augmented solow model essay, 1300 words

Journal of Comparative Economics 37 (2009) 432–452

Contents lists available at ScienceDirect

Journal of Comparative Economics

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / j c e

Can the augmented Solow model explain China’s remarkable economic growth? A cross-country panel data analysis

Sai Ding *, John Knight University of Oxford, Manor Road Building, Oxford OX1 3UQ, UK

a r t i c l e i n f o a b s t r a c t

Article history: Received 17 July 2008 Revised 24 March 2009 Available online 8 June 2009

JEL classification: O11 O47

Keywords: China Economic growth Augmented Solow model Cross-country growth regression Structural change

0147-5967/$ - see front matter � 2009 Association doi:10.1016/j.jce.2009.04.006

* Corresponding author. Fax: +44 20 704 39951. E-mail addresses: [email protected] (S

1 Influential work on the (un)reliability of China’s G thank an anonymous referee for raising this point.

2 The figure is calculated from Ravallion and Chen

Ding, Sai, and Knight, John—Can the augmented Solow model explain China’s remarkable economic growth? A cross-country panel data analysis

China’s economy grew at an average annual rate of 9% over the last three decades. Despite the vast empirical literature on testing the neoclassical model of economic growth using data on various groups of countries, very few cross-country regressions include China and none of them particularly focuses on the explanation of China’s remarkable economic growth. We attempt to fill this gap by utilising panel data on 146 countries over the period 1980–2004 to examine the extent to which the rapid growth of China and the huge gap in the growth rate between China and other countries can be explained by the augmented Solow model. Using system GMM estimation techniques, we find that, in spite of the restrictive assumptions involved, the Solow model augmented by both human capital and structural change provides a fairly good account of international variation in economic growth. In particular, China’s relative success in economic growth is due to high physical capital investment, conditional convergence gain, dramatic changes in the structure of employment and output, and low population growth. Journal of Comparative Economics 37 (3) (2009) 432-452. University of Oxford, Manor Road Building, Oxford OX1 3UQ, UK. � 2009 Association for Comparative Economic Studies. Published by Elsevier Inc. All rights

reserved.

1. Introduction

Few countries have been able to match the pace of China’s sustained economic growth. Since the start of economic reform in 1978, China has maintained a remarkable growth rate. According to official Chinese statistics, the average GDP growth rate during the period 1979–2000 was 9.2%, and it accelerated to 10.1% between 2000 and 2006. Despite the controversy over the reliability of the official figures of real output growth,1 the fact that China has grown fast is beyond dispute. Being the world’s largest developing country, with one-fifth of the world population, China’s growth has contributed significantly to the reduction of global income poverty and inequality by lifting over 300 million people out of one-dollar-a-day poverty in the past three decades.2

While China’s economic growth over the reform period has received much attention in the economic literature, research has focused mainly on relatively narrow topics such as issues of growth convergence or divergence among provinces, deter- minants of cross-provincial growth variation, and assessment of the sustainability of China’s growth, all of which are based

for Comparative Economic Studies. Published by Elsevier Inc. All rights reserved.

. Ding), [email protected] (J. Knight). DP statistics includes Maddison (1998), Rawski (2001), Lardy (2002), Young (2003), and Holz (2006). We

(2007).

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 433

on data for China only. Although these studies are crucial for understanding the growth patterns within China, they can only hint at why China as a whole could grow so rapidly and at the key factors driving the persistent growth disparity between China and other countries.

This paper attempts to fill the gap in the literature by incorporating China into a cross-country growth study based on a neoclassical framework. The novelty and original contribution of the study lie in the following three aspects. Firstly, the empirical growth model we adopt is a revised cross-country specification of the augmented Solow model which allows for cross-country differences in productivity growth as measured by structural change. Unlike the growth accounting ap- proach, it is a useful means of estimating growth in aggregate efficiency without using capital stock data, thus avoiding mak- ing strong assumptions about unknown parameters of production functions. The structural model used in this paper also reduces the problem of model uncertainty faced by informal growth regressions.

Secondly, concerning the econometric methodology, we extend the previous cross-section work of Temple and Wößmann (2006) to the dynamic panel data context, in which the unobserved country-specific effects and potential endogeneity and measurement error of regressors are controlled for. By classifying countries at similar levels of development into the same sample, we control partially for the differences in technology and institutions and alleviate the problem of parameter het- erogeneity. Robust regression techniques are used to isolate the influence of potential outliers so that we are able to concen- trate on the most coherent part of the dataset. Our efforts in jointly dealing with the problems of omitted variables, parameter heterogeneity, measurement error, endogenous regressors and influential outliers provide new insights on empir- ical estimation of cross-country growth regressions.

Thirdly, to explain China’s relative growth success, we use the estimated equation to decompose the sources of differ- ences in growth rates between China and other countries. To our knowledge, this is the first attempt to explain China’s exceptional growth performance by means of cross-country growth regressions.

We find that the Solow model augmented by human capital and structural change predicts China’s economic growth rate quite accurately, and that there are four main determinants of China’s extraordinary growth performance. Capital formation has played a major role in China’s economic growth, and this view of investment-driven growth is consistent with the out- of-equilibrium neoclassical growth theory and in line with explanations for the East Asian ‘miracle’ (Krugman, 1994; Young, 1995). Conditional convergence contributes significantly to growth differences between China and other countries. Eco- nomic growth has been intertwined with productivity-enhancing structural change throughout the reform period. Lastly, the low population growth rate resulting from the restrictive population policy makes an important contribution to China’s growth performance relative to many other developing countries.

The paper is organised as follows: Section 2 provides some background on China’s economic reform and places its growth in comparative perspective. Section 3 briefly summarises the neoclassical growth theory and its empirical formulation in a cross-country growth context. Section 4 discusses econometric methodology in estimating cross-country growth regressions and describes the data and sample classification. Section 5 interprets the estimation results and explains China’s growth on the basis of the model predictions. Section 6 draws conclusions.

2. Background to China’s remarkable economic growth

2.1. China’s gradualist approach to reform

The growth of the Chinese economy since the start of its economic reform has been a process of ‘crossing the river by groping for the stepping stones’, as described by Deng Xiaoping: no stereotype reform package was adopted in advance. One reform begat the need, or the opportunity, for another, and the process became cumulative. The reforms were incremen- tal but hardly slow: huge changes have occurred in less than three decades, as China has moved from central planning to- wards a market economy. It is relevant that, when reform began, China had a labour surplus economy par excellence: labour was underemployed in the farms and in the urban state enterprises: government preferred unemployment to be disguised and shared rather than open and threatening (Knight and Song, 2005, chapters 2, 6, and 8). New sectors could thus be ex- panded without loss of output elsewhere.

The first stage of economic reform (1978–1984) concentrated on the rural areas. The communes were disbanded and individual incentives were restored. Farming households (then 82% of the population) were given use-rights to collec- tively-owned land under long term leases, and the right to sell their marginal produce on the open market. Rural non-farm enterprises were permitted, and they stepped in to produce the light manufactures that the urban state-owned enterprises (SOEs) generally failed to supply. Rural credit constraints encouraged household saving. Rural production rose rapidly as farms became more efficient, as surplus labour was used more productively in rural industry, and as rural entrepreneurship, saving and investment responded to the new opportunities.

The second stage of economic reform (1985–1992) was an incremental process of reforming the urban economy, in par- ticular the SOEs, which were gradually given greater managerial autonomy. The principal-agent problem inherent in state ownership limited the efficiency of SOEs but competition from other market participants – initially village and township enterprises and later domestic and foreign privately owned enterprises as well as from imports – grew steadily.

The third stage of economic reform (1993–) was ignited by Deng Xiaoping’s ‘Southern Tour’ to mobilise support for more radical reforms. The private sector – for the first time acknowledged and accepted – was invigorated. Moreover, administrative and regulatory reform of rural–urban migration, the banking system, the tax system, foreign trade, and

434 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

foreign investment lifted various binding constraints on economic growth. For instance, when the delayed effects of the ’one-child family policy’ slowed down the growth of the urban-born labour force from the mid-1990s onwards, the relaxa- tion of restrictions on temporary rural–urban migration permitted continued rapid growth of the urban economy.

Fig. 1 reflects China’s rapid growth of GDP per capita, averaging 8.5% per annum over the period 1978–2005. The figure also shows a cyclical pattern of growth, more marked in the first and second stages of reform than in the third stage. Two peaks are evident, in 1984–1985 and 1992–1993, respectively, reflecting the outcome of agricultural reforms and the green light given to capitalism. The growth rate troughed in 1989–1990 owing to a surge of inflation and social unrest.

This background information moulds the hypotheses for our study. The reforms created institutions and incentives that had been lacking in the socialist planned economy. They improved both static allocative efficiency and dynamic factor accu- mulation. Growth was also facilitated by the absorption of the abundant resource, labour, into the expanding, more produc- tive, activities. There was drastic movement towards the economy’s production frontier and dramatic movement of the frontier. It is plausible that together they were responsible for China’s remarkably high rate of economic growth over the reform period.

2.2. China’s economic growth in comparative perspective

We compare China with the main regions of the world economy in Table 1. We do so in terms of variables that suggest hypotheses for testing. The table provides information at 10-year intervals over the period 1980–2000, the average for that period, and the change between 1980 and 2000. For China, as for other countries, our measure of GDP is based on the World Bank’s constant price (year 2000) US dollar equivalents, rather than official Chinese statistics, for the purpose of international comparison.

China’s annual average growth of GDP per capita over the 20 years (8.4%) is four times that of the high-income economies (2.1%), and is much higher than that of Sub-Saharan Africa and Latin America and the Caribbean (�0.6% and 0.6%, respec- tively). China’s sustained growth rate is indeed remarkable.

In 1980, China had a lower level of GDP per capita than any of the regions included in the table, although by 2000 it had overtaken South Asia and Sub-Saharan Africa. The intuition is that China was initially further away from its equilibrium GDP per capita, and that forces of convergence would thus enable it to grow relatively fast. This hypothesis requires testing.

China’s growth performance has been associated with an extremely high investment rate. Gross capital formation as a proportion of GDP (averaging 36%) is remarkable for such a poor country, reflecting high household and enterprise saving rates and large capital inflows. We see that the four other regions managed to invest only about 20% of their GDP. A large part of the answer to the question ‘why does China grow so fast?’ might be ‘because it invests so much’. That must be a core hypothesis of our enquiry.

Rapid economic growth was inevitably associated with rapid structural change in the Chinese economy, as industrialisa- tion proceeded. There is a huge rural–urban income divide in China, which provides a strong incentive for rural workers to migrate out of agriculture (Knight and Song, 1999). We calculate that even in 2000 the ratio of GDP per worker in non-agri- culture to that in agriculture was 4.9 in China, higher than in any other region except Sub-Saharan Africa (5.7). The share of agriculture in GDP fell from 30% (higher than elsewhere) in 1980 to only 15% in 2000. The fall (by 15% points) was greater than in other regions; indeed, beyond South Asia, the fall was less than 4% points in the slow-growing regions of the world. The change in China’s sectoral composition of output involved the reallocation of labour from low average labour produc- tivity (and possibly zero marginal productivity) agriculture to high productivity industry.

China has implemented a draconian population policy since the late 1970s. Despite the controversy over the humanity of the ‘one-child family policy’, it has been efficient in reducing fertility and slowing down the rate of population growth. This reduced the pressure on the land and on other scarce resources. By contrast, other regions of the developing world have experienced higher rates of population growth. We hypothesise that China’s growth of GDP per capita benefited from its restrictive population policy.

Data source: World Bank, World Development Indicators 2006 Database.

0 2 4 6 8

10 12 14 16

P er

ce nt

ag e

(% )

Fig. 1. China’s annual GDP per capita growth rates.

Table 1 International comparison of key variables

1980 1990 2000 Average during 1980–2000 Change between 1980 and 2000

GDP per capita growth rate per annum (%) China 6.46 2.29 7.64 8.43 1.18 South Asia 3.87 3.42 2.36 3.33 �1.50 Sub-Saharan Africa 1.07 �1.76 0.79 �0.59 �0.28 Latin America and the Caribbean 3.87 �1.39 2.41 0.57 �1.46 High-income economies 0.46 2.26 2.83 2.06 2.37

GDP per capita per annum (constant 2000 $) China 186.44 391.65 949.18 476.44 762.74 South Asia 235.32 327.86 449.60 329.11 214.28 Sub-Saharan Africa 589.60 530.75 515.38 528.72 �74.22 Latin America and the Caribbean 3565.73 3258.70 3852.41 3481.26 286.68 High-income economies 17304.14 21916.68 26368.33 21419.46 9064.19

Share of gross capital formation in GDP (%) China 35.19 34.74 32.76 36.04 �2.43 South Asia 18.73 22.83 23.53 21.98 4.80 Sub-Saharan Africa 24.76 17.75 17.28 18.99 �7.48 Latin America and the Caribbean 24.54 19.39 21.07 20.91 �3.47 High-income economies 24.62 22.94 22.03 22.25 �2.59

Share of agriculture in GDP (%) China 30.09 27.05 14.83 24.35 �15.26 South Asia 37.15 30.67 24.16 30.90 �12.98 Sub-Saharan Africa 18.72 19.61 18.49 19.61 �0.23 Latin America and the Caribbean 10.16 8.97 6.67 9.01 �3.49 High-income economies 3.97 2.81 1.79 2.84 �2.18

Population growth rate per annum (%) China 1.25 1.47 0.71 1.06 �0.54 South Asia 2.46 2.14 1.83 2.08 �0.63 Sub-Saharan Africa 3.11 2.88 2.49 2.77 �0.62 Latin America and the Caribbean 2.31 1.84 1.51 1.87 �0.80 High-income economies 0.84 0.84 0.82 0.76 �0.02

Average years of schooling over age 15 (year) China 4.77 5.85 6.36 5.61 1.59 South Asia 2.48 3.24 3.76 3.13 1.28 Sub-Saharan Africa 2.24 2.93 3.40 2.89 1.16 Latin America and the Caribbean 4.86 5.54 6.18 5.53 1.32 High-income economies 7.82 8.64 9.30 8.58 1.48

Average annual growth rate of average years of schooling (%) China 1.66 3.36 0.78 1.48 �0.87 South Asia 6.31 3.73 1.96 3.25 �4.35 Sub-Saharan Africa 3.48 2.97 1.27 2.47 �2.22 Latin America and the Caribbean 2.65 1.41 0.93 1.56 �1.72 High-income economies 2.01 1.36 0.71 1.18 �1.31

Data source: Human capital variables are from Barro and Lee (2001); and other variables are from World Bank, World Development Indicators 2006 Database.

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 435

Human capital can raise the individual productivity of workers and improve the adaptability, allocative efficiency, and technical level of an economy. Based on the Barro and Lee (2001) data on international educational attainment, we find that China’s average years of schooling in the population aged over 15 (5.6 years) was much lower than that of high-income econ- omies (8.6 years), but higher than that of South Asia (3.1 years) and Sub-Saharan Africa (2.9 years), and on a par with that of Latin America and the Caribbean (5.5 years) over the period 1980–2000. The pattern of annual growth rate of average years of schooling shows opposite results: the average annual growth rate of China (1.5%) was faster only than that of high-income economies (1.2%) and slower than those of other developing country groups. Therefore, we expect that China’s rapid eco- nomic growth relative to other developing countries is partly due to the level of education, while that relative to the high-income economies can be partly explained by the growth rate of human capital over the reform period.

3. The Solow model in cross-country growth regressions

There is an enormous literature on cross-country growth research. One type of cross-country growth regressions is based on structural growth models. In their influential work, Mankiw et al. (1992), hereafter MRW, found that a neoclassical Solow model with exogenous technology and diminishing returns to capital provides an excellent explanation for international in- come disparities, i.e. about 80% of the cross-country variation in income per capita can be explained by accumulation of

436 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

human and physical capital. Their cross-section analysis has been extended to a panel data framework by Islam (1995), Cas- elli et al. (1996), Bond et al. (2001), and Hoeffler (2002) using various estimation methods to deal with the problems of omit- ted variables, endogenous regressors and measurement error. By contrast, Lichtenberg (1993), Howitt (2000), and Bond et al. (2007) found evidence for endogenous growth models in which capital is an input to the production technology for innova- tions and long-run differences in productivity are endogenous.

A major criticism of structural growth models is that variables that plausibly affect growth are left out. Another approach is to use ad hoc regressions incorporating all relevant variables. Such informal growth regressions have been referred to as ‘Barro-type regressions’ after Barro (1991). In the absence of a formal theoretical derivation, the reduced-form growth regressions are subject to the problem of model uncertainty, which is acknowledged by Levine and Renelt (1992), Mankiw et al. (1995), Sala-i-Martin (1997), Fernández et al. (2001), Hendry and Krolzig (2004), and Hoover and Perez (2004) using one or more of extreme bounds analysis, Bayesian model averaging and general-to-specific methods.

Another strand of research based on the cross-country growth accounting approach tends to focus on the role of techno- logical efficiency in determining economic growth. Klenow and Rodrı́guez-Clare (1997) claimed that total factor productivity (TFP) growth accounts for 90% of the international variation in output growth. Particularly sharp rejections of the importance of capital accumulation also come from Hall and Jones (1999) and Easterly and Levine (2001). However, besides the draw- back of making several assumptions about unknown parameters when estimating production functions, two further argu- ments make us less inclined to use the growth accounting approach. Firstly, when TFP growth is measured as a residual, i.e. as the growth rate in GDP that cannot be accounted for by the growth of the observable inputs, it should not be equated with technological change as many researchers have done. Rather it is ‘a measure of ignorance’ (Abramovitz, 1986), covering many factors like structural change, improvement in allocative efficiency, economies of scale, and other omitted variables and measurement errors. Secondly, although growth accounting provides a convenient way of breaking down observed growth of GDP into components associated with changes in factor inputs and in production technologies, we are not con- vinced that technological change and investment are separable in reality, i.e. changing technology requires investment, and investment inevitably involves technological change. This is consistent with the view of Scott (1989) that technological change and investment are part and parcel of the same thing and that separation has little meaning.

In this paper, we adopt the methodology developed by Temple and Wößmann (2006) to incorporate structural change terms into the augmented Solow model so as to capture the role of both factor accumulation and productivity growth in international variations on output growth. This model has a number of relative advantages. Firstly, compared with the con- ventional MRW models, it allows for cross-country variation in productivity growth by taking into account the effect of la- bour reallocation between sectors with different productivity. Secondly, unlike growth accounting methods, it does not involve the task of measuring the capital stock. Thirdly, the structural model is less subjective to the problem of model uncertainty than Barro-type regressions.

Starting with the framework of MRW, the dynamics of a country’s growth rate towards the steady state can be expressed as

ln YðtÞ LðtÞ � ln

Yð0Þ Lð0Þ

¼�hln Yð0Þ Lð0Þ

þ h a

1 � a lnðsÞ� h

a 1 � a

lnðn þ g þ dÞþ hlnAð0Þþ gt; ð1Þ

where Y is output, L is labour (growing exogenously at rate n), YðtÞLðtÞ and Yð0Þ Lð0Þ are output per worker at time t and at some initial

date, respectively; A is labour-augmenting technological progress (growing exogenously at rate g), and Að0Þ is the initial level of efficiency; s is the constant fraction of output that is saved and invested; d is the depreciation rate of physical capital; a is the elasticity of output with respect to physical capital; h ¼ 1 � e�kt , where k is the rate of convergence, given by k ¼ðn þ g þ dÞð1 � aÞ. Thus, the growth of income per worker is a function of the initial level of income and the determinants of the ultimate steady state.

In order to capture the explicit role of human capital in determining economic growth, MRW augmented the Solow model by including accumulation of human capital as well as physical capital. They provided two possible ways to examine the effect of human capital on economic growth. The first is to estimate the reduced form of the augmented model by including the rate of human capital accumulation. Approximating around steady state, MRW showed that growth of output per worker in this model is given by

ln YðtÞ LðtÞ � ln

Yð0Þ Lð0Þ

¼�hln Yð0Þ Lð0Þ

þ h a

1 � a � b lnðskÞþ h

b 1 � a � b

lnðshÞ� h a þ b

1 � a � b lnðn þ g þ dÞþ hlnAð0Þþ gt; ð2Þ

where sk and sh are the fractions of income invested in physical capital and human capital, respectively. The convergence rate is given by k ¼ðn þ g þ dÞð1 � a � bÞ, where b is the elasticity of output with respect to human capital. The assumption a þ b < 1 implies that there are decreasing returns to capital as a whole.

The second way is to express the role of level of human capital in determining economic growth as

ln YðtÞ LðtÞ � ln

Yð0Þ Lð0Þ

¼�hln Yð0Þ Lð0Þ

þ h a

1 � a lnðskÞþ h

a 1 � a

lnðn þ g þ dÞþ h b

1 � a lnðh�Þþ hlnAð0Þþ gt; ð3Þ

where h� is the steady-state level of human capital. Note that these alternative regressions predict different coefficients on the saving and population growth variables in the augmented Solow model.

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 437

A major criticism of MRW’s specification is their assumption of a common exogenous rate of technological progress (e.g. Easterly and Levine, 2001; Gundlach, 2007; Klenow and Rodrı́guez-Clare, 1997; McQuinn and Whelan, 2007). MRW’s justi- fication is that technology, as a public good, is freely available to individuals and can be transferred instantaneously across national borders. However, this assumption contradicts the fact that diffusion of new technology can be costly or time-con- suming, especially for developing countries. Therefore, it is argued that models for growth in GDP per worker should allow productivity growth to vary across countries.

Temple and Wößmann (2006) developed an empirical model to examine the impact of labour reallocation on aggregate productivity growth and they augmented the conventional growth regressions based on the MRW framework so as to allow for structural change. Their basic idea is that changes in the structure of employment will raise aggregate productivity when the marginal product of labour varies across sectors. If the marginal product of labour is lower in agriculture, then the move- ment of agricultural workers to sectors where the marginal product is higher will raise total output. Since this additional output is produced without change in the total input of capital and labour, the reallocation of labour raises aggregate productivity.

It is a general equilibrium model of production with two sectors (a rural agricultural and an urban non-agricultural sec- tor) and two factors (capital and labour). Total output is given by

Y ¼ Y a þ qY m Xð1; qÞ

; ð4Þ

where q is the relative price of the urban sector good; Y a and Y m are output quantities in agriculture and non-agriculture; and Xð1; qÞ is a GDP price deflator.

The production function in each sector has constant returns to scale and is given by

Y a ¼ AaFðK a; LaÞ; ð5aÞ Y m ¼ Am FðK m; LmÞ; ð5bÞ

where Aa and Am are TFP in agriculture and non-agriculture, respectively. Assuming that workers are paid their marginal products gives

wa ¼ Aa FL; ð6aÞ wm ¼ qAm GL ð6bÞ

where wa and wm are wages in agriculture and non-agriculture, respectively; and the L subscript denotes the partial deriv- ative with respect to labour. Capital also receives its marginal product in both sectors, i.e. AaF K ¼ qAmGK ¼ r, where r is the rental rate on capital and the K subscript is the partial derivative with respect to capital.

This model assumes that any observed effects of reallocation arise because of marginal product differentials and that the propensity to migrate depends on the ratio of wages in the two sectors. Migration will cease when the intersectoral wage ratio falls to a level denoted by k, so the long-run migration equilibrium is

wm ¼ kwa; ð7Þ

where k P 1. The relationship between the extent of structural change and wage ratio can be expressed as

x ¼ p

1 � p ¼ w

wm kwa � 1

� � ; ð8Þ

where p is the migration propensity, defined by p ¼�Daa , where a is the share of agricultural employment in total employ- ment; and w is the speed of adjustment to the long-run equilibrium. The ‘odds ratio’ for migration is increasing in the wage gap between the two sectors. Rearranging (8) gives

wm wa ¼ k 1 þ

1 w

p 1 � p

� � ; ð9Þ

so the extent of current wage ratio can be deduced using information on the observed pace of structural change. In this mod- el, the wage differential varies across countries according to the value of p.

By assuming that the speed of adjustment ðwÞ, the equilibrium differential (k) and the labour share in total output / ¼ wa LY � �

are constant across economies, Temple and Wößmann (2006) derived the following expression for the aggregate Solow residual

_Z Z ¼ sðtÞ

_Aa Aa þð1 � sðtÞÞ

_Am Am ðk � 1Þ/ð1 � aÞ

_m m þ k/

1 w

p 1 � p

ð1 � aÞ _m

m ; ð10Þ

where Z is aggregate efficiency; _X is time derivative of variable X, or _X ¼ dXtdt ; sðtÞ is the nominal output share for agriculture at time t, or sðtÞ¼ Y aY aþqY m ; / is the labour share in total output, or / ¼

wa L Y ; and m is the share of non-agricultural employment in

total employment, or m ¼ 1 � a.

438 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

In the presence of an intersectoral wage differential, the aggregate Solow residual can thus be decomposed as a weighted average of the sectoral TFP growth rates plus the ‘growth bonus’ obtained by reallocating labour to a sector where its mar- ginal product is higher. Since the migration propensity p is related to the extent of structural change as measured by _mm, Eq. (10) implies a convex relationship between growth and structural change. The intuition is that the growth impact of a given extent of structural change will be greatest in those countries experiencing more rapid structural change, as these are also the countries in which the intersectoral wage differential is greatest. Note that the two structural change terms in Eq. (10) will disappear when there is no wage differential in equilibrium, k ¼ 1, and the adjustment process in response to disequi- librium is instantaneous, w ! 1.

Since it was not possible to measure capital stocks at the sectoral level, Temple and Wößmann (2006) treated sectoral TFP as unobservable and relied on a vector V to capture the cross-section variation in aggregate TFP growth that is not due to structural change as follows:

_z z ¼ b0V þðk � 1Þ/MGROWTH þ k/

1 w

DISEQ; ð11Þ

where V is a vector of determinants of aggregate TFP growth including initial level of aggregate TFP and regional differences in technology and institutions proxied by regional dummies; and the structural change terms are defined as

MGROWTH ¼ð1 � aÞ _m

m � Dm ð12aÞ

DISEQ ¼ p

1 � p ð1 � aÞ

_m m �

p 1 � p

Dm: ð12bÞ

Temple and Wößmann (2006) then extended MRW’s model by including the structural change terms derived above to proxy the varying productivity growth across countries. Given the Cobb–Douglas production technology, TFP growth is equal to the growth rate of efficiency (g) times the exponent on the efficiency index ð1 � a � bÞ. In the presence of wage differentials, TFP growth is a function of structural change terms as shown in Eq. (11). Then the extension of MRW’s model takes the form

ln YðtÞ LðtÞ � ln

Yð0Þ Lð0Þ

¼ f þ tðk � 1Þ

1 � a � b MGROWTH þ

tk/ ð1 � a � bÞw

DISEQ þ hc0W � hln Yð0Þ Lð0Þ

; ð13Þ

where f is a constant and W is a vector of explanatory variables including rates of saving, population growth, physical and human capital accumulation. Thus, the specification of Eq. (13) is a hybrid of the Solow model with an aggregate production function and a two-sector framework with sectoral product differentials.

Despite its approximations and limits, this model has a number of comparative advantages. Firstly, compared with the conventional MRW models, Eq. (13) allows for cross-country variation in productivity growth by taking into account the ef- fect of labour reallocation between sectors with different productivity. Secondly, unlike the use of accounting methods to measure TFP growth, this model does not involve the task of measuring the capital stock, which might be problematic for developing countries.

When replacing the assumption that the labour share in output, /, is the same across countries by an assumption that all countries have the same Cobb–Douglas technologies in agriculture, Temple and Wößmann (2006) constructed a second set of structural change terms

MGROWTH2 ¼ð1 � aÞ s a

_m m

ð14aÞ

DISEQ 2 ¼ p

1 � p ð1 � aÞ

s a

_m m ; ð14bÞ

where s is the share of agriculture in total value added. This alternative set of structural change terms adds sa, i.e. the share of agriculture in value added divided by the share of employment.

4. Data and econometric methodology

4.1. Data and sample

Our empirical analysis is based on several datasets of worldwide aggregate series, including Penn World Table (PWT), World Bank World Development Indicators (WDI) and Statistical Database of the Food and Agriculture Organization of the United Nations (FAO). Given the presence of cyclical effects as shown in Fig. 1, we opt for non-overlapping 5-year time intervals which are less sensitive to temporary factors associated with business cycles. The sample period we choose is 1980–2004, corresponding roughly to the reform period in China.

We employ real GDP chain per worker (RGDPWOK) from PWT 6.2, which is adjusted for purchasing power parities (PPPs) and constant price. PPPs are believed to be superior to exchange rates when comparing real incomes and growth rates across countries, as the use of commercial exchange rates tends to overstate the magnitude of income disparities (Temple, 1999a;

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 439

Bosworth and Collins, 2003). However, there is concern about the PPP estimates for China3 and no consensus on the reliability of China’s GDP figures from various sources.4 To address the potential measurement error problems we run a large number of sensitivity tests of the robustness of our findings.

The dependent variable is the change in the logarithm of real GDP per worker at 5-year intervals, and the initial level of income on the right-hand side is measured by real GDP per worker data, starting in 1980, 1985, etc. and ending in 2000. We prefer the per worker variable to the per capita variable because the Solow model is based on a Cobb–Douglas production function in which only the economically active population is involved.

Following MRW, Islam (1995), Caselli et al. (1996), and Hoeffler (2002), we proxy the share of saving by the share of investment in real GDP, which can be obtained from PWT 6.2. The time series are averaged over each 5-year interval. Heston and Sicular (2008) argued that conversion to a common price base may lower the capital formation proportions for China. We therefore, conduct a sensitivity test using China’s official data of the investment share in real GDP to check whether our results are robust.

WDI (September 2006 edition) data on total population and fraction of the population in the age group 15–64 allow us to calculate the working-age population for each country. The average rate of growth of the workforce is computed as the dif- ference between the natural logarithms of the working-age population at the end and beginning of each period and dividing this difference by the number of years.

Rather than follow MRW, Caselli et al. (1996) and Bond et al. (2001) in using the secondary-school enrolment rates to proxy the rates of investment in schooling, we rely on the average level of human capital data provided by Barro and Lee (2001). Both Gemmell (1996) and Temple (1999a) argued that school enrolment rates may conflate human capital stock and accumulation effects and can be a poor proxy for either. The human capital measure we use is average years of school- ing in the population aged over age 15, which provides a direct measure of the stock of human capital at 5-year intervals. For our sensitivity tests we adopt Wang and Yao (2003) estimates of average years of schooling. They argued that the num- ber of graduates from the educational system is a more accurate flow measure than the enrolment rates used by Barro and Lee (2001), and accordingly used each year’s school graduates to obtain the addition to the human capital stock.

FAO provides annual data on total labour force and agricultural labour force, respectively, making the calculation of agri- cultural share of employment possible for most countries. After comparing the employment data for China from FAO with those from China Labour Statistical Yearbook compiled by the National Bureau of Statistics of China (NBS), we find a large dis- crepancy between these two sources (see Table A1 in Appendix). The FAO data correspond closely in most years to the NBS data for total rural employment (including rural non-agricultural employment, amounting to 170 million in 2000). The origin of the FAO’s error is probably the official classification of all rural people (including rural non-farm employment, e.g. workers in the township and village enterprises) as the ‘agricultural population’ under central planning, implying that the error is China-specific. We therefore, use the NBS data for agricultural labour force in our analysis. Brandt et al. (2008) argued that the NBS employment data series contains a major discontinuity in 1990, and that it too may underestimate the rate of de- cline in the primary sector labour force.5 Given the important role of structural change variables played in this model, we adopt Brandt et al.’s data as a further sensitivity test. The annual data on agricultural share of value added are available from WDI. The quinquennial beginning-period data on both employment share and value added share for each country are used to construct the structural change terms.

We consider three samples of countries in this paper.6 Sample I comprises all countries available from PWT 6.2 except those receiving a grade ‘D’ in terms of data quality. As pointed out by MRW, the problem of measurement error is likely to be extre- mely serious for these countries and variables can be badly measured. By eliminating these least reliable data from our sample, we are left with a sample of 146 developed and developing countries.

Sample II contains all developing countries and four East Asian Tigers7 in our non-grade-D sample. Temple (1999a) men- tioned that integrating developed and developing countries in a single empirical framework is not without its problems since institutions and growth processes in developing countries can be different from those in countries already near the technolog- ical frontier. We incorporate four East Asian Tigers into this sample because China shares some economic growth patterns with these countries owing to cultural similarities, geographic location and similar economic development strategies. This sample contains 111 countries after excluding OECD and non-OECD high-income economies of Sample I.

Sample III comprises 61 large developing countries with more than 5 million population and four East Asian Tigers, where grade D data are also excluded. These countries are believed to have much in common with China. By grouping countries

3 In the ‘Treatment of China in PWT 6’, Heston pointed out that the basis for PPP estimates for China is very little improved over previous versions of PWT. The World Bank and IMF recently revised downward their estimates for China’s PPP-based GDP by around 40%, based on new statistical calculations of PPP exchange rates published in December 2007 by the International Comparison Program (ICP). Since our focus is growth rather than level of income, we expect the impact of any inaccuracy of the PPP figures to be relatively minor in our research.

4 Maddison (1998) questioned Chinese official growth statistics and constructed an alternative time series, in which the average real GDP growth rate for China over the period 1978–1995 is 2.39% points below the official one. Rawski (2001) also claimed that official Chinese statistics contain major exaggerations of real output growth beginning with 1998. The PWT roughly follows Maddison’s adjustments to the Chinese national accounts but creates an alternative constant price series based on expenditure deflators. However, Holz (2006) argued that the use of expenditure deflators instead of the official implicit deflators reduces China’s real GDP growth rate, i.e. the PWT figure for China may be biased downwards on that account.

5 In Table A1 in Appendix, the Brandt et al. (2008) data show a more dramatic decline in the share of agriculture in employment than do the NBS data. 6 Full list of countries included in different samples are provided in Table A2 in the Appendix. 7 Hong Kong, Taiwan, South Korea, and Singapore.

440 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

with similar features into the same sample, we expect to control for the difference in technology and institutions and alle- viate the problem of parameter heterogeneity.

4.2. Empirical methodology

Since judging whether countries are in their steady states is problematic in practice, our empirical analysis will focus on the growth rather than the income equation in order explicitly to consider the transitional dynamics through the inclusion of initial income. The growth regression approach encounters the omitted variable problem associated with the unobservable initial level of technology. In a single cross-section growth regression, this omitted term Að0Þ is left within the residual term. Since variations in technical efficiency across countries are likely to be correlated with other explanatory variables, estimates of regressors in a conditional convergence regression are biased and inconsistent. Panel data methods make it possible to control for the unobserved country-specific effect by treating initial efficiency as a time-invariant fixed effect and eliminat- ing its influence through a time-dimensional transformation. Another advantage of the panel over the cross-section regres- sion is the alleviation of the endogeneity problem through the inclusion of lags of regressors as instruments. We therefore rely on panel data methods to estimate the cross-country growth regressions.

Following Bond et al. (2001), our Eqs. (1)–(3) and (13), can be generalised in the following panel data model:

Dyi;t ¼ða � 1Þyi;t�1 þ x 0 i;t b þ gi þ ct þ mi;t; ð15Þ

for i ¼ 1; . . . ; N and t ¼ 2; . . . ; T, where Dyi;t is the log difference in real GDP per worker over a 5-year period, yi;t�1 is the log- arithm of real GDP per worker at the beginning of each period, and xi;t is a vector of other characteristics measured either at the beginning of each period or as an average over each 5-year period including physical and human capital accumulation, population growth and structural change variables. In this paper, we maintain the MRW idea of a common world technology trend representing advancement of knowledge, but we allow for the variation in productivity growth associated with struc- tural change. In addition, the unobserved heterogeneity in the initial level of efficiency is reflected by the country-specific effects, gi. The time dummy, ct , is expected to capture productivity changes that are common to all countries. Both the coun- try- and time-effects may also reflect country-specific and period-specific components of measurement errors (Bond et al., 2001).

Estimating Eq. (15) is equivalent to estimating a dynamic panel data model with a lagged dependent variable on the right- hand side as

yi;t ¼ ayi;t�1 þ x 0 i;t b þ gi þ ct þ mi;t; ð16Þ

for i ¼ 1; . . . ; N and t ¼ 2; . . . ; T . In the context of cross-country growth regressions, our data are featured by a large number of countries N over a small averaged time-series period T.

The presence of country-specific effects, gi, implies several econometric problems relating to the estimation of dynamic panel data models. The correlation between the lagged dependent variable and the time-invariant country-specific effects renders the OLS estimator biased and inconsistent. In the cross-country growth regressions, the OLS estimate of the coeffi- cient of initial income term, â, is likely to be biased upward owing to the positive correlation between Y i;t�1 and gi (Hsiao, 1986). For the fixed effects estimator, the within-groups transformation wipes out the time-invariant gi, but ðY i;t�1 � Y i:�1Þ, where Y i:�1 ¼

PT t¼2 Y i;t�1=ðT � 1Þ, can still be correlated with mi;t � �mi: even if the mi;t are not serially correlated. Nickell (1981)

showed that the unbiasedness and consistency of within-groups estimator in a dynamic panel will depend upon T being large. However, in the typical growth regression with small T, the estimate of the coefficient of initial income term, â, is likely to be seriously biased downwards (Nickell, 1981).

The growth regression using the first-differenced GMM estimator is first differenced to eliminate the effect of initial efficiency and then lagged levels of the right-hand-side variables are used as instruments in the first-differenced equa- tions. However, Bond et al. (2001) indicated that the first-differenced GMM estimator is subject to a large downward finite sample bias particularly when the number of time series observations is small, as the lagged levels of variables are only weak instruments for subsequent first-differences. Instead, they recommended using a system GMM estimator with supe- rior finite sample property developed by Arellano and Bover (1995) and Blundell and Bond (1998). By adding the original equation in levels to the system, Arellano and Bover (1995) and Blundell and Bond (1998) found dramatic improvement in efficiency and significant reduction in finite sample bias through exploiting these additional moment conditions. Bond et al. (2001) also claimed that the potential for obtaining consistent parameter estimates even in the presence of measurement error and endogenous right-hand-side variables is a considerable strength of the GMM approach in the context of empirical growth research. As a consequence, a panel-data system GMM estimator will be our preferred estimation method.

Detection of outliers is important in the cross-country growth regression when a large number of heterogeneous coun- tries are included in the sample (Temple, 1999b). In the dynamic panel data framework the use of a lagged dependent var- iable guarantees that an outlier in the dependent variable will also show up as a bad leverage point in the independent variables. Temple (1999a) suggested that single-case diagnostics like Cook’s distance measure, the Studentized residuals and DFITS are likely to miss groups of outliers or wrongly identify representative observations as outlying. Therefore, we rely on the robust regression technique, iteratively reweighted least squares (RWLS), to identify possible outliers and then omit

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 441

these from our estimation. RWLS assigns a different weight to each observation with zero or lower weights given to observations with large residuals. After removing 13 unrepresentative observations8 from our sample (with weights less than 0.5), we try to restrict the influence of outliers and focus on the most coherent part of the dataset.9

5. Empirical results

5.1. The textbook Solow model

We start by estimating the textbook Solow model as described by Eq. (1) in Table 2. Note that all estimated standard errors are corrected for heteroskedasticity and time dummies are included in each regression. In the system GMM estima- tion initial level of income is treated as a predetermined variable and both investment rates and population growth rates are treated as potentially endogenous variables. Since the p values of over-identifying tests may be inflated when the number of moment conditions is large (Bowsher, 2002), we restrict the number of instruments used for each first-differ- enced equation by including a subset of instruments for each predetermined or endogenous variable. After applying a cor- rection to the two-step covariance matrix derived by Windmeijer (2005), we find very similar results obtained from the one-step and two-step GMM estimators. Therefore, we report only the heteroskedasticity-robust one-step system GMM results in this paper.

The coefficients on initial income have the expected negative sign and are highly significant for all three samples using various estimation methods, indicating strong evidence of conditional convergence. Hence a lower starting value of real in- come per worker tends to generate a higher growth in GDP per worker, once the determinants of steady states are controlled for. In line with the prediction of Bond et al. (2001) and Hoeffler (2002), we find that our system GMM estimator yields a consistent estimate which lies in between the upper bound provided by the OLS estimator and lower bound given by the within-groups estimator.

The investment rate has a significantly positive effect on the growth of GDP per worker in all regressions even after con- trolling for unobserved country-specific effects and allowing for the likely endogeneity of investment.10 However, we fail to identify a significantly negative correlation between population growth and the growth of income per worker. The restriction that the coefficients of the investment and population growth variables are equal in magnitude but opposite in sign is rejected in Samples I and II. Moreover, the estimated elasticity of output with respect to capital ðaÞ obtained from a restricted version of Eq. (1)11 is found to be above 0.5 for all three samples, which is higher than the model-suggested-value of capital share of in- come, 0.33. Therefore, for reasons similar to those of MRW, we reject the textbook Solow model based on our robust system GMM panel data estimation.

5.2. The Solow model augmented with human capital

The role of education in determining economic growth is an area of dispute in the cross-country growth empirics. MRW found a significantly positive effect of human capital on growth, while other studies (Benhabib and Spiegel, 1994; Pritchett, 1999) claimed that increases in measured educational attainment are not related to output growth especially in developing countries.

In Table 3, we estimate Eq. (2) which incorporates the log difference of average years of schooling as a proxy for the accu- mulation of human capital as well as Eq. (3) which includes the logarithm of average years of schooling as a measure of the level of human capital, respectively. In addition, following Gemmell (1996) who argued that higher output growth may re- sult from both larger initial stocks of human capital and a faster rate of human capital accumulation, we test a third spec- ification which augments the Solow model with both the stock and accumulation of human capital.

From now on, we report only the results of system GMM estimation. We treat the initial level of human capital as a pre- determined variable and the growth rate of human capital as a potentially endogenous variable, as fast-growing economies are likely to devote a higher proportion of their resources to educational investment. We find that when the two human cap- ital variables enter the Solow formulation individually, neither of them proves to be significant, which is consistent with many studies that have failed to find a robust correlation between educational attainment and output growth. Interestingly, when we simultaneously incorporate both the stock and accumulation of human capital into the regressions, the level of human capital becomes highly significant and positive for all samples; besides, Wald tests suggest strong evidence of joint

8 The removed outliers are Kuwait, 1995; Congo, Republic of, 2000; Cameroon, 1985; Zambia, 1995; Jordan, 1990; Rwanda, 1980; Sierra Leone, 2000; Iran, 1980; Swaziland, 1990; Jamaica, 1980; Ireland, 2000; Uruguay, 1985; and Paraguay, 1980.

9 In the working paper version of this paper (Ding and Knight, 2008), we show that omitting the outliers from our samples results in better goodness of fit of the regressions: a rise in the R-squares of OLS and within-groups estimation and a decline in the estimated standard errors of most regressors. Besides, we find that our prediction standard errors are bigger in all cases if the outliers are not removed, resulting in wider confidence intervals for all periods. To save space here, we report only the estimation results after the outliers are removed.

10 In our system GMM estimation the original investment variables lagged by 10-year and 15-year periods are used as instruments in the first-differences equations, and first-differenced investment variables lagged by 5-year period are used as additional instruments in the levels equations.

11 The detailed estimation results for the restricted model are reported in Ding and Knight (2008), Table 4.

Table 2 The textbook Solow model

Sample I: 146 countries Sample II: 111 countries Sample III: 61 countries

OLS Within groups System GMM OLS Within groups System GMM OLS Within groups System GMM

Constant 0.181** (0.081) 2.962** (0.438) 0.826** (0.380) 0.292** (0.114) 2.988** (0.477) 0.813* (0.462) 0.256 (0.215) 1.883** (0.595) 0.727 (0.475) lnðY i;t�1Þ �0.051

** (0.008) �0.307** (0.042) �0.145** (0.029) �0.046** (0.009) �0.318** (0.047) �0.110** (0.051) �0.074** (0.011) �0.284** (0.057) �0.176** (0.057) lnðsitÞ 0.124

** (0.015) 0.114** (0.034) 0.256** (0.048) 0.118** (0.016) 0.106** (0.039) 0.237** (0.048) 0.154** (0.020) 0.171** (0.050) 0.198** (0.047) lnðnit þ g þ dÞ �0.029 (0.036) 0.117

** (0.047) 0.025 (0.116) 0.025 (0.038) 0.131** (0.045) 0.129 (0.144) �0.045 (0.085) �0.096 (0.109) �0.159 (0.141) R2 0.239 0.359 0.218 0.358 0.339 0.381 m1 �3.73 [0.000] �3.28 [0.001] �2.64 [0.008] m2 �0.53 [0.598] �0.73 [0.468] �1.20 [0.231] Hansen test p value 0.213 0.290 0.364 Difference Sargan p value 0.637 0.675 0.261 Implied k 0.009 0.073 0.031 0.009 0.077 0.023 0.015 0.067 0.039 Adding-up restriction p value 0.018 0.000 0.037 0.001 0.001 0.011 0.245 0.543 0.798 No. of observations 511 511 511 368 368 368 230 230 230

Note: Heteroskedasticity-consistent standard errors are in parentheses; the test statistics for first and second order correlation are given by m1 and m2 , respectively, and the p-values are in brackets; In the system GMM estimation, lnðY i;t�1Þ is treated as a predetermined variable; lnðsitÞ and lnðnit þ g þ dÞ are treated as endogenous variables; g þ d is assumed to be equal to 0.05; k is the convergence rate; the adding-up restriction refers to the hypothesis that the coefficients of investment and population growth rate are equal in magnitude but opposite in sign as predicted by Eq. (1); and �� and � indicate that the coefficient is significantly different from zero at the 5% and 10% significance level, respectively.

Table 3 System GMM estimation of the augmented Solow model with human capital

Sample I: 146 countries Sample II: 111 countries Sample III: 61 countries

(1) (2) (3) (1) (2) (3) (1) (2) (3)

Constant �0.442 (0.330) �0.699**

(0.340) �0.351 (0.301) �0.213 (0.324) �0.402 (0.298) �0.163 (0.342) �0.177 (0.328) �0.359 (0.243) �0.144 (0.315)

lnðY i;t�1Þ �0.092 **

(0.029) �0.087**(0.022) �0.101**

(0.025) �0.085**

(0.033) �0.061**

(0.029) �0.075**

(0.029) �0.075**

(0.038) �0.059**

(0.027) �0.093**

(0.028) lnðsitÞ 0.173

**(0.042) 0.139**(0.039) 0.121** (0.032) 0.157** (0.038) 0.146** (0.033) 0.132** (0.029) 0.137** (0.038) 0.122** (0.037) 0.116** (0.032) lnðnit þ g þ dÞ �0.348

**

(0.131) �0.461**

(0.156) �0.346**

(0.127) �0.239**

(0.119) �0.253**

(0.107) �0.181 (0.133) �0.218**

(0.117) �0.269**

(0.089) �0.263**

(0.087) lnðhitÞ 0.014 (0.044) 0.084

** (0.041) 0.053 (0.045) 0.086** (0.041) 0.028 (0.051) 0.073**(0.036) DlnðhitÞ 0.241 (0.159) 0.258

* (0.141) 0.219 (0.146) 0.189 (0.127) 0.011 (0.159) 0.077 (0.102) Joint significance test for lnðhitÞ and

DlnðhitÞ 5.51 [0.064] 5.49 [0.064] 5.66 [0.059]

m1 �3.70 [0.000] �4.05 [0.000] �4.18 [0.000] �3.63 [0.000] �3.72 [0.000] �4.00 [0.000] �2.68 [0.007] �2.64 [0.008] �2.80 [0.005] m2 �1.03 [0.301] �1.12 [0.262] �1.11 [0.265] �1.11 [0.269] �1.02 [0.308] �1.03 [0.303] �0.82 [0.414] �0.90 [0.371] �0.94 [0.348] Hansen test p value 0.365 0.610 0.963 0.998 0.992 0.998 0.999 0.999 0.999 Difference Sargan p value 0.430 0.492 0.476 0.870 0.833 0.701 0.707 0.933 0.999 Implied k 0.019 0.018 0.021 0.018 0.013 0.016 0.016 0.012 0.020 Adding-up restriction p value 0.198 0.734 0.517 0.529 0.497 0.458 No. of observations 378 375 375 266 263 263 184 184 184

Note: Heteroskedasticity-consistent standard errors are in parentheses; the test statistics for first and second order correlation are given by m1 and m2, respectively, and the p-values are in brackets; lnðY i;t�1Þ and lnðhitÞ are treated as predetermined variables; lnðsitÞ; lnðnit þ g þ dÞ and DlnðhitÞ are treated as endogenous variables; g þ d is assumed to be equal to 0.05; k is the convergence rate; when the growth rate of human capital is included, the adding-up restriction refers to the hypothesis that the three coefficients other than the one on lagged output sum to zero; when the level of human capital is included, the restriction is that the coefficients on the rates of investment and population growth are opposite in sign and equal in absolute value; and �� and � indicate that the coefficient is significantly different from zero at the 5% and 10% significance level, respectively.

4 4

2 S.

D in

g, J.

K n

igh t/Jo

u rn

a l

o f

C o

m p

a ra

tive E

co n

o m

ics 3

7 (2

0 0

9 )

4 3

2 –

4 5

2

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 443

significance of both human capital variables even in Samples II and III where only developing countries are included. Hence, our results provide support for a role of both the initial stock and subsequent growth of human capital in fostering faster output growth even in less developed countries, which is in contrast to the so-called ’Pritchett hypothesis’ (Pritchett, 1999; Temple, 2001).

Compared with the textbook Solow model, inclusion of human capital in the regressions leads to several major changes. First, the population growth term which has been wrongly signed previously becomes negative and strongly significant for all samples except for one regression in Sample II. Second, these unrestricted regressions do not lead to rejection of the add- ing-up hypotheses as predicted by Eqs. (2) and (3). Moreover, in the restricted models,12 the calculated physical capital’s share of income ðaÞ and human capital’s share of income ðbÞ suggest a þ b < 1, justifying decreasing returns to the set of repro- ducible factors of production, a key assumption of the neoclassical Solow model. In brief, all results suggest better performance of the augmented Solow model with human capital than the textbook one.

5.3. The augmented Solow model with structural change

We now further supplement the augmented Solow model with the structural change terms to test whether labour real- location makes a significant contribution to economic growth. Table 4 presents the system GMM results with the first set of structural change terms, MGROWTH and DISEQ. Being aware of the endogenous nature of the extent of structural change, i.e. periods of more rapid economic growth are also periods of expanding opportunity for rural workers and of rapid structural transformation, we treat both the linear and non-linear structural change terms as potentially endogenous variables.13 In line with the findings of Temple and Wößmann (2006), although MGROWTH and DISEQ are not individually significant, there is strong evidence of joint significance for all samples according to the Wald test.

The results become even better when we add the second set of structural change terms, MGROWTH2 and DISEQ2, into the cross-country growth regressions. Recall that the alternative set of structural change variables captures both structural change in employment and sectoral transformation in total value added. In Table 5, not only are both structural change terms jointly significant but also the non-linear term, DISEQ2, itself remains highly significant and positive in every regression and each structural change term appears strongly individually significant in Samples II and III where only developing countries are involved. Reflecting very different sectoral structures and patterns of structural change in developed and developing countries, our results show that the role of structural change in determining economic growth is stronger in the case of developing countries. In addition, the persistently significant DISEQ2 term further justifies Temple and Wößmann (2006)’s hypothesis that the growth effect of structural change is nonlinear.

Inclusion of the second set of structural change terms further improve the performance of the human capital variables. There is now a significant and positive association between cross-country differences in the initial endowment level of edu- cation and subsequent output growth even when the stock of human capital enters the regression individually. Besides, all other parameters are correctly signed and highly significant in every regression. There is also no evidence of second order serial correlation in the first-differenced residuals and neither Hansen test nor Difference Sargan test rejects the validity of instruments, suggesting the consistency of the system GMM estimators being used. The estimated convergence rate, k, remains stable at around 2% per annum for each sample, which is consistent with the evidence commonly found in the cross-country growth literature (for instance, Bond et al., 2001; Barro and Sala-i-Martin, 2004).

In brief, our system GMM results strongly support the extended version of the augmented Solow model with both human capital and structural change, as developed by Temple and Wößmann (2006). The movement of labour across sectors is an essence of the development process, and this needs to be captured in the cross-country growth regressions.

5.4. Growth and growth difference predictions

The good performance of the augmented Solow model with structural change allows us to predict China’s growth rate and to examine the reasons why China has grown faster than other countries. The prediction is made by introducing China’s val- ues of the explanatory variables into the estimated equation. The model we employ for prediction is the one augmented by both level and growth rate of human capital as well as structural change terms given in Table 4. We choose the estimates of Sample III when we compare China with Sub-Saharan Africa and opt for Sample I when we later account for the predicted growth difference between China and all other country groups in the world. Our results remain robust when the sensitivity tests described in Section 4.1 are conducted.14 To save space, we report only the prediction results based on the sample data originally used.

12 The detailed estimation results for the restricted models with human capital are reported in Ding and Knight (2008), Table 6. 13 The levels of structural change variables lagged by 10-year and 15-year periods are used as instruments in the first-differences equations, and first-

differenced structural change variables lagged by 5-year period are used as additional instruments in the levels equations. 14 We ran a large number of sensitivity tests for China using Brandt et al. (2008) employment data, China’s official real GDP data, China’s official data of

investment over GDP, Wang and Yao (2003) human capital data, and a sub-sample test for the period 1980–1994 when the reliability of China’s GDP data is less subject to dispute. Our results proved to be consistent and robust. Details of these results are available from the authors upon request.

Table 4 System GMM Estimation of the augmented Solow model with human capital and the first set of structural change terms

Sample I: 146 countries Sample II: 111 countries Sample III: 61 countries

(1) (2) (3) (1) (2) (3) (1) (2) (3)

Constant �0.408* (0.246) �0.661** (0.259) �0.406 (0.294) �0.016 (0.389) �0.184 (0.345) �0.147 (0.306) �0.212 (0.271) �0.166 (0.272) �0.057 (0.276) lnðY i;t�1Þ �0.113

** (0.018) �0.095** (0.017) �0.085** (0.023) �0.118** (0.024) �0.095** (0.018) �0.085**(0.024) �0.101** (0.021) �0.101** (0.018) �0.096** (0.023) lnðsitÞ 0.206

** (0.035) 0.206** (0.032) 0.113** (0.026) 0.176** (0.030) 0.192** (0.028) 0.116** (0.026) 0.148** (0.031) 0.153** (0.028) 0.109** (0.027) lnðnit þ g þ dÞ �0.345

** (0.090) �0.404** (0.096) �0.312** (0.114) �0.237** (0.114) �0.242** (0.103) �0.216** (0.095) �0.296** (0.080) �0.285** (0.078) �0.245** (0.079) lnðhitÞ 0.045 (0.032) 0.067

** (0.033) 0.071* (0.038) 0.071** (0.035) 0.022 (0.041) 0.053* (0.033) DlnðhitÞ 0.057 (0.076) 0.166 (0.108) 0.026 (0.078) 0.092 (0.083) 0.057 (0.102) 0.062 (0.088) MGROWTH 0.561 (1.130) 0.475 (1.194) 1.225 (1.109) 0.413 (1.145) 0.369 (1.173) 1.225 (0.928) 1.135 (0.902) 0.701 (1.046) 0.397 (0.879) DISEQ 2.484 (3.184) 2.916 (3.177) 2.057 (3.055) 2.971 (2.974) 3.457 (3.099) 2.074 (2.568) 2.484 (2.624) 3.841 (3.062) 3.732 (2.473) Joint significance test for

lnðhitÞ and DlnðhitÞ 4.98 [0.083] 4.57 [0.101] 2.74 [0.255]

Joint significance test for MGROWTH and DISEQ

7.18 [0.028] 8.54 [0.014] 20.67 [0.000] 9.56 [0.008] 9.72 [0.008] 23.11 [0.000] 17.00 [0.000] 15.91 [0.000] 18.51 [0.000]

m1 �3.69 [0.000] �3.81 [0.000] �4.29 [0.000] �3.45 [0.000] �3.56 [0.000] �3.76 [0.000] �2.90 [0.004] �2.91 [0.004] �2.89 [0.004] m2 �0.83 [0.408] �1.07 [0.284] �1.17 [0.240] �0.96 [0.337] �1.11 [0.266] �1.12 [0.264] �0.75 [0.456] �0.85 [0.393] �0.90 [0.366] Hansen test p value 0.595 0.532 0.999 0.982 0.990 0.999 0.999 0.999 0.999 Difference Sargan p value 0.200 0.336 0.707 0.359 0.326 0.999 0.985 0.991 0.999 Implied k 0.024 0.020 0.018 0.025 0.020 0.018 0.021 0.021 0.020 No. of observations 373 370 370 261 258 258 179 179 179

Note: MGROWTH and DISEQ are treated as endogenous variables; other definitions are the same as those in the previous tables.

Table 5 System GMM estimation of the augmented Solow model with human capital and the second set of structural change terms

Sample I: 146 countries Sample II: 111 countries Sample III: 61 countries

(1) (2) (3) (1) (2) (3) (1) (2) (3)

Constant �0.235 (0.274) �0.659** (0.276) �0.156 (0.251) 0.036 (0.380) �0.386 (0.310) 0.166 (0.341) 0.169 (0.306) �0.155 (0.286) 0.395 (0.262) lnðY i;t�1Þ �0.106

** (0.029) �0.107** (0.021) �0.106** (0.029) �0.093** (0.029) �0.071** (0.016) �0.081** (0.028) �0.107** (0.027) �0.094** (0.022) �0.118** (0.021) lnðsitÞ 0.153

** (0.032) 0.135** (0.037) 0.120** (0.028) 0.139** (0.029) 0.138** (0.029) 0.114** (0.028) 0.096** (0.031) 0.096** (0.035) 0.105** (0.027) lnðnit þ g þ dÞ �0.297

** (0.159) �0.513** (0.119) �0.276** (0.133) �0.170 (�0.149) �0.241* (0.125) �0.087 (0.134) �0.197** (0.101) �0.314** (0.125) �0.133* (0.082) lnðhitÞ 0.066

** (0.033) 0.089** (0.034) 0.066* (0.036) 0.078** (0.037) 0.068* (0.037) 0.064** (0.026) DlnðhitÞ 0.110 (0.111) 0.159 (0.124) 0.088 (0.101) 0.120 (0.118) �0.035 (0.132) 0.045(0.075) MGROWTH2 1.333 (1.167) 1.029 (0.936) 1.262 (1.086) 1.533 (1.228) 1.699* (0.985) 1.467 (1.129) 2.732** (1.071) 3.092** (1.152) 2.941** (0.696) DISEQ2 0.143** (0.062) 0.223** (0.061) 0.188** (0.069) 0.118** (0.059) 0.111** (0.057) 0.131** (0.056) 0.196** (0.058) 0.214** (0.061) 0.226** (0.054) Joint significance test for

lnðhitÞ and DlnðhitÞ 6.72 [0.035] 4.74 [0.093] 6.61 [0.037]

Joint significance test for MGROWTH2 and DISEQ2

9.09 [0.011] 17.85 [0.000] 12.80 [0.002] 8.34 [0.015] 10.40 [0.006] 9.53 [0.009] 13.83 [0.001] 14.73 [0.001] 36.77 [0.000]

m1 �3.74 [0.000] �3.82 [0.000] �3.99 [0.000] �3.54 [0.000] �3.41 [0.001] �3.69 [0.000] �2.66 [0.008] �2.48 [0.013] �2.56 [0.010] m2 �0.92 [0.360] �1.06 [0.288] �1.01 [0.314] �0.96 [0.337] �1.06 [0.287] �0.96 [0.338] �0.77 [0.442] �0.82 [0.414] �0.78 [0.433] Hansen test p value 0.538 0.986 0.873 0.968 0.999 0.999 0.999 0.999 0.999 Difference Sargan p value 0.246 0.393 0.395 0.437 0.969 0.899 0.979 0.997 0.999 Implied k 0.022 0.023 0.022 0.020 0.015 0.017 0.023 0.020 0.025 No. of observations 355 352 352 250 247 247 170 170 170

Note: MDROWTH2 and DISEQ2 are treated as endogenous variables; other definitions are the same as those in the previous tables.

4 4

4 S.

D in

g, J.

K n

igh t/Jo

u rn

a l

o f

C o

m p

a ra

tive E

co n

o m

ics 3

7 (2

0 0

9 )

4 3

2 –

4 5

2

Table 6 Growth prediction for China and growth difference prediction bet een China and Sub-Saharan Africa (based on Sample III estimation)

Variable (1) Parameter estimates

(2) Mean val of China (per an m)

(3) Mean value of SSA (per annum)

(4) Mean difference (China vs. SSA) (4) = (2) � (3)

(5) Difference in predicted growth (China vs. SSA) (5) = (1) � (4)

(6) Percentage of total predicted growth difference (China vs. SSA)

lnðY i;t�1Þ �0.096 1.554 1.655 �0.101 0.010 17.2 lnðsitÞ 0.109 0.662 0.386 0.276 0.030 53.7 lnðnit þ g þ dÞ �0.245 �0.760 �0.704 �0.056 0.014 24.3 lnðhitÞ 0.052 0.345 0.232 0.113 0.006 10.6 DlnðhitÞ 0.062 0.015 0.016 �0.001 0.000 �0.2 MGROWTH 0.397 0.011 0.004 0.007 0.003 4.9 DISEQ 3.733 0.001 0.000 0.001 0.004 7.3 Actual annual growth rate or

growth rate difference 0.072 0.004 0.068

Predicted annual growth rate or growth rate difference

0.063 0.007 0.056

Residual 0.009 0.012

Note: The predicted growth for China equals the sum of all contr utions to growth by the regression variables including constant and time dummies. The difference in the sample means of a variable between regions X and Y equals the average value in region X minus the av ge value in region Y; the difference in predicted growth between region X and Y attributable to a certain variable is equal to the difference in the sample means of that variable between region X and Y times the timated coefficient on that variable from the regression; the total difference in predicted growth between region X and Y equals the sum of the differences in predicted growth attributable to all variables cont ned in the regression including constant and time dummies.

Table 7 Growth prediction for China and growth difference prediction bet een China and other country-groups (based on Sample I estimation)

Variable Parameter estimates

Mean value of China

P centage components of the difference in predicted growth rates

C na vs. all o er countries

China vs. high- income economies

China vs. all other developing countries

China vs. Sub- Saharan Africa

China vs. Latin America and the Caribbean

China vs. East Asia and the Pacific

China vs. South Asia

lnðY i;t�1Þ �0.085 1.554 6 8 93.4 46.9 24.8 52.6 46.8 21.2 lnðsitÞ 0.113 0.662 2 8 10.9 31.3 42.9 27.5 40.7 40.8 lnðnit þ g þ dÞ �0.313 �0.760 � 3.1 �47.1 �3.6 24.2 7.2 1.0 3.4 lnðhitÞ 0.067 0.345 � .8 �11.0 2.8 10.8 �0.2 �2.9 16.2 DlnðhitÞ 0.165 0.015 � .7 1.0 �1.4 �1.7 0.0 �0.4 �6.1 MGROWTH 1.225 0.011 1 3 19.8 11.0 10.9 10.3 13.7 13.8 DISEQ 2.057 0.001 2 2.6 2.1 2.9 1.8 3.7 3.5 Actual annual growth rate or

growth rate difference 0.072 0 61 0.058 0.059 0.066 0.067 0.051 0.044

Predicted annual growth rate or growth rate difference

0.067 0 56 0.052 0.056 0.064 0.061 0.044 0.049

Residual 0.005 0 05 0.006 0.003 0.002 0.006 0.007 �0.005

Note: All other countries consist of 145 countries in Sample I exce China, which also includes Europe and Central Asia that is not reported in this table; the high-income economies include 39 high-income OECD and non-OECD members; All other developing countries consist o 06 countries except China; Sub-Saharan Africa includes 26 countries; Latin America and the Caribbean contains 26 countries; East Asia and the Pacific comprises 14 countries excluding China; and South Asia cludes 7 countries.

S. D

in g,

J. K

n igh

t /Jo

u rn

a l

o f

C o

m p

a ra

tive E

co n

o m

ics 3

7 (2

0 0

9 )

4 3

2 –

4 5

2 4

4 5

w

ue nu

ib era es ai

w

er

hi th

4. 5. 1 1 0 3. .3 .0

.0

.0

pt f 1

in

446 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

Table 6, using PWT 6.2, shows that the actual average annual growth rate of output per worker in China over the per- iod 1980–2004 was 7.2%. Our model predicts that output per worker in China grew at an average rate of 6.3% per annum, implying an unexplained residual of 0.9% per annum. Given the average prediction standard error of 0.7%, China’s annual growth rate falls within a 90% confidence interval for the prediction.15 The part of the unexplained residual that is not due to measurement error might represent the remaining improvement in allocative and technical efficiency that is not ac- counted for by the model – resulting, for instance, from economic reforms and marketization. Thus China’s growth is due to a combination of both pushing out the production frontier and moving towards the frontier. We are able to claim that our augmented Solow model – capturing initial income, investment, population growth, level and growth of human capital, as well as structural change – is successful in predicting China’s growth rate.

A meaningful decomposition of China’s absolute growth rate is not possible owing to the negative convergence term (comparing initial output with zero output) and the logarithmic form of some explanatory variables, but a decomposition of China’s relative growth is informative. Table 6 presents a detailed decomposition of the differences in the growth predictions for China and Sub-Saharan Africa. We choose Sub-Saharan Africa for research focus as a country group that is representative of least developed countries in 1980 (of which China was one) and which subse- quently experienced growth failure. The actual and predicted annual growth difference between China and Sub-Sah- aran Africa are 6.8% and 5.6%, respectively, i.e. the unexplained residual is 1.2% per annum. When the predicted growth difference is decomposed, we find that capital investment is the most important component (accounting for 54% of the total).

Capital accumulation has traditionally been viewed as an inferior source of growth, in that capital deepening is subject to diminishing returns and will eventually run out of steam. This has not been true for China’s high investment rates for two reasons. Firstly, investment is a major carrier of structural change: structural transformation requires investment in new, normally high-productivity activities. In China, employment growth in the high-productivity industrial and service sectors is determined by the rate of investment in those sectors. The new job opportunities are largely filled by migrant workers from the low-productivity agricultural sector. Secondly, the slow convergence rate predicted by our model, roughly 2% per annum, implies that the average time an economy spends to cover half of the distance between its initial position and its steady state is about 35 years. Therefore, given the diminishing returns to investment, the role of capital accumulation in driving economic growth can persist for decades during the economic transition to the long-run equilibrium.

The role of capital accumulation in driving economic growth in East Asia has been emphasised in several influential studies. For instance, Krugman (1994) argued that the ‘myth of East Asia’s growth miracle’ lay in a mobilisation of re- sources and an extraordinary growth in inputs rather than gains in efficiency. Young (1995) adopted the growth account- ing approach and identified the fundamental role played by factor accumulation in explaining the growth of four East Asian tigers. Our results are in line with these findings but add to the literature by highlighting the role of capital accu- mulation for China in the context of cross-country growth regressions.

As far as other variables are concerned, the fact that the average growth rate of human capital in China was slightly below that in Sub-Saharan Africa over the period 1980–2004 leads us to predict that their growth difference would be smaller by 0.2%. Contributions to that difference came from China’s slower population growth (24%), higher level of hu- man capital (11%), conditional convergence gain (17%) and its more dramatic structural change (12%).

Using this methodology, we are able to account for the predicted growth differences between China and other major country groups as shown in Table 7, based on estimates of Sample I. Growth prediction for China and growth difference pre- diction between China and Sub-Saharan Africa are also reported as a robustness check: we find that our prediction results remain stable when different sample estimates are employed. Our main findings for the other country groups are as follows.

Firstly, conditional convergence, the basic property of the Solow model, has considerable explanatory power for the growth difference across countries, ranging from 21% between China and South Asia to 93% between China and the high-in- come economies.

Secondly, China invests more than other economies, which accounts for more than 40% of the predicted growth difference between China and both East Asia and the Pacific and also South Asia, and 28% in the case of Latin America and the Carib- bean. By stimulating structural change, high investment rates are not only a cause of economic growth but also a symptom of productivity improvement.

Thirdly, reallocation of labour from low- to high-productivity sectors is another source of China’s economic growth. The joint contribution of the linear and non-linear structural change terms to the predicted growth difference ranges from 12% between China and Latin America and the Caribbean to 22% between China and the high-income economies. The role of the structural change terms is consistent with the common view (e.g. World Bank, 1993; Young, 1995) that an important part of productivity growth in low-income economies is attributable to improvement in allocative efficiency through inter- sectoral reallocation of labour from agriculture to industry.

Fourthly, the slower population growth rate of China also contributes to its faster growth relative to other developing countries (accounting for 7% of the predicted growth difference between China and Latin America and the Caribbean; 3% in the case of South Asia; and 1% for East Asia and the Pacific).

15 This is calculated based on the mean value of the prediction for each 5-year interval. Detailed information for each data point is available from the authors.

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 447

Fifthly, the level of human capital explains 16% of the predicted growth difference between China and South Asia. Com- pared with the other country groups shown in Table 7, the level of human capital in China is still quite low.

Lastly, the growth rate of human capital contributes positively to the predicted growth difference of China with the high- income economies, but its contribution is tiny (1%).

6. Conclusions

We have examined the role of the augmented Solow model in explaining China’s remarkable economic growth rate, both absolute and relative to the rest of the world. Following Temple and Wößmann (2006)’s introduction of two sectors into the model, we allowed productivity growth to vary across countries. We extended their cross-section analysis to a dynamic pa- nel data analysis using a robust and consistent system GMM estimator. We have shown the value of cross-country compar- ative analysis in our attempt to explain a phenomenon of great importance to the world.

Firstly, we found that the extended version of the augmented Solow model provides a good explanation of China’s eco- nomic growth, i.e. China’s annual growth rate of GDP per worker (7.2%) falls within the 95% confidence interval for its pre- dicted value (6.3%). The unexplained residual might represent China’s efficiency gains from economic reform and marketization that are not captured by the structural change terms.

Secondly, our model is a valuable means of understanding the large and persistent differences in growth rates between China and other countries. China’s relatively good performance is mainly due to accumulation of physical capital, conditional convergence, improvements in factor productivity through structural change, and slower population growth. The level of hu- man capital contributes to the growth difference between China and other developing countries, but not the growth of hu- man capital. There is room for China to expand its investment in human capital. Our identification of this set of variables provides a framework for the development of growth-promoting policies.

China’s experience shows that rapid growth is indeed possible with imperfect institutions, but it is important in these circumstances that the government addresses the institutional obstacles to growth as they become apparent. The reform of rural and urban institutions from 1978 onwards loosened various binding constraints on growth and helped unleash pre- viously untapped market forces. The preferred model captures the effects of factor accumulation and structural change. Our analysis is silent on the role of several underlying and policy-relevant variables such as institutions, research and develop- ment, financial depth and openness of the economy, each of which is potentially important in the growth process (see, for instance, Quah, 2000). Moreover, since some variables in our growth equations may themselves need to be explained if we are to discover the ultimate drivers of growth (see, for instance, Blomström et al., 1996), it is sensible also to investigate their determinants in China. However, data requirements favour a cross-province rather than a cross-country analysis of such variables.

Acknowledgments

We are grateful to two anonymous referees for their valuable comments on an earlier version of this paper. We also wish to thank participants in the conference ‘China in the World Economy’ in Stockholm, in the Royal Economic Society 2008 annual meeting in Warwick, and in the Chinese Economy and the CSAE seminars in Oxford for their constructive comments and insightful discussion of our paper. The financial support of the Leverhulme Trust is gratefully acknowledged.

Appendix

Table A1 Employment data for China

FAO Data NBS Data Brandt et al. (2008) data

Total labour force (Unit: 10,000 person)

Labour force in agriculture (Unit: 10,000 person)

Share of agricultural labour force (%)

Total employment (Unit: 10,000 person)

Employment in agriculture (Unit: 10,000 person)

Share of agricultural employment (%)

Total employment (Unit: 10,000 person)

Employment in primary sector (Unit: 10,000 person)

Share of primary sector employment (%)

1978 52,946 39,565 74.73 40,152 28,318 70.53 46,843 32,445 69.26 1979 54,017 40,140 74.31 41,024 28,634 69.80 47,967 31,416 65.50 1980 55,104 40,718 73.89 42,361 29,122 68.75 49,397 30,593 61.93 1981 56,258 41,459 73.70 43,725 29,777 68.10 51,039 29,890 58.56 1982 57,429 42,209 73.50 45,295 30,859 68.13 52,618 29,138 55.38

(continued on next page)

Table A2 Samples in this paper

All countries in PWT 6.2 (188 countries) Country isocode PWT order Region Sample 1 (146) Sample II (111) Sample III (61)

Afghanistan AFG 1 SA 1 1 1 Albania ALB 2 EUCA 1 1 0 Algeria DZA 3 MENA 0 0 0 Angola AGO 4 SSA 0 0 0 Antigua ATG 5 HIE 1 0 0 Argentina ARG 6 LAC 1 1 1 Armenia ARM 7 EUCA 1 1 0 Australia AUS 8 HIE 1 0 0 Austria AUT 9 HIE 1 0 0 Azerbaijan AZE 10 EUCA 1 1 1 Bahamas BHS 11 HIE 1 0 0 Bahrain BHR 12 HIE 1 0 0 Bangladesh BGD 13 SA 1 1 1 Barbados BRB 14 LAC 1 1 0 Belarus BLR 15 EUCA 0 0 0 Belgium BEL 16 HIE 1 0 0 Belize BLZ 17 LAC 1 1 0 Benin BEN 18 SSA 1 1 0 Bermuda BMU 19 HIE 1 0 0 Bhutan BTN 20 SA 0 0 0 Bolivia BOL 21 LAC 1 1 1 Bosnia and Herzegovina BIH 22 EUCA 1 1 0 Botswana BWA 23 SSA 1 1 0 Brazil BRA 24 LAC 1 1 1 Brunei BRN 25 HIE 1 0 0

Table A1 (continued)

FAO Data NBS Data Brandt et al. (2008) data

Total labour force (Unit: 10,000 person)

Labour force in agriculture (Unit: 10,000 person)

Share of agricultural labour force (%)

Total employment (Unit: 10,000 person)

Employment in agriculture (Unit: 10,000 person)

Share of agricultural employment (%)

Total employment (Unit: 10,000 person)

Employment in primary sector (Unit: 10,000 person)

Share of primary sector employment (%)

1983 58,636 42,980 73.30 46,436 31,151 67.08 54,117 28,338 52.36 1984 59,897 43,787 73.10 48,179 30,868 64.07 55,810 27,634 49.51 1985 61,224 44,636 72.91 49,873 31,130 62.42 57,551 26,946 46.82 1986 62,630 45,538 72.71 51,282 31,254 60.95 59,151 27,704 46.84 1987 64,106 46,485 72.51 52,783 31,663 59.99 60,744 27,726 45.64 1988 65,618 47,452 72.32 54,334 32,249 59.35 62,240 28,232 45.36 1989 67,116 48,404 72.12 55,329 33,225 60.05 63,561 29,913 47.06 1990 68,563 49,312 71.92 64,749 34,117 52.69 64,749 30,107 46.50 1991 69,524 49,652 71.42 65,491 34,956 53.38 65,491 30,044 45.88 1992 70,413 49,928 70.91 66,152 34,795 52.60 66,152 29,943 45.26 1993 71,248 50,151 70.39 66,808 33,966 50.84 66,808 29,219 43.74 1994 72,056 50,343 69.87 67,455 33,386 49.49 67,455 28,155 41.74 1995 72,858 50,518 69.34 68,065 33,018 48.51 68,065 27,759 40.78 1996 73,657 50,678 68.80 68,950 32,909 47.73 68,950 26,827 38.91 1997 74,446 50,820 68.26 69,820 33,095 47.40 69,820 26,734 38.29 1998 75,223 50,939 67.72 70,637 33,232 47.05 70,637 26,625 37.69 1999 75,979 51,034 67.17 71,394 33,493 46.91 71,394 25,991 36.41 2000 76,711 51,100 66.61 72,085 33,355 46.27 72,085 25,446 35.30

Notes: (1) NBS Data are from China Labour Statistical Yearbook (2003), complied by National Bureau of Statistics of China (NBS) and Ministry of Labour and Social

Security of China; NBS definition of agricultural employment is the number of employment in farming, forestry, animal husbandry and fishery; and NBS definition of total employment is the number of employment in farming, forestry, animal husbandry and fishery; mining and quarrying; manufacturing; production and supply of electricity, gas and water; construction; geological prospecting and water conservancy; transport, storage, post and telecom- munications; wholesale and retail trade and catering services; finance and insurance; real estate trade; social services; health care, sporting and social welfare; education, culture and arts, ratio, film and television; scientific research and polytechnic services; government agencies, party agencies and social organizations; and others. (2) FAO data are from Statistical Database of the Food and Agriculture Organization of the United Nations. FAO definition of labour force in agriculture

(economically active population in agriculture) is that part of the economically active population engaged in or seeking work in agriculture, hunting, fishing or forestry; and FAO definition of total labour force (economically active population) is the number of all employed and unemployed persons (including those seeking work for the first time). It covers employers, self-employed workers, salaried employees, wage earners, unpaid workers assisting in a family, farm or business operation, members of producers’ cooperatives, and members of the armed forces. (3) Brandt et al. (2008) data come from Brandt et al. in Brandt and Rawski (2008, pp. 683–728).

448 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

Table A2 (continued)

All countries in PWT 6.2 (188 countries) Country isocode PWT order Region Sample 1 (146) Sample II (111) Sample III (61)

Bulgaria BGR 26 EUCA 1 1 1 Burkina Faso BFA 27 SSA 1 1 1 Burundi BDI 28 SSA 1 1 0 Cambodia KHM 29 EAP 0 0 0 Cameroon CMR 30 SSA 1 1 1 Canada CAN 31 HIE 1 0 0 Cape Verde CPV 32 SSA 0 0 0 Central African Republic CAF 33 SSA 0 0 0 Chad TCD 34 SSA 0 0 0 Chile CHL 35 LAC 1 1 1 China CHN 36 EAP 1 1 1 Colombia COL 37 LAC 1 1 1 Comoros COM 38 SSA 0 0 0 Congo, Dem. Rep. ZAR 39 SSA 0 0 0 Congo, Republic of COG 40 SSA 1 1 0 Costa Rica CRI 41 LAC 1 1 0 Cote d’lvoire CIV 42 SSA 1 1 1 Croatia HRV 43 EUCA 1 1 0 Cuba CUB 44 LAC 0 0 0 Cyprus CYP 45 HIE 0 0 0 Czech Republic CZE 46 EUCA 1 1 1 Denmark DNK 47 HIE 1 0 0 Djibouti DJI 48 MENA 0 0 0 Dominica DMA 49 LAC 1 1 0 Dominican Republic DOM 50 LAC 1 1 1 Ecuador ECU 51 LAC 1 1 1 Egypt EGY 52 MENA 1 1 1 El Salvador SLV 53 LAC 1 1 0 Equatorial Guinea GNQ 54 SSA 0 0 0 Eritrea ERI 55 SSA 0 0 0 Estonia EST 56 EUCA 1 1 0 Ethiopia ETH 57 SSA 1 1 1 Fiji FJI 58 EAP 1 1 0 Finland FIN 59 HIE 1 0 0 France FRA 60 HIE 1 0 0 Gabon GAB 61 SSA 1 1 0 Gambia, The GMB 62 SSA 1 1 0 Georgia GEO 63 EUCA 1 1 1 Germany GER 64 HIE 1 0 0 Ghana GHA 65 SSA 1 1 1 Greece GRC 66 HIE 1 0 0 Grenada GRD 67 HIE 1 0 0 Guatemala GTM 68 LAC 1 1 1 Guinea GIN 69 SSA 1 1 0 Guinea-Bissau GNB 70 SSA 0 0 0 Guyana GUY 71 LAC 0 0 0 Haiti HTI 72 LAC 0 0 0 Honduras HND 73 LAC 1 1 0 Hong Kong HKG 74 HIE 1 1 1 Hungary HUN 75 EUCA 1 1 1 Iceland ISL 76 HIE 1 0 0 India IND 77 SA 1 1 1 Indonesia IDN 78 EAP 1 1 1 Iran IRN 79 MENA 1 1 1 Iraq IRQ 80 MENA 0 0 0 Ireland IRL 81 HIE 1 0 0 Israel ISR 82 HIE 1 0 0 Italy ITA 83 HIE 1 0 0 Jamaica JAM 84 LAC 1 1 0 Japan JPN 85 HIE 1 0 0 Jordan JOR 86 MENA 1 1 0 Kazakhstan KAZ 87 EUCA 1 1 1 Kenya KEN 88 SSA 1 1 1 Kiribati KIR 89 EAP 1 1 0 Korea, Dem. Rep. PRK 90 EAP 1 1 1 Korea, Republic of KOR 91 HIE 1 1 1 Kuwait KWT 92 HIE 1 0 0 Kyrgyzstan KGZ 93 EUCA 1 1 0 Laos LAO 94 EAP 0 0 0

(continued on next page)

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 449

Table A2 (continued)

All countries in PWT 6.2 (188 countries) Country isocode PWT order Region Sample 1 (146) Sample II (111) Sample III (61)

Latvia LVA 95 EUCA 1 1 0 Lebanon LBN 96 MENA 1 1 0 Lesotho LSO 97 SSA 0 0 0 Liberia LBR 98 SSA 0 0 0 Libya LBY 99 MENA 1 1 0 Lithuania LTU 100 EUCA 1 1 0 Luxembourg LUX 101 HIE 1 0 0 Macao MAC 102 HIE 1 0 0 Macedonia MKD 103 EUCA 1 1 0 Madagascar MDG 104 SSA 1 1 1 Malawi MWI 105 SSA 1 1 1 Malaysia MYS 106 EAP 1 1 1 Maldives MDV 107 SA 1 1 0 Mali MU 108 SSA 1 1 1 Malta MLT 109 HIE 0 0 0 Mauritania MRT 110 SSA 1 1 0 Mauritius MUS 111 SSA 1 1 0 Mexico MEX 112 LAC 1 1 1 Micronesia, Fed. Sts. FSM 113 EAP 1 1 0 Moldova MDA 114 EUCA 1 1 0 Mongolia MNG 115 EAP 0 0 0 Morocco MAR 116 MENA 1 1 1 Mozambique MOZ 117 SSA 0 0 0 Namibia NAM 118 SSA 0 0 0 Nepal NPL 119 SA 1 1 1 Netherlands NLD 120 HIE 1 0 0 Netherlands Antilles ANT 121 HIE 1 0 0 New Zealand NZL 122 HIE 1 0 0 Nicaragua NIC 123 LAC 1 1 0 Niger NER 124 SSA 0 0 0 Nigeria NGA 125 SSA 1 1 1 Norway NOR 126 HIE 1 0 0 Oman OMN 127 MENA 1 1 0 Pakistan PAK 128 SA 1 1 1 Palau PLW 129 EAP 1 1 0 Panama PAN 130 LAC 1 1 0 Papua New Guinea PNG 131 EAP 0 0 0 Paraguay PRY 132 LAC 1 1 0 Peru PER 133 LAC 1 1 1 Philippines PHL 134 EAP 1 1 1 Poland POL 135 EUCA 1 1 1 Portugal PRT 136 HIE 1 0 0 Puerto Rico PRI 137 HIE 0 0 0 Qatar QAT 138 HIE 1 0 0 Romania ROM 139 EUCA 1 1 1 Russia RUS 140 EUCA 1 1 1 Rwanda RWA 141 SSA 1 1 1 Samoa WSM 142 EAP 1 1 0 Sao Tome and Principe STP 143 SSA 0 0 0 Saudi Arabia SAU 144 HIE 0 0 0 Senegal SEN 145 SSA 1 1 1 Serbia and Montenegro SCG 146 EUCA 1 1 1 Seychelles SYC 147 SSA 0 0 0 Sierra Leone SLE 148 SSA 1 1 0 Singapore SGP 149 HIE 1 1 1 Slovak Republic SVK 150 EUCA 1 1 0 Slovenia SVN 151 HIE 1 0 0 Solomon Islands SLB 152 EAP 1 1 0 Somalia SOM 153 SSA 0 0 0 South Africa ZAF 154 SSA 1 1 1 Spain ESP 155 HIE 1 0 0 Sri Lanka LKA 156 SA 1 1 1 St. Kitts and Nevis KNA 157 LAC 1 1 0 St. Lucia LCA 158 LAC 1 1 0 St.Vincent and Grenadines VCT 159 LAC 1 1 0 Sudan SDN 160 SSA 0 0 0 Suriname SUR 161 LAC 0 0 0 Swaziland SWZ 162 SSA 1 1 0 Sweden SWE 163 HIE 1 0 0 Switzerland CHE 164 HIE 1 0 0

450 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

Table A2 (continued)

All countries in PWT 6.2 (188 countries) Country isocode PWT order Region Sample 1 (146) Sample II (111) Sample III (61)

Syria SYR 165 MENA 1 1 1 Taiwan TWN 166 HIE 1 1 1 Tajikistan TJK 167 EUCA 0 0 0 Tanzania TZA 168 SSA 1 1 1 Thailand THA 169 EAP 1 1 1 Togo TGO 170 SSA 0 0 0 Tonga TON 171 EAP 1 1 0 Trinidad and Tobago TTO 172 LAC 1 1 0 Tunisia TUN 173 MENA 1 1 1 Turkey TUR 174 EUCA 1 1 1 Turkmenistan TKM 175 EUCA 0 0 0 Uganda UGA 176 SSA 0 0 0 Ukraine UKR 177 EUCA 1 1 1 United Arab Emirates ARE 178 HIE 0 0 0 United Kingdom GBR 179 HIE 1 0 0 United States USA 180 HIE 1 0 0 Uruguay URY 181 LAC 1 1 0 Uzbekistan UZB 182 EUCA 0 0 0 Vanuatu VUT 183 EAP 1 1 0 Venezuela VEN 184 LAC 1 1 1 Vietnam VNM 185 EAP 1 1 1 Yemen YEM 186 MENA 0 0 0 Zambia ZMB 187 SSA 1 1 1 Zimbabwe ZWE 188 SSA 1 1 1

Note: ‘EAP’: East Asia and Pacific; ‘EUCA’: Europe and Central Asia; ‘LAC’: Latin America and the Caribbean; ‘MENA’: Middle East and North Africa; ‘SA’: South Asia; ‘SSA’: Sub-Saharan Africa; and ‘HIE’: high-income economies including both high-income OCED members and non-OECD members. Country groups by region and income are based on definitions of the World Bank.

S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452 451

References

Abramovitz, M., 1986. Catching up, forging ahead, and falling behind. Journal of Economic History 46 (2), 385–406. Arellano, M., Bover, O., 1995. Another look at the instrumental variable estimation of error-components models. Journal of Econometrics 68, 29–52. Barro, R.J., 1991. Economic growth in a cross section of countries. The 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 3, 541–563. Barro, R.J., Sala-i-Martin, X., 2004. Economic Growth, second ed. The MIT Press. Benhabib, J., Spiegel, M.M., 1994. The role of human capital in economic development: evidence from aggregate cross-country data. Journal of Monetary

Economics 34 (2), 143–173. Blomström, M., Lipsey, R., Zejan, M., 1996. Is fixed investment the key to economic growth? Quarterly Journal of Economics 111 (1), 269–276. Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics 87 (1), 115–143. Bond, S., Hoeffler, A., Temple, J., 2001. GMM Estimation of Empirical Growth Models. CEPR Discussion Paper No. 3048. Bond, S., Leblebicioglu, A., Schiantarelli, F., 2007. Capital Accumulation and Growth: A New Look at the Empirical Evidence. University of Oxford, mimeo. Bosworth, B.P., Collins, S.M., 2003. The empirics of growth: an update. Brookings Papers on Economic Activity 2, 113–179. Bowsher, C.G., 2002. On testing overidentifying restrictions in dynamic panel data models. Economics Letters 77, 211–220. Brandt, L., Hsieh, C.T., Zhu, X.D., 2008. Growth and structural transformation in China. In: Brandt, L., Rawski, T. (Eds.), China’s Great Economic

Transformation. Cambridge University Press, New York, pp. 683–728. Caselli, F., Esquivel, G., Lefort, F., 1996. Reopenning the convergence debate: a new look at cross-country growth empirics. Journal of Economic Growth 1,

363–389. Ding, S., Knight, J., 2008. Can the augmented Solow model explain China’s economic growth? a cross-country panel data analysis. Department of Economics

Discussion Paper No. 380, University of Oxford. Easterly, W., Levine, R., 2001. It’s not factor accumulation: stylized facts and growth models. World Bank Economic Review 15, 177–219. Fernández, C., Ley, E., Steel, M.F.J., 2001. Model uncertainty in cross-country growth regressions. Journal of Applied Econometrics 16 (5), 563–576. Gemmell, N., 1996. Evaluating the impacts of human capital stocks and accumulation on economic growth: some new evidence. Oxford Bulletin of

Economics and Statistics 58 (1), 9–28. Gundlach, E., 2007. The Solow model in the empirics of growth and trade. Oxford Review of Economic Policy 23 (1), 25–44. Hall, R., Jones, C., 1999. Why do some countries produce so much more output per worker than others? Quarterly Journal of Economics 114 (1), 83–116. Hendry, D.F., Krolzig, H.M., 2004. We ran one regression. Oxford Bulletin of Economics and Statistics 66 (5), 799–810. Heston, A., Sicular, T., 2008. China and development economics. In: Brandt, L., Rawski, T. (Eds.), China’s Great Economic Transformation. Cambridge

University Press, New York, pp. 27–67. Hoeffler, A., 2002. The augmented Solow model and the African growth debate. Oxford Bulletin of Economics and Statistics 64 (2), 135–158. Holz, C.A., 2006. China’s reform period economic growth: how reliable are Augus Maddison’s estimates? Review of Income and Wealth 52 (1), 85–119. Hoover, K.D., Perez, S.J., 2004. Truth and robustness in cross-country growth regressions. Oxford Bulletin of Economics and Statistics 66 (5), 765–798. Howitt, P., 2000. Endogenous growth and cross-country income differences. American Economic Review 90 (4), 829–846. Hsiao, C., 1986. Analysis of Panel Data. Cambridge University Press, Cambridge. Islam, N., 1995. Growth empirics: a panel data approach. Quarterly Journal of Economics 110 (4), 1127–1170. Klenow, P.J., Rodrı́guez-Clare, A., 1997. A neoclassical revival in growth economics: has it gone too far? NBER Macroeconomics Annual 12, 73–103. Knight, J., Song, L., 1999. The Rural–Urban Divide. Economic Disparities and Interactions in China. Oxford University Press, Oxford. Knight, J., Song, L., 2005. Towards a Labour Market in China. Oxford University Press, Oxford. Krugman, P., 1994. The myth of Asia’s miracle. Foreign Affairs 73 (6), 62–79.

452 S. Ding, J. Knight / Journal of Comparative Economics 37 (2009) 432–452

Lardy, R.N., 2002. China will keep on growing. The Wall Street Journal Asia, 14. Levine, R., Renelt, D., 1992. A sensitivity analysis of cross-country growth regressions. American Economic Review 82 (4), 942–963. Lichtenberg, F.R., 1993. R&D investment and international productivity differences. In: Giersch, H. (Ed.), Economic Growth in the World Economy. J.C.B.

Mohr, Tuebingen, Germany, pp. 89–110. Maddison, A., 1998. Chinese Economic Performance in the Long Run. OECD, Paris. Mankiw, N.G., Romer, D., Weil, D.N., 1992. A contribution to the empirics of economic growth. Quarterly Journal of Economics 107 (2), 407–437. Mankiw, N.G., Phelps, E.S., Romer, P.M., 1995. The growth of nations. Brookings Papers on Economic Activity 1995 (1), 275–326. McQuinn, K., Whelan, K., 2007. Solow (1956) as a model of cross-country growth dynamics. Oxford Review of Economic Policy 23 (1), 45–62. Nickell, S., 1981. Biases in dynamic models with fixed effects. Econometrica 49 (6), 1417–1426. Pritchett, L., 1999. Where has all the education gone? World Bank Economic Review 15 (3), 367–391. Quah, D., 2000. Cross-country growth comparison: theory to empirics. LSE Department and Economics Working Paper. Ravallion, M., Chen, S.H., 2007. China’s (uneven) progress against poverty. Journal of Development Economics 82, 1–42. Rawski, T., 2001. What is happening to China’s GDP statistics? China Economic Review 12, 347–354. Sala-i-Martin, X.X., 1997. I just ran two million regressions. American Economic Review 87 (2), 178–183. Scott, M.F., 1989. A New View of Economic Growth. Clarendon Press, Oxford. Temple, J., 1999a. The new growth evidence. Journal of Economic Literature 37 (1), 112–156. Temple, J., 1999b. A positive effect of human capital on growth. Economics Letters 65, 131–134. Temple, J., 2001. Generalizations that aren’t? evidence on education and growth. European Economic Review 45, 905–918. Temple, J., Wößmann, L., 2006. Dualism and cross-country growth regressions. Journal of Economic Growth 11, 187–228. Wang, Y., Yao, Y.D., 2003. Sources of China’s economic growth, 1952–1999: incorporating human capital accumulation. China Economic Review 14, 32–52. Windmeijer, F., 2005. A finite sample correction for the variance of linear efficient two-step GMM estimators. Journal of Econometrics 126, 25–51. World Bank, 1993. The East Asian Miracle: Economic Growth and Public Policy. Oxford University Press. Young, A., 1995. The tyranny of numbers: confronting the statistical realities of the east Asian growth experience. The Quarterly Journal of Economics 110

(3), 641–680. Young, A., 2003. Gold into base metal: productivity growth in the People’s Republic of China during the reform period. Journal of Political Economy 111 (6),

1220–1261.