each of the

regimes identified by our conditional latent class model. To examine

convergence we estimate the following convergence equation from

Mankiw et al. (1992):

lny t ð Þ

y 0

ð Þ= 1− eλT

− 1 − eλT

??

α

1 − α

ð

?

Þln s ð Þ− 1− eλT

ln y0

ð

??

α

1 − α

ðÞn + g + δ

ðÞ

?

Þ;

ð6Þ

where y is output per worker, t and 0 indicate the end and start of the

respective period, s is the share of income saved and assumed to be

invested,andδ andgdenote theexogenous depreciation, respectively.

Following the convergence literature, we choose a joint rate for δ and

g of 6%.36The working population evolves at rate n, which we observe

from the data. According to the last term in Eq. (6), countries with

lower initial output per worker should grow faster. The pace of

convergence is implied by λ.

We estimate Eq. (6) for all countries during the whole sample

period, as in Mankiw et al. (1992), as well as for each five-year period,

as in Barro and Sala-i-Martin (1992). Subsequently, we repeat the

estimations for each regime.

The top panel of Table 9 reports the estimates of Eq. (6) for the full

sampleof77countriesduringtheperiod1970–2000andforthesixsub-

periods.Theinitialincomecoefficientisnegativeandinlinewiththeory.

The diagnostics (sample size and R2) indicate fairly good explanatory

power. For the whole sample period the annual convergence rate

implied by a λ of 2.4% is in line with previous evidence.37

Our conjecture that the existence of different regimes has

important implications for convergence also receives support from

the regime-specific parameters. Long-run estimates of initial income

coefficients, reported in the rightmost column of Table 9, imply

convergence rates that range from 2.1% in the fairly developed regime

B to 4.4% in the poorly performing regime C.38Since laggard countries

are predicted to converge faster, this illustrates that imposing the

assumption of an identical steady state underestimates convergence

ratesforsome,mostlylessdeveloped,countries.Thekeyfindingofthis

analysisis thatcountries convergeto their ownregime-specific steady

state and the rate of convergence differs from regime to regime.

Convergence results per regime across the whole sample period

are also subject to a caveat. Grouping countries to one regime for the

entire period requires allocations based on mean regime member-

ships. A country that appears in regime A for three intervals and then

three intervals in group C would be allocated to regime B.39Most

35Christensen and Cummings (1981), Dollar and Sokoloff (1990), King and Levine

(1994), and Kim and Lau (1996) report similar results.

Table 8

Migrating countries and their regime determinants.

Movement

A to BA to CB to AB to CC to AC to B

(N)(210)(25)(185)(10)(40)(30)

Human capital

Sign

t-test

Rank test

+

0.3239

0.1523

−

0.0104⁎⁎

0.0042⁎⁎⁎

−

0.0000⁎⁎⁎

0.0003⁎⁎⁎

−

0.0000⁎⁎⁎

0.0000⁎⁎⁎

+

0.0056⁎⁎⁎

0.9465

−

0.9765

0.3484

Financial development

Sign

t-test

Rank test

−

0.0841⁎

0.8435

−

0.0015⁎⁎⁎

0.0000⁎⁎⁎

−

0.6749

0.8185

−

0.0000⁎⁎⁎

0.0000⁎⁎⁎

−

0.5140

0.9233

+

0.0002⁎⁎⁎

0.0005⁎⁎⁎

Primary sector share

Sign

t-test

Rank test

−

0.1276

0.0062⁎⁎⁎

+

0.0003⁎⁎⁎

0.0001⁎⁎⁎

−

0.1293

0.3351

+

0.0000⁎⁎⁎

0.0000⁎⁎⁎

−

0.0000⁎⁎⁎

0.0000⁎⁎⁎

−

0.0378⁎⁎

0.0415⁎⁎

Openness to trade

Sign

t-test

Rank test

+

0.9494

0.4874

+

0.0000⁎⁎⁎

0.0000⁎⁎⁎

−

0.1817

0.0007⁎⁎⁎

+

0.8936

0.8955

+

0.7739

0.4589

−

0.0231⁎⁎

0.0227⁎⁎

Significance at the 10/5/1% levels (⁎/⁎⁎/⁎⁎⁎).

36Changing (δ+g) between 2.5% and 9.5% in increments of fifty basis points does

not alter our results qualitatively.

37Overall, the other coefficients are also consistent with theoretical predictions:

faster population growth reduces the growth of per worker income and higher savings

facilitate growth significantly.

38See Islam (1995), Caselli et al. (1996), and Jones (1997) for evidence.

39Results shown here are based on the arithmetic mean of the groups to which the

countries are allocated during the maximum of the six periods, as shown in Table A.3

in the Appendix A.

123

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 12

countries do not switch regimes, which can mitigate some of these

concerns. We can account more explicitly for possibly different

convergence speeds. To this end we estimate Eq. (6), similar to Barro

and Sala-i-Martin (1992), for each of the different episodes in the

latent class frontier model to identify the regimes.

The pertinent columns in Table 9 confirm the dispersion of

convergence rates.40In regime B there is convergence across most

periods at a pace of 2%. In contrast, the fully efficient and primarily

technological change-driven regime A only exhibits signs of strong

convergence during the 1980s. The absence of significant estimates of

convergence in each period in regime C likely reflects the low number

of observations.

Overall, aggregate convergence might be deceptive to the extent

that within-regime estimates differ significantly across both regimes

and time periods. Our results thus might explain why some studies

continuetofindmixedevidenceabouttheexistenceandmagnitudeof

convergence.Thatis,itmaybeduetotheirneglectofdifferentregimes.

5. Conclusion

Thestandardneoclassicalgrowthliteratureassumesthat:(i)coun-

tries use resources efficiently, and (ii) that the underlying production

technologyis the same for all countries.In this paper we address these

issues by estimating a stochastic frontier model augmented with a

latent class structure. Hence, we explicitly account for inefficiency and

allow for production technologies to differ across groups of countries.

In contrast with many cross-country growth studies, we estimate

membership in groups instead of determining ex ante which countries

should be compared.

Our empirical analysis is based on a sample of 77 countries over a

thirty-year period. The results support the existence of three regimes

of countries. First, a large, or mature regime that is comprised of many

mature economies, such as the U.S. and European countries. It

is characterized by high output per capita, high human capital

accumulation, high capital elasticity, and some level of inefficiency.

The second regime, the emerging regime, contains primarily emerg-

ing (developing) countries, mostly Asian, and is characterized by

productive labor, relatively well-developed financial system, and a

high efficiency level. Finally, the third regime, the developing regime,

includes the least developing countries, mostly African, and is

characterized by very low output per capita, a high degree of open-

ness to trade, a large primary sector share, and high inefficiency.

The driving forces of growth vary across regimes also. Growth in

the mature regime depends primarily on factor accumulation in

general and capital in particular, whereas the key generator of growth

in the emerging regime is (pure) technical change. Growth in the

developing regime depends on both (pure and labor-augmenting)

technical change and the accumulation of labor. Overall, our results

support a rather pluralistic explanation of growth experiences in the

countries in our sample. We also find evidence that input accumu-

lation is a reasonable description of the growth process for some

countries (in certain regimes) and of productivity (efficiency and

technology) developments for others.

Our findings strongly suggest several different growth processes,

which means that one-size-fit-all policy prescriptions based on

standard one-class results cannot prescribe the right medicine for

40The additional effects of saving and population and technical change, as well as

depreciation, differ too. Here we focus on the implications of different technology

regimes for income convergence.

Table 9

Convergence regressions.

Variable1970–1974 1975–19791980–19841985–19891990–19941995–2000 1970–2000

All

ln y0

ln (n+g+δ)

ln (s)

ln α

N

R2

Implied λ

−0.029

0.104

0.075⁎

0.423

45

0.060

0.006

−0.062⁎⁎⁎

0.052

0.050

0.471⁎

50

0.181

0.013

−0.085⁎⁎⁎

−0.319⁎

0.100

−0.674⁎⁎

58

0.198

0.018

−0.086⁎⁎⁎

−0.495⁎⁎⁎

0.130⁎⁎⁎

−1.234⁎⁎⁎

65

0.521

0.018

−0.024

−0.105⁎

0.143⁎⁎

−0.484⁎⁎⁎

74

0.259

0.005

0.005

−0.340⁎⁎⁎

−0.085⁎⁎⁎

−0.477⁎⁎⁎

76

0.196

−0.001

−0.528⁎⁎⁎

−0.532

1.056⁎⁎⁎

−2.119⁎⁎

77

0.563

0.024

Regime A

ln y0

ln (n+g+δ)

ln (s)

ln α

N

R2

Implied λ

−0.037

0.478⁎

0.153

1.204⁎

21

0.201

0.008

0.025

−0.221⁎⁎

−0.110

0.073

13

0.439

−0.005

−0.187⁎⁎⁎

−0.208

0.251⁎⁎⁎

−0.560

21

0.543

0.041

−0.141⁎⁎⁎

−0.694⁎⁎⁎

0.197⁎⁎

−1.737⁎⁎⁎

21

0.642

0.030

−0.053

−0.266⁎⁎⁎

0.074

−0.598⁎⁎⁎

27

0.502

0.011

−0.043

−0.392⁎⁎⁎

−0.027

−0.623⁎⁎

29

0.198

0.007

−0.663⁎⁎⁎

−0.998⁎⁎

0.962⁎⁎⁎

−2.664⁎⁎⁎

51

0.756

0.035

Regime B

ln y0

ln (n+g+δ)

ln (s)

ln α

N

R2

Implied λ

−0.126⁎⁎

0.032

0.139⁎⁎

0.216

16

0.373

0.027

−0.099⁎⁎⁎

0.019

0.043

0.480

26

0.491

0.021

−0.122⁎⁎⁎

−0.269

0.209⁎⁎⁎

−0.759⁎

26

0.445

0.026

−0.100⁎⁎⁎

−0.501⁎⁎⁎

0.121⁎

−1.194⁎⁎⁎

37

0.466

0.021

−0.085⁎⁎⁎

−0.121⁎⁎

0.220⁎⁎⁎

−0.568⁎⁎⁎

36

0.463

0.018

0.005

−0.417⁎⁎⁎

−0.040

−0.812⁎⁎

36

0.301

−0.001

−0.482⁎⁎⁎

−0.485

0.999⁎⁎⁎

−2.066

17

0.584

0.021

Regime C

ln y0

ln (n+g+δ)

ln (s)

ln α

N

R2

Implied λ

−0.293⁎

−1.580⁎⁎

0.164⁎⁎

−3.974⁎⁎

8

0.726

0.069

0.220

0.133

0.116

0.378

11

0.55

−0.040

−0.155

−2.863⁎

0.186

−7.307⁎

11

0.499

0.034

0.026

−0.384

0.081

−0.991

7

0.671

−0.005

−0.007

0.408⁎⁎

0.069

0.824

11

0.434

0.001

0.263

0.080

−0.158

0.546

11

0.443

−0.039

−0.747⁎⁎⁎

−1.302⁎⁎⁎

−0.001

−2.252⁎⁎

9

0.852

0.044

Ordinary least square estimates of Eq. (6); ln y0is the log of initial per capita output; the savings rate, s, is approximated by the investment share of GDP in fixed capital; δ+g=0.06;

n is the observed average annual growth of the working population; the number of observations is in parentheses. Significance at the 10/5/1% levels (⁎/⁎⁎/⁎⁎⁎).

124

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 13

Table A.1

Specification tests for determining the number of regimes.

Model Conditional Log likelihood

AIC BIC

3-regime latent class frontier

2-regime latent class frontier

3-regime latent class frontier

4-regime latent class frontier

No

Yes

Yes

Yes

1018.390

639.791

1119.540

no convergence

−1894.790

−1177.580

−2081.080

−1500.280

−894.205

−1642.120

Akaike Information Criterion (AIC)=ln AIC=2m−2ln LF(z), Schwartz Bayesian Information Criterion (SBIC)=−2ln LF(z)+m·ln(n); LF(j) is the likelihood value for Z groups, m is

the number of parameters in the model, and n the number of observations. The preferred specification has the lowest AIC or the lowest SBIC. See also Orea and Kumbhakar (2004).

Table A.2

Tests for equality of parameters between regimes.

Regimes Variable(s)TestStatistic p-valueHypothesis

Latent regime parameters

A and B

A and C

B and C

A, B and C

Translog production function

Translog production function

Translog production function

Translog production function

Wald

Wald

Wald

Wald

40.983

7.261

14.240

50.456

0.000

0.007

0.000

0.000

Rejected

Rejected

Rejected

Rejected

Ancillary parameters on production traits

A and B

A and C

B and C

A and B

A and C

B and C

Capital elasticity (εit|z

Capital elasticity (εit|z

Capital elasticity (εit|z

Labor elasticity (εit|z

Labor elasticity (εit|z

Labor elasticity (εit|z

K)

K)

K)

L)

L)

L)

t-test

t-test

t-test

t-test

t-test

t-test

−16.754

47.183

52.170

14.316

−10.326

−20.460

0.000

0.000

0.000

0.000

0.000

0.000

Rejected

Rejected

Rejected

Rejected

Rejected

Rejected

Regime membership probability parameters

A and B

A and B

A and B

A and B

A and B

Human capital (H)

Financial development (F)

Primary sector share (P)

Openness to trade (T)

H, F, P and T

Wald

Wald

Wald

Wald

Wald

0.057

144.980

0.020

6.360

183.553

0.811

0.000

0.888

0.012

0.000

Not rejected

Rejected

Not rejected

Rejected

Rejected

Null hypothesis tested at the 5% significance level is the equality of parameters between classes. Means and standard deviations for εit

standard errors for H, F, P, and T can be found in Table 2.

Kand εit

Lcan be found in Table 4. Coefficients and

Table A.3

Regime membership.

Period70–7475–7980–8485–8990–9495–00Period70–7475–7980–8485–8990–9495–00

CountryRegionCountryRegion

AfricaAmericas

Algeria

Benin

Botswana

Cameroon

Congo, Republic of

B

C

B

B

C

Argentina

Bolivia

Brazil

Canada

Chile

B

B

B

B

B

A

A

A

B

A

B

A

C

B

B

C

B

C

B

A

A

A

AAB

A

A

B

A

B

A

A

(continued on next page)

any country. More education alone, for instance, may put countries in

more advanced regimes, but for education to be effective other

development measures, such as enhanced factor allocation, financial

development, or trade policies, are required.

The presence of fairly persistent and economically significant

inefficiencies in the operations of best-practice technologies, espe-

cially in the regime with mostly developing countries, has important

policy implications. For countries in the developing regime, develop-

ment efforts geared toward developing the skills to exploit existing

technologies may be better than promoting the dissemination of new

technologies alone. Further research into the relative costs and bene-

fits of policies promoting either technical change or efficiency im-

provements is warranted.

This implication ties in with our finding regarding the dynamics

of regimes, which we investigate by accounting for regime migra-

tions. Few countries appear able to maintain their upgrades to faster

growing regimes. Because the determinants of group membership do

not differ significantly between countries that shift regimes and those

that stay, we believe that additional research into the determinants

of regime switches, rather than membership would be fruitful. Our

results show that most migrations pertain to countries switching

between the mature and the emerging regime. The intra-group

convergence patterns also reveal that countries from the mature and

emerging regimes primarily improve by catching up with the leader

countries in their own groups. Among the least developing economies

though, such catch-up patterns are absent. Both the periods and pace

ofconvergencediffer,at timessubstantially, rangingbetween1.8%and

4.1% per year. Thus, we find evidence of convergence to their own

regime for most countries, but our results support in particular

different convergence clubs around the world.

Appendix A

125

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 14

References

Abramovitz, M., 1986. Catching up, forging ahead, and falling behind. Journal of Economic

History 46 (2), 385–406.

Acemoglu,D.,Zilibotti, F.,2001.Productivity differences. QuarterlyJournalofEconomics

116 (2), 563–606.

Aigner, D.J., Lovell, K.C., Schmidt, P., 1977. Formulation and estimation of stochastic

frontier production function models. Journal of Econometrics 6 (1), 21–37.

Auerbach, A.J., Hassett, K.A., Oliner, S.D., 1994. Reassessing the social returns to

equipment investment. Quarterly Journal of Economics 109 (3), 789–802.

Azariadis, C., Drazen, A., 1990. Threshold externalities in economic development. The

Quarterly Journal of Economics 105 (2), 501–526.

Baltagi, B.H., Griffin, J.M., 1988. A general index of technical change. Journal of Political

Economy 96 (1), 20–41.

Barro, R.J., 1991. Economic growth in a cross section of countries. Quarterly Journal of

Economics 106 (2), 407–443.

Barro, R.J., Lee, J.-W., 2001. International data on educational attainment: updates and

implications. Oxford Economic Papers 53 (3), 541–563.

Barro, R.J., Sala-i-Martin, X., 1992. Convergence. Journal of Political Economy 100 (2),

223–251.

Basu,S.,Weil,D.,1998.Appropriatetechnologyandgrowth.QuarterlyJournalofEconomics

113 (4), 1025–1054.

Battese,G.E., Corra, G.S., 1977. Estimation of a production frontiermodel, with application

to the pastoral zone of Eastern Australia. Australian Journal of Agricultural Economics

21 (3), 169–179.

Beck, T., Demirgüç-Kunt, A., Levine, R., 2000. A new database on financial development

and structure. World Bank Economic Review 14, 597–605.

Ben-David, D., Loewy, M.B., 1998. Free trade, growth, and convergence. Journal of

Economic Growth 3 (2), 143–170.

Benhabib, J., Spiegel, M.M., 1994. The role of human capital in economic development:

evidence from aggregate cross-country data.Journal of MonetaryEconomics 34(2),

143–173 Oct.

Bernard, A.B., Jones, C.I., 1996a. Productivity across industries and countries: time series

theory and evidence. Review of Economics and Statistics 78 (1), 135–146.

Bernard, A.B., Jones, C.I., 1996b. Technology and convergence. Economic Journal 106

(437), 1037–1044.

Bloom, D., Canning, D., Sevilla, J., 2002. Technological diffusion, conditional conver-

gence, and economic growth. NBER Working Paper No. 8713. National Bureau of

Economic Research, Cambridge, Mass.

Bloom, D.E., Canning, D., Sevilla, J., 2003. Geography and poverty traps. Journal of

Economic Growth 8 (4), 355–378.

Brock, W.A., Durlauf, S.N., 2001. What have we learned from a decade of empirical

research on growth? growth empirics and reality. World Bank Econ Rev 15 (2),

229–272.

Cameron, G., Proudman, J., Redding, S., 2005. Technological convergence, R&D, trade

and productivity growth. European Economic Review 49 (3), 775–807 Apr.

Cannon, E., 2000. Human capital: level versus growth effects. Oxford Economic Papers

52 (4), 670–676.

Canova, F., 2004. Testing for convergence clubs in income per capita: a predictive

density approach. International Economic Review 45 (1), 49–77.

Caselli, F., 2005. Accounting for cross-country income differences. In: Handbook of

Economic Growth, vol. 1. Elsevier, pp. 679–741.

Caselli, F., Gerardo, E., Fernando, L., 1996. Reopening the convergence debate: a new

look at cross-country growth empirics. Journal of Economic Growth 1 (3), 363–389.

Christensen, L.R., Cummings, D., 1981. Real product, real factor input, and productivity in

theRepublicofKorea,1960–1973.JournalofDevelopmentEconomics8(3),285–302.

Coelli, T., Rao, D.P., Battese, G.E., 2005. An Introduction to Efficiency Analysis. 2nd

edition. Springer, New York.

Connolly, M.P., 1998. The dual nature of trade: measuring its impact on imitation and

growth. Federal Reserve Bank of New York, Staff Reports No. 44.

Davis, L.S., Owen, A.L., Videras, J., 2007. Do all countries follow the same growth

process? Mimeo (SSRN eLibrary).

Desdoigts, A., 1999. Patterns of economic development and the formation of clubs.

Journal of Economic Growth 4 (3), 305–330.

Dollar, D., Kraay, A., 2004. Trade, growth, and poverty. The Economic Journal 114,

F22–F49.

Dollar, D., Sokoloff, K., 1990. Patterns of productivity growth in South Korean

manufacturing industries, 1963–1979. Journal of Development Economics 33 (2),

309–327 Oct.

Table A.3 (continued)

Period 70–74 75–7980–84 85–8990–9495–00Period 70–74 75–79 80–8485–89 90–9495–00

Country RegionCountry Region

Egypt

Gambia, The

Ghana

Kenya

Lesotho

Malawi

Mali

Mauritania

Mauritius

Mozambique

Rwanda

Senegal

South Africa

Togo

Tunisia

Uganda

Zambia

A

C

C

C

C

A

C

C

C

C

C

A

C

B

C

C

C

C

A

A

B

B

C

C

C

C

B

A

C

B

A

C

A

C

C

A

C

B

B

C

C

C

Colombia

Costa Rica

Dominican Republic

Ecuador

El Salvador

Guatemala

Guyana

Honduras

Mexico

Panama

Paraguay

Peru

Trinidad and Tobago

United States

Uruguay

Venezuela

A

B

B

A

B

B

A

B

B

B

A

A

A

A

B

B

A

A

A

B

B

B

A

A

A

B

B

B

A

A

A

B

A

B

A

A

C

B

B

B

B

B

A

B

A

B

C

C

C

A

B

C

A

A

A

A

B

B

A

B

B

B

A

B

B

B

B

A

A

A

B

B

B

A

B

AAA

A

C

B

A

C

A

C

C

C

C

A

C

A

C

B

C

B

C

A

C

B

B

A

A

B

C

C

BB

B

B

A

B

BB

CAB

Europe

B

B

B

A

A

Asia

Austria

Belgium

Denmark

Finland

France

Germany

Greece

Hungary

Iceland

Italy

Netherlands

Norway

Poland

Portugal

Spain

Sweden

Switzerland

United Kingdom

B

B

B

B

A

B

B

A

A

B

B

B

B

B

B

B

B

B

A

B

B

B

B

C

B

B

A

B

A

B

B

A

A

B

A

B

B

A

A

B

B

A

B

A

B

B

B

B

B

B

B

Bangladesh

Hong Kong

India

Indonesia

Iran

Japan

Jordan

Malaysia

Pakistan

Philippines

Sri Lanka

Syria

Thailand

Turkey

AA

A

B

A

A

A

B

A

A

B

B

A

B

A

A

A

A

A

A

B

A

A

C

B

C

B

B

B

C

A

A

B

B

B

B

B

B

A

B

A

A

B

A

B

B

A

B

A

A

A

B

A

B

A

A

B

B

A

B

A

ABA

B

C

B

B

B

B

B

A

B

B

B

A

C

A

A

B

A

B

B

A

A

B

B

CCC

A

A

B

B

B

B

B

A

B

B

B

B

Oceania

BAustralia

New Zealand

B

B

B

B

B

B

B

B

A

B

ABAA

Notes: most likely regime membership allocations per five-year period to the emerging (A), mature (B), and developing (C) groups. Probabilities obtained from Eq. (4) conditional

on period averages of human capital, financial development, primary sector share, and openness to trade. Total (annual) observations per regime A, B, and C, respectively, are 689,

919, and 305.

AfricaAmericas

126

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127

Page 15

Durlauf,S.N.,Johnson,P.A.,1995.Multipleregimesandcross-countrygrowthbehaviour.

Journal of Applied Econometrics 10 (4), 365–384.

Durlauf, S.N., Johnson, P.A., Temple, J.R., 2005. Growth econometrics. In: Handbook of

Economic Growth, vol. 1. Elsevier, pp. 556–677. Ch. 8.

Easterly, W., 1999. The ghost of financing gap: testing the growth model used in the

international financial institutions. Journal of Development Economics 60 (2),

423–438.

Edwards, S., 1998. Openness, productivity and growth: what do we really know? The

Economic Journal 108 (447), 383–398.

Eicher, T.S., 1999. Trade, development and converging growth rates: dynamic gains

from tarde reconsidered. Journal of International Economics 48 (1), 179–198.

Färe, R., Grosskopf, S., Norris, M., Zhang, Z., 1994. Productivity growth, technical

progress, and efficiency change in industrialized countries. American Economic

Review 84 (1), 66–83.

Frankel, J.A., Romer, D., 1999. Does trade cause growth? American Economic Review

89 (3), 379–399.

Greene, W.H., 2002a. Alternative Panel Data Estimators for Stochastic Frontier Models.

Mimeo (available at http://pages.stern.nyu.edu/ wgreene/).

Greene, W.H., 2002b. Econometric Modeling Guide. Econometric Software, Inc., New

York.

Greene, W.H., 2005. Reconsidering heterogeneity in panel data estimators of the

stochastic frontier model. Journal of Econometrics 126 (2), 269–303.

Greenwood, J., Jovanovic, B., 1990. Financial development, growth, and the distribution

of income. Journal of Political Economy 98 (5), 1076–1107.

Grossman, G., Helpman, E., 1991. Innovation and Growth in the Global Economy. MIT

Press, Cambridge, MA and London.

Hall, R.E., Jones, C.I., 1999. Why do some countries produce so much more output per

worker than others? Quarterly Journal of Economics 114 (1), 83–116.

Hausmann,R.,Hwang,J.,Rodrik,D.,2007.Whatyouexportmatters.JournalofEconomic

Growth 12 (1), 1–25.

Hobijn, B., Franses, P.H., 2000. Asymptotically perfect and relative convergence of

productivity. Journal of Applied Econometrics 15 (1), 59–81.

Islam,N., 1995.Growth empirics:a paneldata approach. QuarterlyJournalof Economics

110 (4), 1127–1170.

Jerzmanowski, M., 2006. Empirics of hills, plateaus, mountains and plains: a Markow-

switching approach to growth. Journal of Development Economics 81(2), 357–385.

Jondrow, J., Knox Lovell, C.A., Materov, I.S., Schmidt, P., 1982. On the estimation of

technical inefficiency in the stochastic frontier production function models. Journal

of Econometrics 19 (2–3), 233–238.

Jones, C., 1997. On the evolution of the world income distribution. Journal of Economic

Perspectives 11 (3), 19–36.

Jones, C.I., 2005. The shape of production functions and the direction of technical

change. Quarterly Journal of Economics 120 (2), 517–549.

Kejak, M., 2003. Stages of growth in economic development. Journal of Economic

Dynamics and Control 27 (5), 771–800.

Kim, J.I., Lau, L.J., 1996. The sources of Asian Pacific economic growth. The Canadian

Journal of Economics / Revue Canadienne d'Economique 29, S448–S454.

King, R., Levine, R., 1993. Finance and growth: Schumpeter might be right. Journal of

Monetary Economics 32 (3), 513–542.

King, R.G., Levine, R., 1994. Capital fundamentalism, economic development, and

economic growth. Carnegie-Rochester Conference Series on Public Policy 40 (1),

259–292.

Kneller, R., Stevens, P., 2006. Frontier technology and absorptive capacity: evidence

from OECD manufacturing industries. Oxford Bulletin of Economics and Statistics

68 (1), 1–21.

Koop, G., 2001. Cross-sectoral patterns of efficiency and technical change in

manufacturing. International Economic Review 42 (1), 73–103.

Koop, G., Osiewalski, J., Steel, M.F., 1999. The components of output growth: a stochastic

frontier analysis. Oxford Bulletin of Economics and Statistics 61 (4), 455–487.

Koop, G., Osiewalski, J., Steel, M.F., 2000. Modeling the sources of output growth in a

panel of countries. Journal of Business and Economic Statistics 18 (3), 284–299.

Krueger, A.B., Lindahl, M., 2001. Education for growth: why and for whom? Journal of

Economic Literature 39 (4), 1101–1136.

Krueger, A.O., 1998. Why trade liberalization is good for growth. The Economic Journal

108, 1513–1522.

Kumar, S., Russell, R.R., 2002. Technological change, technological catch-up, and capital

deepening: relative contributions to growth and convergence. American Economic

Review 92 (3), 527–548.

Kumbhakar, S.C., Lovell, K.C., 2000. Stochastic Frontier Analysis. Cambridge University

Press, Cambridge.

Levine, R., Renelt, D., 1992. A sensitivity analysis of cross-country growth regressions.

American Economic Review 82 (4), 942–963.

Limam, Y.R., Miller, S.M., 2004. Explaining Economic Growth: Factor Accumulation,

Total Factor Productivity Growth, and Production Efficiency Improvement.

University of Connecticut, Department of Economics, Working Paper Series No.

2004-20. Forthcoming in Quarterly Review of Economics and Finance.

Los, B., Timmer, M.P., 2005. The appropriate technology explanation of productivity

growth differentials: an empirical approach. Journal of Development Economics

77 (2), 517–531.

Lucas, R.E., 1988. On the mechanics of economic development. Journal of Monetary

Economics 22 (1), 3–42.

Mankiw, G.N., 1995. The growth of nations. Brookings Papers on Economic Activity 1,

276–326.

Mankiw, G.N., Romer, D., Weil, D.N., 1992. A contribution to the empirics of economic

growth. Quarterly Journal of Economics 107 (2), 407–437.

Meeusen, W., van den Broeck, J., 1977. Efficiency estimation from Cobb–Douglas

production functions with composed error. International Economic Review 18 (2),

435–444.

Nelson, R., Phelps, E., 1966. Investment in humans, technological diffusion, and economic

growth. American Economic Review: Papers and Proceedings 56 (1–2), 69–75.

Orea, L., Kumbhakar, S.C., 2004. Efficiency measurement using a latent class stochastic

frontier model. Empirical Economics 29 (1), 169–183.

Paap, R., Franses, P.H., van Dijk, D., 2005. Does Africa grow slower than Asia, Latin

America and the Middle East? Evidence from a new data-based classification

method. Journal of Development Economics 77 (2), 553–570.

Pack, H., Page, J.M., 1994. Reply to Alwyn Young. Carnegie-Rochester Conference Series

on Public Policy 40, 251–257.

Pack, H., Saggi, K., 2001. Vertical technology transfer via international outsourcing.

Journal of Development Econmics 65 (2), 389–415.

Pagano, M., 1993. Financial markets and growth: an overview. European Economic

Review 37 (2–3), 613–622.

Papageorgiou, C., 2002. Trade as a threshold variable for multiple regimes. Economics

Letters 77 (1), 85–91.

Persson, T., Tabellini, G., 1994. Is inequality harmful for growth? American Economic

Review 84 (3), 600–621.

Quah, D., 1996. Regional convergence clusters across Europe. European Economic

Review 40, 951–958.

Redding, S., 1996. The low-skill, low-quality trap: strategic complementarities between

human capital and R & D. The Economic Journal 106 (435), 458–470.

Rodriguez,F.,Rodrik,D.,2001.In:Bernanke,B.,Rogoff,K.(Eds.),Tradepolicyandeconomic

growth: a sceptic's guide to the cross-national evidence. NBER Macroeconomics

Annual 2000, vol. 15. MIT Press, Cambridge MA, pp. 261–325.

Romer, P.M., 1989. Human capital and growth: theory and evidence. National Bureau of

Economic Research Working Paper Series No. 3173.

Sachs, J.D., Werner, A., 1999. The big push, natural resource booms and growth. Journal

of Development Economics 59 (1), 43–76.

Schumpeter, J., 1934. The Theory of Economic Development. Harvard University Press,

Cambridge, MA, USA.

Sianesi, B., Reenen, J.V., 2003. The returns to education: macroeconomics. Journal of

Economic Surveys 17 (2), 157–200.

Sirimaneetham, V., Temple, J., 2006. Macroeconomic policy and the distribution of

growth rates. Tech. Rep. 06/584. Department of Economics, University of Bristol.

Solow, R.M., 1956. A contribution to the theory of economic growth. Quarterly Journal

of Economics 70 (1), 65–94.

Solow, R.M., 1957. Technical change and the aggregate production function. Review of

Economics and Statistics 39 (3), 312–320.

Solow, R.M., 1994. Perspectives on growth theory. Journal of Economic Perspectives

8 (1), 45–54.

Tallman, E.W., Wang, P., 1994. Human capital and endogenous growth: evidence from

Taiwan. Journal of Monetary Economics 34 (1), 101–124.

Temple,J.,1999.Thenewgrowthevidence.JournalofEconomicLiterature37(1),112–156.

Temple, J., 2005. Dual economy models: a primer for growth economists. The

Manchester School 73 (4), 435–478.

Trefler, D., 1993. International factor price differences: Leontief was right! Journal of

Political Economy 101 (6), 961–987.

Tsionas, E.G., Kumbhakar, S.C., 2004. Markov switching stochastic frontier model.

Econometrics Journal 7 (2), 398–425.

Vollrath, D., forthcoming. How important are dual economy effects for aggregate

productivity? Journal of Development Economics.

World Development Indicators, 2006. (WDI). World Bank.

Young, A., 1991. Learning by doing and the dynamic effects of international trade. The

Quarterly Journal of Economics 106 (2), 369–405.

Young, A., 1994. Lessons from the East Asian NICS: a contrarian view. European Economic

Review 38 (3–4), 964–973.

127

J.W.B. Bos et al. / Journal of Development Economics 91 (2010) 113–127