WPS3699
Half a World:
Regional inequality in five great federations
Branko Milanovic1
World Bank and Carnegie Endowment for International Peace
The paper studies regional (spatial) inequality in five most populous
countries in the world: China, India, the United States of America,
Indonesia and Brazil in the period 1980-2000. They are all federations or
quasi-federations composed of entities with substantial economic
autonomy. Two types of regional inequalities are considered: Concept 1
inequality which is inequality between mean incomes (GDPs per capita) of
states/provinces and Concert 2 inequality which is inequality between
population-weighted regional mean incomes. The first inequality speaks to
the issue of regional convergence, the second, to the issue of overall
inequality as perceived by citizens within a nation. All three Asian
countries, show rising inequality in terms of both concepts in the decade
of the 1990's. Divergence in income outcomes is particularly noticeable
for the most populous states/provinces in India and China. United States,
where regional inequality is the least, shows further convergence. Brazil,
with the highest level of regional inequality, displays no trend. A
regression analysis fails to establish robust association between the usual
macro variables and the two types of regional inequality.
Key words: China, India, USA, Brazil, Indonesia, regional inequality, world inequality
JEL classification: R12, I3, 057
World Bank Policy Research Working Paper 3699, September 2005
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the
exchange of ideas about development issues. An objective of the series is to get the findings out quickly,
even if the presentations are less than fully polished. The papers carry the names of the authors and should
be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely
those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors,
or the countries they represent. Policy Research Working Papers are available online at
http://econ.worldbank.org.
1I am grateful to Tamar Manuelyan Atinc, Roberto Zagha, Bala Bhaskar Naidu and Bert Hofman for
providing most of the regional GDP data used in this paper. Gouthami Padam was an excellent research
assistant, and I am grateful to her. I am grateful to Isabelle Bensidoun, Guillaume Daudin, Bert Hofman,
Ravi Kanbur, Bert Keidel, Christian Morrisson, Mansoob Murshed, Martin Ravallion, Erik Thorbecke,
Xiaobo Zhang, an anonymous referee, and participants of a conference on inequality in China at Cornell
University and at Brookings Institution for very useful comments. Work was financed by a grant from
McArthur Foundation and by the World Bank Research Grant RPO 85725. Email:
bmilanovic@worldbank.org.
1
2
1. Introduction
Global inequality is strongly influenced by what happens to populous countries,
both to their average income levels, and inequality within each country. If we look at the
former, fast growth rates of China and more recently of India, by reducing their relative
income distance from the United States and other rich countries, have lowered global
inequality. Fast (average) growth in these nations is the most important element
contributing both to lower global inequality and to lower global poverty. At the same
time however within these nations, there are increasing cleavages: some provinces grow
faster than others, so mean income differences between the provinces increase. In
addition, the process of growth is accompanied in both China and India by increasing
inequality between individuals in the country as well as within most provinces and states
(see for example, Jha (2004) for India; Chen and Wang ( 2001) for China).
The objective of this paper is to explore the subnational changes in the five most
populous countries in the world: China, India, United States, Indonesia, and Brazil. We
examine these changes for three reasons. First, as already mentioned, subnational
changes in populous countries are important because they are a significant determinants
of global inequality. The second reason lies in what subnational changes imply for
national unity. One of the concerns of the Chinese leadership, for example, is how rising
income differences between the prosperous coastal provinces and the less dynamic North
East may affect China's political unity (see Renard, 2002). The third reason to look at
2
the subnational level is that it provides us with some lessons regarding what we might
expect at the global level were most of the obstacles to free circulation of labor, capital
and goods to be lifted. For obviously although these obstacles still exist in a number of
countries even between their constituent units, they are dramatically lower than between
nations. For example, shipment of goods between different Indian states is, in many
2In June 2004, nine Chinese provinces (Yunnan, Guizhou, Guangdong, Fujian, Jiangxi, Hunan, Guangxi,
Hainan and Sichuan) and the special administrative regions of Hong Kong and Macau created a Pan Pearl
River Regional Cooperation Council whose objective is to eliminate nontariff barriers between the
provinces and to coordinate provincial economic policies. One of the objectives was to speed up growth in
order to catch up with Shanghai and the booming Yangtze River basin (see International Herald Tribune,
"Chinese provinces form a trading bloc", June 2, 2004, p. 17).
3
cases, subject to fees or border check but this is a far cry from the obstacles to
international trade. Similarly, labor mobility in China is formally restricted as people
need to have a permit in order to live in large cities but these rules are most often ignored,
and even if they were fully implemented, the rules are much less stringent than the rules
governing circulation of labor between nations. It is thus important to find out whether a
generally unhindered movement of goods, labor and capital leads provinces toward
convergence in their mean incomes or not.
A comparison between subnational inequalities in these five countries is
meaningful despite the fact that the number and size of regional units vary. If states and
provinces were based on randomly drawn border lines, it could be indeed argued that
greater regional inequality in one country compared to another is simply an artifact,
possibly due to the way such regional units are defined. However, in most or all cases
with which we deal here, provinces/states possess both a significant amount of autonomy
and specificity. Inequality between them thus does have a bearing on the issues of
horizontal equity, and particularly so where states differ in their religious and ethnic
composition.
The structure of the paper is as follows. In Section 2, we present some data that
show the importance of the five countries in total world income, population and global
inequality. In Section 3, we move to the subnational level of analysis and present a brief
review of the literature dealing with the issue of regional inequality in the five countries.
In Sections 4, we look at the evolution of the differences between mean incomes of the
regions during the last twenty years in each of the five nations.3 It is here that we address
the issues of income convergence or divergence within each country. In Section 5, we
look at the differences in mean regional incomes but now with each region weighted by
its population share. Once we weigh them by their populations, we come closer to
3We use the term "region" to indicate a subnational unit whether its exact appellation is "state" or
"province" or "republic"; the term "nation" or "country" is used to designate the subjects of the
international law, that is the five countries included in the analysis here.
4
explaining the actual "feeling" of inequality within each nation particularly when
differences in mean regional incomes coincide with other horizontal cleavages like
ethnicity or religion. We also come closer to explaining the overall national inequality
between individuals because the only part which is not included now is the part of
inequality--at times substantial, it is true--which is due to inter-personal income
differences within each region. In Section 6, we try to identify the policies or factors at
national level which might explain the process of regional convergence or divergence; in
other words, we try to find some regularities explaining why regions are converging or
diverging. Section 7 presents the conclusions.
2. The importance of the five countries in the world
Table 1 shows the share of the five most populous countries in world population
and income in 1980, 1990 and 2000. Their share in world population has throughout
remained at just above one-half. However, their share in world output has expanded from
a bit over one-third to 44 percent.
Table 1.Share of the five largest countries in world population and world income (in %)
1980 1990 2000
Share in world 51.7 50.6 50.8
population
Share in world PPP 34.3 36.6 44.0
output
The increase in the five countries' share in world income was driven by the
growth (compared to world mean income) of all countries save Brazil (Table 2). The
most extraordinary was of course China's growth. In 1980, China's GDP per capita was
just under 17 percent of the world average. In 2000, it was in excess of 60 percent. The
United States likewise grew faster than world mean income: its GDP per capita was 4.3
times greater than the world mean in 1980, while in 2000, it was almost 4.6 times as high.
5
Table 2. The five large countries' GDPs per capita compared to the world mean
1980 1990 2000 Change
(1980-2000)
India 0.164 0.206 0.246 +0.082
China 0.169 0.307 0.602 +0.433
Indonesia 0.364 0.495 0.527 +0.163
Brazil 1.029 0.869 0.822 -0.207
United States 4.308 4.470 4.579 +0.271
Note: All calculations done for GDP per capita expressed in 1995 international (PPP) dollars. Source:
World Bank.
What is the importance of these countries for global inequality? Being the most
populous countries in the world, they will obviously influence, sometimes decisively,
global inequality. However when we address the issues of global inequality we have to be
very careful in what exactly we are measuring. The first concept of inequality (dubbed
Concept 1 by Milanovic, 2005)) measures differences in mean incomes between
countries. No weighting is involved here and each country counts the same. This concept
is important if we are interested whether there is unconditional divergence or
convergence of mean (countries') incomes in the world, but there our five countries
count as much as any other five countries in the world.
Concept 2 inequality is different. It likewise takes mean national incomes but
weighs them by countries' populations. Now, of course, our five countries will matter a
lot because they account for almost one-half of world income and population. But note
that when we use Concept 2 inequality, we disregard the entire inequality due to the
differences in personal incomes within nations, that is we assume that each individual has
the mean income of his/her country. Concept 3 inequality (inequality between all
individuals in the world) is accordingly underestimated by Concept 2 inequality. 4
Table 3 shows that the differences in mean incomes (GDPs per capita) between
the five countries (that is, disregarding their interactions with other countries) account for
between a fifth and a third of total Concept 2 inequality in the world (depending on
4However about ¾ of Concept 3 inequality in the world is accounted for by the differences in countries'
mean incomes (Milanovic, 2002 and 2005).
6
whether we use Gini or Theil index). It is important to explain how this calculations is
done. Consider the Gini coefficient. In the case of Concept 2 inequality, it is equal to
1 n n
(1)
( yj - yi)pipj
i j>i
where yi = GDP per capita of i-th country (and yj>yi), pi = population share of i-th
country, and = mean world income. Thus, Gini Concept 2 for the world is calculated by
taking the absolute differences in mean incomes between all countries in the world,
normalizing them by the world mean income and weighing by the population shares. To
calculate the part due to the five countries we "extract" only those individual inter-
1
country terms (ICT), that is
(yj - yi)pipj , where both i and j belong to "our" five
countries. Thus, for example, in 1980, the US GDP per capita was $PPP 21,755 and
China's $PPP 852. The difference between the two amounted to $PPP 20,900. In terms of
world mean income (which was then $5,050), the difference was 4.13 times. This is in
turn weighted by US and Chinese population shares (0.054 and 0.234 respectively), and
the obtained value 5.3 Gini points represents the contribution of the difference in mean
incomes between the United States and China to Concept 2 world Gini. The same
calculation is repeated for all five countries (there is in total 10 such calculations). The
sum gives the total amount of the direct contribution due to the interaction between the
five countries. As Table 3 shows, it did not change much in the last twenty years: it
oscillated around 11 Gini points throughout.
Note however that these calculations leave out all the similar interactions between
each of our five countries and the rest of the world (say, China vs. Germany, or US vs.
Nigeria, or India vs. Senegal etc.) which also contribute to world inequality. If we add all
these contributions, we notice that they have declined from almost 17 Gini points in 1980
to 14 Gini points in 2000. The decline was almost entirely due to China's fast growth
that diminished the distance, and hence the inequality terms, between herself and the rest
of the (rich) world. Overall the five largest countries contribute around one-half of total
global Concept 2 inequality as measured by the Gini coefficient.
7
Differently, we can measure inequality by using the Theil index. The Theil index
for Concept 2 inequality can be written:
p n yi yi
i ln (2)
i=1
yi yi
Here we take the same approach: we "extract" only the pi ln terms
belonging to our five countries. The contribution of the US is obtained (again using the
1980 data) as the product of its population share (0.054) with its income relative to the
world mean (20,900/5050=4.308) times the log of the last amount. To get the total
contribution of the five countries, we simply add such amounts. 5 Their contribution to
the global Theil-measured Concept 2 inequality varies between one-third in 1980 and 37
percent in 2000 (see Table 3).
Table 3. Direct and total contribution of the five largest countries to
Concept 2 global inequality
1980 1990 2000
Gini
Concept 2 Gini (world) 54.1 52.9 50.2
Interaction between the five countries 11.6 10.8 11.3
(Gini points)
Interaction of the five with other 16.9 15.6 14.0
countries (Gini points)
Total percentage contribution 53 50 50
Theil entropy
Concept 2 Theil (world) 63.3 59.8 52.3
Contribution of five largest countries 21.0 17.8 19.4
(Theil points)
Percentage contribution 33 30 37
Note: Concept 2 inequality is inequality between mean countries' GDPs per capita weighted by
countries' populations. Both Gini and Theil expressed in percentages.
5Unlike the Gini coefficient which is based on bilateral comparisons of income between all countries and
allows us to distinguish the contribution of each and every pair of countries, Theil index just sums
individual countries' contributions. In other words, it cannot distinguish between the contribution due to the
interaction between the five countries themselves, and the five countries and the rest of the world.
8
3. A brief review of regional (within-country) inequality studies6
The issues we address here--Concept 1 and Concept 2 inequality within nations--
have been, in slightly different contexts, addressed before. This was done in two contexts.
The first is the issue of regional inequality within countries. There are two views in the
literature that are often juxtaposed. The first is due to Williamson (1965) who argued
that in the early stages of economic development, regional inequality would tend to rise
as growth occurs in discrete locales, but that later inequalities will decline as equilibrating
forces such as better infrastructure, technological diffusion, decreasing returns to capital
in richer and high-wage areas, diseconomies of agglomeration etc. become stronger.
Thus, regional inequality is expected to follow an inverted U shape as income level
grows. Williamson's reasoning is closely related to the idea of the Kuznets curve where
similar development although not in spatial terms produces first an increase and then a
decline in inequality. It is also based on the neoclassical (Solow-type) assumptions which
include decreasing marginal returns. A different view has been proposed more recently
within the context of the new economic geography school (Krugman and Venables,
1995) and endogenous growth (Romer 1986; see also a recent review by Easterly and
Levine 2002). There the argument is that that increasing returns to scale and advantages
of agglomeration of capital and knowledge will tend to perpetuate, or even increase,
spatial inequalities. Yet in Krugman and Venables (1995), decreasing transportation costs
may play an offsetting role: assume that transportation costs are zero, then the advantage
of cheap labor in less developed countries (or regions) will, to some extent, tend to offset
the advantages of increasing returns to scale.
The key difference between these two approaches seems not to lie in their view of the
short-run developments, where they all, including the earlier development theories such
as Myrdal's (1957), Rosenstein-Rodan's (1943), Hirschmann (1958) or Perroux (1970;
1988), seem to agree that growth is disequalizing, but in their view of the long-run
developments where either traditional neoclassical assumptions dominate--rendering
6Since we deal with regional inequality, we do not review studies of the most common Concept 3 (inter-
personal) inequality.
9
growth ultimately equalizing in spatial terms as well--or where such assumptions are
rejected or made less potent thus weakening the forces which make for spatial equality. 7
Recently, the short- and long-run aspects have been combined in a paper by Petrakos,
Rodriguez-Pose and Roviolis (2003), which looks at the regional inequalities within the
European Union (with several regions defined within each country). The authors find
that that the short-term effects of growth are disequalizing in the sense that higher growth
rate tends to increase regional inequality (controlling for all other country-relevant
attributes), while higher income level is associated with lower regional inequality. The
authors interpret the second finding as implying the long-run equilibrating effects of
growth along the lines of the Solow and Williamson models. Their measure of regional
inequality, as in several other papers (e.g. Akita and Kataoka (2003) regarding Japan;
Akita and Kawamura (1992) regarding Indonesia and China; Bhalla, Yao and Zhang
(2003) and Kanbur and Zhang (2003) for China) is the population-weighted coefficient of
variation or population weighted Theil index. This is what we called Concept 2
inequality, and the justification for using Concept 2 (rather than Concept 1) inequality is
that it reflects regional inequality as experienced by an "average" person; in other words,
regional divergence which may be due to a few sparsely populated regions' either very
fast or very slow growth is rather irrelevant for the actual feeling of spatial (horizontal)
inequality as experienced by the people in the country.
But the issue of regional inequality--using Concept 1 inequality--has also recently
received quite a lot of prominence due to the popularity enjoyed by the so-called growth
convergence literature. While the convergence issues have originally been defined and
studied at the level of countries (that is, convergence of national economies), they have
recently been studied at the level of subnational regions. For the countries included here,
such papers are Zhang, Liu and Yao (2001) for China, Azzoni (2001) for Brazil, Ram
(1992) for the United States, and Jha (2004) and Dreze and Srinivasan (2000) for India.8
The rationale for the interest in Concept 1 inequality is very different from the interest in
7There is an obvious link between these views, as couched in terms of regional developments within
individual countries and regional developments in the world as in Krugman and Venables (1995) or
Krugman (1991).
10
Concept 2 inequality. The issue of convergence of (unweighted) regions within a country,
or (unweighted) countries in the world, addresses the problem of whether the same or
similar economic policies produce similar results or not. Consider the example of a
country which has a single national economic policy, that is where there is no significant
regional freedom of economic policy-making. Suppose that Concept 2 regional inequality
is decreasing. But if we still find that Concept 1 inequality is increasing, it raises the
interesting question of what characteristics enjoyed by some regions are responsible for
their not catching up (or for their growing too fast). Thus both Concept 1 and Concept 2
inequality are of interest.
We shall now briefly review some of key (representative) regional inequality studies
that deal with the five great federations included here. 9 The studies of China are the most
numerous. There are two reasons for this. First, the extremely fast growth of Chinese
economy over the last quarter century has been associated with increasing regional
inequalities. This has obvious implications both for political stability and for economic
theory, that is for figuring out why and how certain regions grow and others don't, and
whether the dominant feature of China's inequality is rural vs. urban inequality, or inter-
provincial inequality. The consensus seems to be that it is the former. 10 For example,
Bhalla, Yao and Zhang (2003) calculate inequality in per capita consumption across
provincial and urban-rural partitions (that is, they use data for mean annual incomes for
rural and urban areas for each province, that is 28 provinces times 2 = 56 observations11)
and find that in 1995, more than ¾ of thus calculated Theil Concept 2 inequality is
accounted for by the rural-urban split.12 This is a result similar to the one obtained by
Kanbur and Zhang (2003) who find that the same urban-rural split (that is, the difference
8 For other countries, see Goerlich and Mas (2001) for Spanish provinces, Yemtsov (2002) for the
subjects of the Russian Federation.
9 We use the term "federation" in a technical sense, to indicate that the constituent parts do have some
power of economic decision-making and represent meaningful entities in historic, ethnic or religious sense.
Not all of the five countries studied here are federations in a juridic sense of the word (see also Annex 1).
10When it comes to inter-provincial inequality, it seems to be more the case of "club" inequality, that is of
three clubs (East, West, Center) diverging from each other (see Yao and Zhang 2001).
11More exactly, they have the data on mean per capita consumption of peasants and non-peasants (by
province) as obtained from Chinese household surveys. They interpret peasant consumption to be rural, and
likewise non-peasant to be equivalent to urban (see Bhalla et al., p. 945).
12It is notable that the share of the urban-rural difference in total appears constantly high, that is between
70 and 80 percent, from 1952 to the end of century (see Bhalla et al (2003, Table 2, p. 947)).
11
between mean urban and mean rural income by province) explains 56 percent of
Concept 2 inequality (calculated from the same 56 observations).13 The results of these
two studies and a few others are compared in Annex 2.
The second reason lies in the lack of individual-level data on income inequality in
China, that is lack of data on Concept 3 inequality. Concept 2 regional inequality can be
14
used, if our partitioning is sufficiently fine, to approximate the evolution in Concept 3
inequality. In other words, if we think that most of inequality is spatial, and use a very
fine partition (divide the country in meaningful but also numerous regional units) then
thus calculated Concept 2 inequality should approximate, if not necessarily the level then
the evolution, of Concept 3 inequality. To see this consider that, at the extreme, every
individual can be treated as a "region": then Concept 2 inequality collapses into Concept
3 inequality. This was, for example, the approach underlying Kanbur and Zhang (2003)
paper on regional inequality in China. As in the Bhalla et al. (2003), Kanbur and Zhang
divide China into 28 provinces and each of the provinces into its rural and urban areas.
They have the data for mean incomes for each of thus defined 56 regions for the period
1952-2000. They calculate Concept 2 inequality from these means, find that the rural-
urban split accounts for the bulk of total inequality (much more than the inland-coastal
split)15 but use the Concept 2 inequality as a proxy for the Concept 3 inequality. Then
they try to relate changes in Concept 2 inequality levels over the last 50 years to various
policy episodes (Great Leap Forward, Cultural Revolution, agricultural liberalization,
urban and industrial liberalization etc.) They find that the Concept 2 inequality, with this
relatively fine partition, amounts in 2000 to a Gini of 37.2 which, of course, sets a rather
high lower bound to total personal income inequality in China.16
13Why these two results are not the same is puzzling. A comparison of the results of these two studies and
a few others can be found in Annex 2 of the Internet version of this paper available at
http://econpapers.hhs.se/RAS/pmi44.htm
14Of course, there are many studies of inter-personal inequality in China. They are however all
approximations based on the published group data from national surveys since Chinese authorities have
been unwilling to share micro (individual-level) data.
15Although after 1993, there is a rapid increase in the within-urban and within-rural components indicating
that there are widening income differentiations within urban and within rural areas as well. ("Within" in this
context means "between mean incomes of rural (or urban) provincial incomes.")
16When we use regional GDP per capita as welfare indicators (as here) to derive Concept 2 inequality, its
value cannot be fully taken as the lower bound of Concept 3 inequality because of likely transfers between
12
As for other countries, Jha (2004), in a study of India's inequality over the last fifty
years looks at the issue of Concept 1 convergence between the states and concludes that
divergence has been more common and has accelerated since the reforms in the early
1990's. Ram (1992), and Barro and Sala-i-Martin (1992) have done similar analysis for
the United States. Ram (1992) finds a steadily decreasing inter-state inequality from 1950
to 1980 and an increase in the next decade. For Brazil, Azzoni (2001, p. 137) shows
decreasing Concept 1 inequality throughout the 1950's and 1960's and a stable one in the
last twenty years.
As this brief review shows, regional inequality studies fall into three categories that
closely match our three concepts of inequality. Many of them, in the wake of the
convergence literature, deal with Concept 1 inequality. Others, perhaps equally
numerous, deal with Concept 2 inequality--regional or horizontal inequality as actually
"experienced" by the population. Finally, some use regional partition (Concept 2
inequality) as a proxy for Concept 3 inequality.
It should be noted that the work on regional inequality is not facilitated by the
absence of an accepted or clear terminology. The results are often impenetrable because
the same term, say "regional inequality" may be used to represent Concept 1 or Concept
2 inequality. Even the term "region" is sometimes used for the smallest units (say, states
in a country) and, perhaps in the same paper, for the agglomeration of several such units
into a larger whole which is still not national level (thus, for example, authors often write
of China's three regions: East, Center and West, and of China's regions, meaning in the
latter case provinces). As a consequence, "regional inequality" might mean either one of
the four combinations: Concept 1 or Concept 2 inequality, or inequality between the
provinces or inequality between agglomerations of provinces, that is larger "regions." In
addition, the share of the between (inter-regional) component is quite different--even if
the regions. This is different from the analysis on the global level when redistributive transfers between the
countries are minimal. I owe this point to Christian Morrisson.
13
the partitions are the same--if we look at the "between" share in total interpersonal
(Concept 3) inequality or in Concept 2 inequality.17
Even the measures of inequality are often opaque. While many authors use Theil
indexes because of their decomposition properties, it is not always clear if Theil (1) also
known as Theil entropy measure, or Theil (0) also known as mean log deviation index is
used. While the use of one or another Theil index does not have an impact on results
within each individual study (since the measures move almost always the same way), it
does render difficult comparison of the absolute values of inequality measures from
several different studies.18 We have tried to be as clear as possible in our terminology
here and to the risk of being repetitive will use "Concept 1" and "Concept 2" inequality
often in order to avoid possible misunderstanding. Similarly, we shall try to be clear
about the "partition" we use: states (provinces) adding up to a nation, or states
(provinces) adding up to regions which then add up to a nation.
4. The five federations: descriptive statistics and Concept 1 inequality
The analysis of each individual country's inequality is conducted on the basis of
regional GDPs per capita expressed in nominal prices. 19 Clearly, to the extent that there
are price differences within a country and price levels are higher in richer regions, such
inequality statistics will tend to overestimate intra-regional inequality.
In order to allow for comparison between regions belonging to different countries,
all regional GDPs are also expressed in real terms. First, they are expressed in real units
of the domestic currency and then converted into the 1995 international dollars ($PPP).
The typical pattern of conversion is the one for China as expressed in equation (1) where
Y denotes provincial GDP per capita, r = real provincial growth rate, DD = country-wide
17For example, in a very detailed summary of the results of the "between" component in different country
studies given in Sharrocks and Wan (2004, Table 1A), total inequality against which the "between"
component is measured is sometimes Concept 2 and sometimes Concept 3 inequality.
18Yet another problem seems to be that some authors use natural logs and others use logs with the base of
10. Again, this does not matter for individual studies, but does matter for comparative work.
19Except in Indonesia for which we have only real provincial incomes.
14
deflator, and PPP = PPP-equivalent dollar exchange rate. We have, from the Chinese
statistics, real annual provincial growth rates expressed in 1978 all-China prices. This
means that real income of region i, country j (=China here) and year t (subsequent to
1978) is obtained by applying real growth rates to the nominal 1978 regional income.
Note that the application of real growth rates from region to region implies that
differences in annual inflation between the regions are accounted for. This amount, that
is provincial income in year t expressed in 1978 prices (the denominator in equation 3) is
then, using the country-wide deflator between 1978 and 1995 (DDj,78,95), converted into
all-China 1995 prices.20 This is finally converted into international dollars by using the
1995 purchasing power exchange rate of the country (obtained from the World Bank
data).
Yi, j, t,95,$=Yi, j, t - 1,78, d*(1+ ri, )
j, t* PPP95 (3)
DDj( 78,95)
where the subscript d denotes domestic constant prices, and subscript $ constant
international prices.21
A slightly different approach is used for US regions. Nominal incomes (year t
expressed in dollars of the same year) are converted using the overall country's deflator
(DDj) which converts t year's dollars into the 1995 dollars. This implies that price
differences across US states are non-existent.22 Since the US prices are, by construction,
made to equal to international prices, the PPP convertor is equal to 1 (equation 4)
Yi, j, t,95,$=Yi, j, t, t, ,$*1 (4)
DDj (t,95)
Table 4 shows the descriptive statistics and our data coverage of the five
countries. More detailed information regarding countries' administrative structure is
given in Annex 1. Figure 1 shows Concept 1 inequality for the five countries calculated
using Gini and Theil indexes.
20Note that the deflator has only country (j) subscript.
21A special caveat must be made regarding Chinese national account data: see the discussion in Annex 1.
22There is also no state-level CPI in the United States.
15
We see that the unweighted level of inter-regional inequality is the highest in
China, next come Brazil and Indonesia,23 followed by India and the lowest level of inter-
regional inequality is in the United States, despite the fact that the US has the highest
number of states and that there is a presumption that Concept 1 inequality will increase
with greater number of partitions (see Proposition 1 in Shorrocks and Wan, 2004). As we
can see from Figure 1, the Concept 1 Gini is about 30 for China and Brazil, 25 for
Indonesia, less than 20 for India, while for the United States is it around 10. The ranking
of the countries is the more or less same if we use Theil index--except that in the year
2000 Indonesia displays higher inequality than Brazil.
But the figures also reveal very different time patterns in the evolution of Concept
1 inequality. China which has the highest levels of inequality displays also significant
reductions throughout the 1980's. The decline in regional inequality is particularly
impressive if we look at the Theil index which puts greater emphasis on what happens at
the tails of the distribution. The period of regional equalization however comes to a close
around 1990 and is reversed afterwards. On the contrary, Indonesia and India show
increasing regional inequality throughout the entire period; Brazil, no noticeable trend.
Finally, the United States which almost throughout has the lowest level of inter-regional
differences, shows a burst of increased inequality in the early 1980s, followed by a
sustained if weak reduction since.
When we look at the evolution of inter-regional inequality within a country, we
need to distinguish between inter-regional inequality calculated in nominal and real
prices. As discussed earlier, the former is simply a comparison of regional nominal GDPs
per capita; the latter is calculated based on GDP per capita and a price structure of a
given year which is then "augmented" by annual real growth rates. The results presented
in Figure 1 are based on nominal magnitudes with the exception of Indonesia where we
23In accordance with the rest of the literature (see Akita and Kawamura, 2002, and Akita and Lukman,
1995) regional inequality statistics for Indonesia are based only on non-oil and gas part of the GDP.
Inequality is significantly greater if we include oil and gas but on the other hand most of these revenues are
not retained at the provincial level but are appropriated by the Center and then redistributed. Hence such
inequality would overestimate actual regional inequality in the standard of living of the population.
16
had access to real data only. However, in China and India nominal and real inequality
measures do not move in parallel.
17
Table 4. Regions in the five countries: descriptive statistics
China India USA Brazil Indonesia
Period included 1978-2001 1980-2000 1977-2001 1985-2001 1983-2001
Welfare indicator GDP per GDP per GDP per GDP per GDP per
capita capita capita capita capita
Constant prices (year) 1978 1980-81 1995 1985 1983
Number of regions 29 14 50 26 26
included in the study
Total number of regions 30 25 50 26 26
in the country
Not included in the study Hong Kong, All Union District of Federal --- 3/
(approx. population in m) Macau, territories Columbia, District, 2
million, (Delhi etc.), Pacific and million
Taiwan 11 million Caribbean
(China) 29 possessions,
million 5 million 2/
Total population 1,271 1,033 283 172 213
(million; in 2000)
Population coverage (in 99 92 100 99 100
percent; 2000) 1/
Most populous region 114.2 170.6 (Uttar 34.0 37.1 (Sao 35.7 (West
(million, 2000) (Sichuan) Pradesh) (California) Paulo) Java)
Least populous region 5.6 (Ningxia) 20.6 (Haryana) 0.5 0.3 (Roraima) 1.2 (Maluku)
(million, 2000) (Wyoming)
Richest region (year 21.7 3.1 42.0 9.1 (Sao 13.7 (East
2000; in tho. 1995 (Shanghai) (Maharashtra) (Connecticut) Paulo) Kalimantan)
international dollars)
Poorest region (year 1.6 (Guizhou) 0.7 (Bihar) 20.0 (West 1.3 1.1 (East Nusa
2000; in tho. 1995 Virginia) (Maranhao) Tenggara)
international dollars)
Ratio: richest-poorest 13.6 4.4 2.1 7.0 12.5
region (2000)
Median region by 3.3 (Anhui) 1.7 (Kerala) 28.8 3.1 2.4 (East Java)
income (year 2000; in (Pennsylvania) (Pernambuco)
tho. 1995 international
dollars)
GDP per capita (year 4.1 1.7 31.5 5.7 3.6
2000; in tho. 1995
international dollars)
Note: For the purposes of comparison, GDP per capita is always expressed in 1995 international dollars. No region-specific
PPP rates are used (e.g. all regions of a given country have their GDP per capita converted into $PPP using a single exchange
rate). 1/ Population in regions included in the study divided by total population as defined here. 2/ Pacific and Caribbean
possessions include 12 territories only 5 of which (Puerto Rico, Guam, US Virgin Islands, American Samoa and Mariana
Islands) have resident population apart from the military. Their total population is 4.4 million (see
http://www.infoplease.com/ipa/A0108295.html). 3/ East Timor was included in the Indonesian data until independence in
1999. For consistency, we have excluded it throughout.
18
Figure 1. Concept 1 (unweighted inter-regional inequality) in the five countries
Gini Theil
0.400 0.300
China
0.300
Brazil
0.200
Indonesia China
i li
0.200
Gin The Brazil
India
Indonesia
0.100
0.100 USA
India
USA
0.000 0.000
1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 1977 19781979 1980 1981 1982 1983 1984 1985 1986 19871988 1989 19901991 1992 19931994 1995 1996 1997 1998 1999 2000 2001
Years Years
Note: Concept 1 inequality is calculated in nominal terms (except for Indonesia where we have real data
only). GDP for Indonesia excludes oil and gas data.
Figure 2. China: Concept 1 inequality calculated in nominal
and real terms (1978 prices)
0.40
0.36 reals
0.32
nominals
0.28
0.24
0.20
1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
Note: Provinces of Qinghai and Hainan (populations of respectively 5 and 7 million) are not
included in the calculations based on real GDP per capita due to the lack of data. They are included in the
nominal GDP data series though.
19
Figure 2 shows very different evolutions of the two inequalities in the case of
China. If we look at nominal differences in GDP per capita, we note a strong reduction
in inter-provincial inequality up to 1990, and a slightly increased level since. The decline
in nominal inequalities (which is even more dramatic if we look at the Theil than at the
Gini measure and thus implies that the catch up has been particularly strong for the poor
provinces) coincides with the period of agricultural liberalization. Since 1990,
approximately around the time when China enters more substantial liberalization in the
urban sector, this trend is reversed. However if we look at the inter-provincial inequality
in real terms, we see that it remained around Gini of 35 throughout the entire period. The
implication of these different movements in real and nominal Gini and Theil is that there
has been a price catch up of the poorer provinces (or that the output mix produced in poor
provinces has moved toward products whose prices have been rising more than the
average). In other words, if in 1978, the poorer provinces had a lower price level,24 then
while their real growth rates did not systematically differ from those of the rich
provinces, their relative price level must have risen in order to observe a decline in
nominal inter-provincial inequality. After 1990, the two measures move the same way
indicating that the price catch-up has stopped.
24This is a very sensible supposition although we cannot prove it since Chinese statistics publish only
growth rates of the provincial CPIs but do not provide provincial price levels.
20
Figure 3. India: Concept 1 inequality calculated in nominal
and real terms (1980-81 prices)
0.3
0.2 nominals
reals
0.2
0.1
0.1
0.0
1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
Note: The data on 1999 real GDPs per capita are incomplete and hence not included.
In India, both real- and nominal-based Concept 1 inequality was rising during the
last twenty years (Figure 3). However, the same phenomenon of the poorer states' catch-
up in terms of price levels is obvious here too as real-based inter-state inequality is higher
than the inequality calculated from nominal state GDPs per capita.
In Brazil, inter-state inequality moves up during the periods of high and volatile
inflation: between 1986 and 1990 inflation accelerated from 150 percent annually to
almost 3000 percent and inter-state inequality, measured by nominal GDPs per capita,
rose. (However inequality measured in real terms stayed remarkably constant; Figure 4).
Deceleration of inflation in 1991 was accompanied by a decline in inter-state income
differences. But during the next three years as inflation accelerated again (from 400
percent in 1991 to more than 2000 in 1994), nominal inter-state inequality rose again.
This is an evolution very similar to what we observe for the United States where also
inter-state inequality rose during the inflationary period 1977-1981. In both cases, it
seems that nominal GDPs per capita are very differently affected from state to state when
inflation is high. This is not unexpected. We know that high inflation and hyperinflation
are accompanied by greater price volatility, that is real prices of goods and services vary
more in high inflation environments (see Parks (1978) and for more recent evidence,
21
Dabus (2000)). It is then not surprising that different states--which produce different
output mixes--will also tend to be affected more unevenly when inflation is high than
when it is low.
Figure 4. Brazil: Concept 1 inequality calculated in nominal and real terms (1995 prices)
0.40
0.35
0.30
reals
nom inals
0.25
0.20
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001
Note: The state of Tocantins, due to lack of data, is not included in years 1985-1988.
5. Concept 2 regional inequality
In principle population-weighted inequality (Concept 2) is interesting mostly as a
stepping stone, or a lower bound to "true" inequality between individuals. As mentioned
before, it may reflect the "feeling" of inequality within a country much better than the
unweighted Concept 1, particularly when regional inequalities coincide with other types
of horizontal cleavages. Changes in the population-weighted inequality are important
25
for two additional reasons. First, we may want to see whether inequality is driven by the
differing population growth rates between the regions. For example, increasing Concept 2
inequality (and very likely increasing inter-personal inequality) may be due to migration
of the population into faster-growing or richer regions.
Secondly, an interesting issue is whether there is correlation between growth rates
and population size of different regions. This is in turn relevant for two reasons.
25As Venables and Kanbur (2003, p. 11) write: "..it may be the case that the distribution of individual
attributes such as ability, is the same in all regions. Spatial [Concept 2] inequality is then particularly
inequitable and--especially when aligned with political, religious or ethnic tensions--may be dangerous to
social and political stability. If spatial divisions align with ethnic, religious or language splits, as so often
happens, then the between group component of inequality takes on a greater significance than its
contribution to interpersonal (Concept 3) inequality."
22
Assuming for the moment that there is no migration between the regions, the evidence of
a positive relationship between population and economic growth may help shed
additional light on the issue of increasing returns to scale: greater population may help
growth if it delays the onset of diminishing marginal returns. Second, higher growth in
more populous regions has obvious (positive) implications for the reduction in poverty. If
income growth rates tend to be higher in more populous regions then--everything else
being the same--poverty reduction will be greater.
We now turn to these two issues: (i) the influence of uneven regional population
growth on Concept 2 inequality, and (ii) the association between regional population size
and per capita GDP growth. But before we do so, we need to quickly look at the results of
Concept 2 inequality for the five countries. They are displayed in Figure 5. Brazil shows
the highest inequality with the Gini of about 30. As mentioned before, Concept 2
inequality sets the lower bound to inter-personal (Concept 3) inequality. This means
that--were all individual within each state of Brazil to have the same income--overall
inequality in Brazil would still be substantial. China has a Gini of about 25, Indonesia
and India about 20 (with the latter steadily catching up) and the US less than 10.26 The
ranking of the countries is basically unchanged throughout the period. In the 1990's,
inequality was on the rise in all three Asian countries. In the 1980's, however, China's
inequality--measured by the Gini--is constant while Theil shows a strong downward
trend. This indicates that the period must have been characterized by the catch-up of the
poorest (population-weighted) provinces.27 For the three Asian countries, there is a clear
contrast between the 1980's when regional inequality was declining or constant and of
the 1990's when it is also very clearly going up in all three countries.
26Since size income inequality in the United States is around 40 (see the 2004 release of World Income
Inequality Database available at http://www.wider.unu.edu/), this means that differences in states' mean
incomes explain only about a quarter of total US inequality.
27This is because Theil is more sensitive than the Gini to what happens at the tails of distribution. A more
detailed discussion of China's concept 2 inequality is available in Annex 2 of the Internet version of this
paper available at http://econpapers.hhs.se/RAS/pmi44.htm
23
Figure 5. Concept 2 (population-weighted inter-regional inequality) in the five countries
Gini Theil
0.400 0.200
Brazil
0.300 China Brazil
China
il
Gini 0.200 Indonesia 0.100
The
Indonesia
India
0.100 USA India
USA
0.000 0.000
1977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001 1977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001
Years Years
Note : Based on nominal state-level GDP per capita except for Indonesia where only real amounts are
available.
We next decompose the change in Gini between the first year for which we have
data (late 1970's or early 1980's depending on the country), 1990 and 2000. Table 5
shows what Concept 2 inequality would have been if the state/provincial distribution of
population had remained unchanged, and if provincial growth rates had been the same.
This enables us to distinguish two effects: uneven population growth, and uneven per
capita economic growth. The results show that the effect of differential population
movements has been fairly minor in all countries (see columns 4 and 9). Its greatest effect
was in the decade of the eighties in Indonesia where it was equalizing and the decade of
the nineties in China where, on the contrary, it contributed to greater inequality. One has
to be careful however in the interpretation of the population change. It is for example
true, in an accounting sense, that had the composition of Chinese population by province
remained the same as in 1990, regional inequality in 2000 would have been less.
However, this takes the per capita incomes by province as given. Yet in a deeper sense
this is wrong. This is because population might have moved in response to higher wages
in richer states bringing let us suppose--in the process the per capita income in the
24
richer states down. Without such an equilibrating movement of labor, inequalities might
have been even greater. Thus for a more meaningful analysis, we would need
information on natural vs. mechanical population changes by state/province.
We turn next to the contribution of differential growth rates to the change in
Concept 2 inequality (see columns 6 and 11). There the most remarkable turnaround is in
the case of China: while in the 1980's, differential growth by province reduced regional
inequality by almost 3 Gini points, in the next decade, differential growth added 2.6 Gini
points to inequality. In the first period, it is the poor and populous provinces that grew
relatively fast; in the second period, the poor and populous provinces tended to grow
slower than average. This is shown in Figure 6 where we note that the position of many
large and relatively poor provinces slipped from being in the NW quadrant (that is,
growing faster than average) in the decade of the 1980's, to the SW quadrant (growing
slower than average) in the decade of the 1990's.
In India, differential growth was disequalizing throughout although more so in the
decade of the nineties. In Indonesia too it was more disequalizing in the 1990's than in
the 1980's. It is only in Brazil, in the 1990's, that differential growth was substantially
equalizing.
The equalizing effect of differential growth rates can occur for two reasons
however. First, poorer regions may grow faster than the rich. Second, although there may
be no income convergence among the regions as such, there could be income
convergence among the populous regions. This would tend to reduce Concept 2
inequality. To explain: assume that populous regions are uniformly distributed across
income distribution and that there is zero correlation between growth rates and initial
incomes. Thus, there is neither convergence nor divergence. However, within the
subgroup of populous regions, let correlation between initial income and growth may be
25
negative. These regions' incomes will therefore become more clustered, and since it is
them who "matter" for Concept 2 inequality, inequality will go down. 28
28Suppose that there are two rich and small regions called A and B, and two poor and populous regions
called C and D. Let there be no income convergence, and let A and C grow at high positive rates, and B
and D decline. Now, Concept 2 inequality will increase since the two key regions which, because of their
sizes determine what happens to Concept 2 inequality, will have become dissimilar. Note that this takes
place while there is neither income divergence not a relationship between population size and growth--but
simply income divergence between populous regions.
26
Table 5. Concept 2 Gini coefficients with actual and hypothetical populations and income
Change between (circa) 1980 and 1990 Change between 1990 and 2000
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Actual Actual 1990 Population 1990 Differential Actual 2000 Population 2000 Differential
(circa)1980 1990 Concept 2 effect Concept 2 growth 2000 Concept 2 effect Concept 2 growth
Concept 2 Concept 2 inequality (2)-(3) inequality effect Concept 2 inequality (7)-(8) inequality effect
inequality inequality if if GDP per (2)-(5) inequality if if GDP per (7)-(10)
population capita population capita
compositio growth compositio growth
n did not rates were n did not rates were
change the same change the same
USA 6.9 8.9 8.9 0 7.1 +1.8 8.3 8.4 +0.1 8.9 -0.6
China 23.9 20.9 20.9 0 23.8 -2.9 24.4 24.0 +0.4 21.8 +2.6
India 14.8 16.1 16.1 0 14.8 +1.3 18.7 18.7 0 16.2 +2.5
Brazil 31.1 30.9 30.8 +0.1 31.3 -0.4 28.0 28.0 0 31.0 -3.0
Indonesia 19.0 18.8 19.6 -0.8 18.5 +0.3 19.9 20.0 -0.1 18.8 +1.1
Note: The "circa 1980" means the first year for which the data are available: it is 1977 for the United States, 1978 for China, 1980 for India, 1985 for Brazil and
1983 for Indonesia. All Concept2 Ginis calculated from nominal data except for Indonesia. Indonesian data exclude oil and gas component of GDP.
27
Figure 6. China: Provincial per capita growth, initial income level and population size
between 1978 and 1990 (left panel) and 1990 and 2000 (right panel)
90 00
1978- 1990-
at 11 at 11
apic apic
per per
PD
G 9 PD 9
G
of of
e e
atr atr
ht ht
ow 7 ow 7
gr gr
age age
er er
av av
5 5
300 500 700 900 1100 1300 800 1300 1800 2300 2800
GDPper capitain1978 GDPper capita in 1990
Note: Vertical axis shows average growth rate (in percent per capita) over each period. The size of each dot
reflects population weight of the province. The city provinces of Beijing, Tianjin and Shanghai are not
shown. Horizontal lines drawn at the level of nation-wide average growth during each period (7.3 percent
per capita in the first, 9.2 percent per capita in the second). Vertical line drawn at respectively 1978 and
1990 nation-wide GDP per capita (expressed in 1995 international dollars).
To check the second possibility we regress regional rates of growth over a period
of five-years on initial regional income levels as in (5)
ln yit - ln yi , t - 5 = 0 + 1 yi , t - 5+ eit (5)
where i is i-th region, t is time, y = real GDP per capita, and the sign of the
coefficient 1 indicates presence or absence of unconditional convergence. The regression
is population-weighted since our interest is not in convergence as such but in
convergence conditional on population size. The regression is calculated for each country
and for all years. This is done in order to avoid biasing the results through choice of the
time period and stage of the economic cycle.29 The clearest results (see Figure 7) are for
India: they show a statistically significant regional divergence since the early 1990's. For
China, it is only in the five-year period ending in 1999, that we come close to having a
29Petrakos et al. (2002, p. 4) for example show how, using the same data, the choice of different time
intervals may yield either convergence or divergence.
28
statistically significant regional divergence. Similarly, Indonesia has a five-year period of
divergence ending in 1997. In contrast to the three Asian countries that all show some
evidence of population-weighted divergence in the 1990's are Brazil and the United
States. Brazil's states were strongly converging until the mid-1990's. The United States
had a period of convergence from about 1990-91 until the late 1990's.
Consider now what is behind these results. As mentioned before, Concept 2
convergence or divergence is in reality a convergence or divergence within the subset of
populous regions. The difference in outcomes among the most populous states is very
clear in the case of India. Consider the three most populous Indian states: Uttar Pradesh
(population 170 million in 1999), Bihar (107 million) and Maharashtra (95 million).
Between 1990 and 1999, GDP per capita in Maharashtra that was the richest of the three
states increased by 60 percent (in real terms). Meanwhile, in Bihar--which is the poorest
state in India--GDP per capita remained the same. Finally, Uttar Pradesh--which has the
median level of income among the states--saw its GDP per capita increase by about 15
percent. The situation is similar in China. The three largest provinces are Sichuan
(population 114 million in 2000), Henan (92 million) and Shandong (91 million). During
the decade of the nineties, Shandong which was the richest of the three in 1990 saw its
real GDP per capita triple. Henan's GDP per capita increased by 2½ times, and Sichuan's
(which was the poorest of the three provinces) "only" doubled. In conclusion, in both
India and China, the most populous regions diverged in the 1990's and that drove
Concept 2 inequality--and possibly inter-personal inequality as well--up.
29
6. How to explain changes in inter-regional inequality?
As we have seen, levels of inter-regional inequality are quite different in the five
countries. Indonesia and Brazil have the highest level of Concept 2 inequality. They are
followed by China. India is in the middle but is the only country showing a consistent
increase in inter-regional inequality while the United States displays the most uniform
levels of GDP per capita between its component states. We shall try to explain within-
national convergence or divergence in more general terms, that is by looking at the
variables of interest (at national level) that might contribute to or reduce inter-regional
inequalities. These variables are not easy to identify in general because of countries'
specificities. Thus, for example, the same overall growth rate might produce regional
divergence in one case, and regional convergence in another depending on what drives
growth. When growth is narrow-based as in case of oil production, the differences
between regions (oil-rich and the rest) are likely to increase. When growth is broad-
based, or is fuelled by agricultural growth (as in the case of China during the early
liberalization), it can be expected to help convergence as poorer and agricultural regions
catch up with rich regions. The general presumption--as we have seen from
Williamson's (1965) hypothesis--is that growth will tend to be regionally disequalizing
(even if ultimately higher income may be associated with lower inequalities). We may
thus expect growth rate to be positively related to regional inequality. This is also a
finding obtained by Petrakos et al. (2003) in their study of regional inequality in the
European Union. The same results are further reported in the case of Indonesia by Akita
and Kawamura (2002, p.12) who find that the period of fast growth between 1993 and
1996 was associated with a slightly increasing Concept 2 inequality while the crisis led to
a decline in regional inequality (p.16). In a long-run 1939-95 study of Brazil's regional
inequality Azzoni (2001, p. 144) finds the same relationship.
30
Figure 7. Unconditional regional population-weighted convergence or divergence
( coefficient with the 95 confidence interval)
United States China India
0.10
0.04 0.12
0.02 0.08
0.00
1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
ue ue
value
a val0.00 val0.04
Bet etaB 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 etaB
-0.10
-0.02 0.00
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.20 -0.04 -0.04
Years Years Years
Indonesia Brazil
0.04
0.04
0.02
0.02
ues
val0.00
a
etB 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 value 0.00
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
Beta
-0.02 -0.02
-0.04 -0.04
Years Years
Note: the regression is ROG(t-5,t)=0+1 log GDP (t-5) (population weighted) where ROG=annualized per
capita growth rate between t-5 and t, and GDP = real GDP per capita in year t-5 (all expressed in constant
international 1995 dollars). Indonesian data exclude oil and gas portion of GDP.
31
Similarly, the effect of greater openness (approximated by the trade to GDP ratio
in nominal terms) may be ambiguous. Openness can help the already rich regions, or can
create income gaps where none existed. But it can also help poorer regions whose output
(e.g. agricultural goods) was artificially held down through price controls. On the other
hand, one can argue that there may be certain policies associated with globalization
whose effects are less likely to be ambivalent. They include policies of financial
liberalization and higher interest rate that tend to favor rich households. To the extent that
rich people are concentrated in certain areas, such policies will increase spatial inequality.
As for the empirical results, they span the entire gamut. Kanbur and Zhang (2003) find
that openness was associated with rising inequality in China. Zhang and Zhang (2003)
similarly decompose Concept 1 inequality between China's states and find that about 20
percent of differences in provincial GDPs per capita can be ascribed to differences in
trade shares (p. 57).30 Petrakos et al. (2002, p. 19) however find that openness (defined as
regional integration within the European Union31) did not have a uniform impact on all
EU countries: in some it was associated with greater regional inequality, in others by
smaller, and in some had no statistically significant effect at all. Wei and Wu (2001)
using urban/rural ratios of mean income for more than 100 cities and their adjoining
areas in China find that in the period 1988-93 increased openness tended to reduce the
urban/rural ratio. Since urban/rural differences are perhaps the principal explanation for
regional inequality in China, increased openness would seem to reduce regional
inequality.
Visual inspection of Figure 8 which displays the relationship between openness
and Concept 1 inequality shows that all of our five countries (with the exception of
Indonesia), have, as we would expect, relatively low openness. The graph also shows that
in China, increased openness was not associated with increased regional inequality while,
on the contrary, in India, this seems to be clearly the case. At the other extreme is
30Their approach is interesting because they run production functions of the same form for all provinces.
One of the arguments in the production function is trade ratio (in addition to education, domestic and
foreign capital accumulation etc.) Then log variance of GDPs per capita is decomposed and the regression
coefficients times the covariance between income and each argument gives an estimate of that particular
argument's contribution to inter-provincial income inequality.
31And measured as share of European trade in country's total trade.
32
Indonesia where increased openness was associated with a decrease in regional
inequality. Finally in Brazil and the United States, there was no apparent relationship
between openness and regional inequality. The conclusion is that the country experiences
differ and that openness as such may not have the same discernable effects on countries
regardless of their level of development, type of economic institutions, and other macro
economic policies.
A somewhat different picture however emerges when we look at the relationship
between Concept 2 inequality and openness (Figure 9). It is now very clear that in both
China and India, increased openness was associated with greater regional inequality. In
Indonesia, the relationship is the opposite and in Brazil and the US the correlations seems
non-existent.
We then run two types of regressions across both Concept 1 and Concept 2
inequalities (see equation 6). The first type are fixed-effect panel regressions, the second
are Arellano-Bover GMM (Generalized Method of Moments) regressions.
Iit = 0 + 1ROGit + 2lnGDPit + 3OPENit + 4RINTit + 5INFLit + uit (6)
where I = inequality concept that can be either Gini or Theil (with e.g. Gini2
denoting Gini coefficient under Concept 2 inequality), i denotes country, t = year, ROG
= real annual rate of growth of per capita GDP, GDP = level of GDP per capita, OPEN
= trade to GDP ratio in current prices, RINT = real interest rate (on deposits) in percent
per annum, INFL = the average annual rate of inflation proxied by the change in
consumer price index. GDP and rate of growth of GDP are included, as discussed in
Section 2, to test for the presence of short- and long-term effects of growth on regional
inequality (as implied by, among others, by Williamson's 1965 hypothesis).32 The
possible role of openness is explained above. We also include real rate of interest and
inflation as possible controls on the assumption that both can contribute to regional
32This is also the reason why a squared GDP per capita term (in the spirit of the Kuznets hypothesis) is not
included. The latter is a test of inter-personal inequality in the presence of structural change (and increase
in income). Here we test regional inequality which, according to the Williamson and Perroux hypotheses,
should be less in the already developed economies where market integration is more advanced. Per capita
income thus serves as a proxy for the extent of integration of capital, labor and goods markets.
33
inequality: the first because a higher rate of interest will increase incomes of property-
owners who also tend to be spatially concentrated; the second, because higher inflation is
often associated with greater variability in relative prices. To the extent that different
regions specialize in producing different output-mixes, then relative price variability will
affect them differently, thus adding to inequality.
All variables in (6) are defined at national level. Inequality that is calculated
across provinces is Concept 1 or Concept 2 inequality at the national level (in the same
way that inequality across individuals in a country is a national level variable). The other
variables are obviously so. It is an unbalanced panel since the number of observations for
each country is not the same (e.g. for China we have the data for the 1978-2000 period,
but for Indonesia only for 1983 to 2001). Since there are unobservable country-specific
effects and the number of regional units per country varies (which in turn means that,
everything else being the same, countries with a greater number of regional units would
tend to have higher inequality), we use the fixed effects model. However if we
acknowledge that inequality is a process characterized with a high degree of persistence,
we have to move to dynamic panels (with the lagged inequality value on the RHS). The
introduction of the lagged dependent variable makes the fixed-effect estimator
inconsistent in small samples. To account for that we use Arellano-Bover GMM system
estimator which similarly to the Arellano-Bond estimator involves first-differencing to
remove possible country fixed effects, instrumenting of predetermined and endogenous
variables by their lagged values, 33but also extracts as much information as possible from
the data by estimating the regressions both in levels and changes and imposing the
equality of the slopes between the two.
33We also use as instruments population size and land area. The Sargan test of identifying restrictions is
statistically insignificant throughout implying acceptance of exogeneity of instruments.
34
Figure 8. The relationship between openness and Concept 1 inequality
BRA CHN IDN
.4
.3
.2
s .1
ealrni1iniG .2 .4 .6 .8 1
IND USA
.4
.3
.2
.1
.2 .4 .6 .8 1 .2 .4 .6 .8 1
trade/GDPall in current dollars
Graphs by contcod
Figure 9. The relationship between openness and Concept 2 inequality
.
BRA CHN IDN
.4
.3
.2
.1
r_2in .2 .4 .6 .8 1
Gi IND USA
.4
.3
.2
.1
.2 .4 .6 .8 1 .2 .4 .6 .8 1
trade/GDPall in current dollars
Graphs by contcod
Note: Both inequality concepts are expressed in real terms. The data for
Indonesia exclude gas and oil. CHN=China; IND=India, IDN=Indonesia;
BRA=Brazil.
35
Table 6. Determinants of regional inequality
Concept 1 inequality Concept 2 inequality
Gini1 Theil1 Gini2 Theil2
Fixed effects Arellano- Fixed effects Arellano- Fixed effects Arellano- Fixed effects Arellano-
Bond Bond Bond Bond
Lagged Gini 0.977 0.973 0.973 0.962
or Theil (0.000) (0.000) (0.000) (0.000)
ROG -0.020 0.031 0.0014 0.036 0.028 0.046 0.038 0.004
(0.69) (0.19) (0.98) (0.15) (0.61) (0.006) (0.32) (0.002)
GDP per 0.009 -0.0014 -0.003 -0.001 0.056 -0.002 0.03 -0.001
capita (0.37) (0.15) (0.77) (0.22) (0.000) (0.011) (0.000) (0.006)
Openness 0.039 0.004 0.034 0.003 0.005 0.000 0.021 0.003
(0.13) (0.50) (0.15) (0.57) (0.87) (0.99) (0.27) (0.39)
Inflation 0.0007 0.0002 0.0008 -0.0001 0.004 0.0004 0.003 0.0001
(0.60) (0.75) (0.52) (0.80) (0.016) (0.44) (0.01) (0.94)
Real interest 0.009 0.014 0.01 0.22 -0.028 -0.01 -0.02 0.004
(0.85) (0.58) (0.81) (0.39) (0.57) (0.56) (0.57) (0.78)
Constant 0.126 0.014 0.116 0.01 -0.297 0.002 -0.19 0.013
(0.14) (0.15) (0.14) (0.24) (0.002) (0.011) (0.003) (0.008)
R2 within 0.14 0.07 0.41 0.39
F 2.5 2010 1.1 1534 10.0 3799 9.4 2540
(0.04) (0.000) (0.38) (0.00)) (0.000) (0.000) (0.00) (0.000)
No. of 84 78 84 78 84 78 84 78
observations
Sargan test 74.8 75.4 73.7 78.6
(0.95) (0.94) (0.95) (0.86)
Note: p values between brackets. Coefficients significant at 5 percent or less are shaded. Real interest = real annual rate of interest in percent p.a.
divided by 1000. Inflation = ln (1+annual inflation rate in percent). Openness = trade/GDP (in nominal US dollars). GDP per capita = ln GDP per capita in 1995
international dollars. ROG = annual rate of growth of GDP per capita (expressed in fractions, e.g. 3 percent is expressed as 0.03). Gini and Theil are expressed
as ratios and calculated across real regional incomes. Arellano-Bond is one-step GMM estimator; calculations done using David Roodman's xtabond2.ado
STATA software.
36
The results of estimation are displayed in Table 6. Regarding Concept 1
inequality, we see that practically no variable in either formulation seems to have any
influence and the quality of the overall results is very weak. Contrary to several other
studies, we do not find evidence that higher rate of growth has a disequalizing effect on
regional incomes. The situation is different when we move to the determinants of
Concept 2 regional inequality. There, according to both Gini and Theil, and when using
dynamic panel specification, higher rate of growth is disequalizing, and higher income
level is equalizing. We thus find the same results as obtained by Petrakos, Rodriguez-
Pose and Roviolis (2003), in the context of the European Union regions. The effect of 34
higher growth rate on Concept 2 inequality is rather small though: an acceleration of
national growth rate by 1 percentage point is associated on average with an increase of
Concept 2 inequality by 0.046 Gini points. Inflation is disequalizing only in the static
(fixed-effects) framework. Openness, however, although it enters all the regressions with
a positive sign, is not statistically significant in none of them. The same is true for real
rate of interest which shows a sign change as well.
7. Conclusions
The three Asian countries show much greater evidence of increasing regional
income inequality than Brazil and the United States. Divergence in China, India and
Indonesia is evident in the decade of the 1990's. The clearest case is that of India. In the
1990's, it shows a consistently increasing regional inequality of both Concept 1 and 2
types--whether it is measured by the Gini or by the Theil index. The largest states
(Maharashtra, Uttar Pradesh and Bihar) are diverging in their incomes, and generally
richer states have registered faster growth.
The experience of the 1990's is in contrast to that of the previous decade for all
three countries. In China during the 1980's, there was a convergence between the regions.
It coincided with the period of agricultural reforms which enabled poorer regions to catch
34In fixed-effects though, income level is positively associated with Concept 2 inequality and rate of
growth is not significant.
37
up. Moreover, over the same period, there was also a strong Concept 2 convergence. The
process comes to a halt in the 1990's with provincial mean incomes growing at relatively
similar rates (Concept 1 inequality stable), but, importantly, Concept 2 inequality
increasing. Similarly to what we noted for India, in China the divergence in the second
half of the 1990's was driven by the divergence in outcomes between populous
provinces. As there is a contrast between Maharashtra and Bihar in India, there is a
contrast between Shandong and Sichuan in China.
As for the two non-Asian countries studied here, the United States, despite
already being most regionally homogeneous, has generally displayed tendency toward
income convergence according to both inequality concepts. Brazil shows least consistent
change. There is only a slight change in both Concept 1 and Concept 2 inequality, but it
is overshadowed by a cyclical effect of inflation. It is therefore difficult to see whether
there is any trend. Nevertheless Brazil has the highest Concept 2 Gini inequality of all
countries included here. To see how large that inequality is, one need simply realize that
were all incomes within each state absolutely equal, Brazil's overall Gini would still be
around 30, a moderate nation-wide inequality level.
When we try to fit our findings into an admittedly Procrustean bed of regression
analysis, we do not find much evidence, even after adjusting for countries' unobservable
characteristics, that the most obvious variables used in the work on regional inequality
explain its Concept 1 change over time. Turning to population-weighted regional
inequality, we find that it tends to be positively associated with faster growth and
negatively with income level. As for openness, although its coefficient is positive, it is
not statistically significant in any of the formulations. High inflation is (in static panel
framework) associated with higher Concept 2 inequality.
Global inequality changes and inequality changes in the five most populous
countries in the world during the last two decades have indeed been quite dramatic but
also extremely complex. Note the following facts. While at the level of these countries,
we find no evidence for the narrowing of (unweighted) inter-regional income differences
38
during the decade of the 1990's, Concept 2 inequality however tended to rise in the two
most important countries (India and China) as well as Indonesia. But at the global level, a
process directly opposite to the one at the national level was simultaneously taking place.
It consisted of widening income differences between the countries (as poor countries
have tended to lag in growth rates behind the rich) and a reduction in Concept 2
inequality thanks mostly to China's phenomenal growth. In other words, while within
India or China during the last ten years or so, Concept 1 inequality was stable and
Concept 2 inequality was increasing, on the global level (mostly thanks to what was
happening precisely in these two countries), Concept 1 inequality was up and Concept 2
inequality went down.
Even without addressing the most difficult task of measurement of inter-personal
inequality at the global level, one can easily see how contradictory are recent changes in
different facets of inequality and how inadequate are simple answers regarding both the
direction of change and even more so regarding the causality. In other words, when we
ask a question as apparently simple as "what was a change in inequality in country X",
we need to define very carefully what we mean by inequality since its different versions
may easily move in the opposite directions.
39
REFERENCES
Akita, Takahito and Lukman, Rizal Affandi (1995), "Interregional inequalities in Indonesia: A
sectoral decomposition analysis for 1975-92", Bulletin of Indonesian Economic Studies, vol. 31, No. 2,
August, pp. 61-81.
Akita, Takahito and Kazumi Kawamura (2002), "Regional income inequality in China and
Indonesia: A comparative analysis", Paper presented at the 46th Congress of European Regional Science
Association, Dortmund, Germany, August 27-31, 2002.
Available at http://www.ersa.org/ and http://econpapers.hhs.se/paper/wiwwiwrsa/ersa02p432.htm.
Akita, Takahito and Mitsuhiko Kataoka (2003), "Regional income inequality in the post-war
Japan", paper presented at the 47th Congress of European Regional Science Association, Jyvaskyla,
Finland, august 27-30, 2003. Available at http://econpapers.hhs.se/paper/wiwwiwrsa/ersa03p480.htm.
Azzoni, Carlos R. (2001), "Economic growth and regional income inequality in Brazil" , The
Annals of Regional Science, vol. 35, pp. 133-152.
Barro, Robert and Xavier Salla-i-Martin (1992), "Convergence", Journal of Political Economy, vol.
100 (2), pp. 223-251.
Bliss, Christopher (1999), "Galton's Fallacy and Economic Convergence", Oxford Economic
Papers, vol. 51, pp. 4-14.
Bhalla, Ajit S., Shujie Yao and Zongyi Zhang (2003), "Causes on inequalities in China", Journal
of International Development, vol. 15, pp. 939-955.
Chen, Shaohua and Yan Weng (2001), "China's growth and poverty reduction: Recent trends
between 1990 and 1999", World Bank Research Working Paper No. 2651, July.
Dabus, Carlos (2000), "Inflationary Regimes and Relative Price Variability: Evidence from
Argentina" Journal of Development Economics vol. 62, No. 2 (August 2000): pp. 535-47
Dreze, Jean and P.V. Srinivasan (2000), "Poverty and inequality in India: Evidence from regional
data", Journal of Quantitative Economics, vol. 16,. No. 1, January, pp. 33-50.
Easterly, William and Ross Levine (2001), "It's Not Factor Accumulation:Stylized Facts and
Growth Models", World Bank Economic Review, vol. 15(2), pp. 177-220.
Goerlich, Francisco J. and Matilde Mas (2001), "Inequality in Spain, 1973-91: Contribution to a
Regional Database", Review of Income and Wealth, Series 47, No.3, September.
Heston, Alan (2001), "Treatment of China in PWT 6.1", December. Available
http://pwt.econ.upenn.edu/.
Hirschman, Albert (1958), The strategy of economic development, New Haven: Yale University
Press.
Holz, Carsten A. (2004), "China's Statistical System in Transitin: Challenges, Data Problems and
Institutional Innovations", Review of Income and Wealth, Number 3, September, pp. 381-409.
Jha, Raghbendra (2004), "Reducing poverty and inequality in India: has liberalization helped", in
Giovanni Andrea Cornia (ed.) Inequality, growth and poverty in an era of liberalization, Oxford, U.K. and
New York: Oxford University Press. Also available at
http://rspas.anu.edu.au/economics/publish/papers/wp2002/wp-econ-2002-04.pdf.
Kanbur, Ravi and Xiaobo Zhang (2003) "Fifty years of regional inequality in China: A journey
through central planning, reform and openness". Available at
http://people.cornell.edu/pages/sk145/papers/Halfcentury81.pdf. forthcoming in Review of Development
Economics, vol. 9, No. 1, 2005.
Krugman, Paul (1991), Geography and Trade, London: MIT Press.
Krugman, Paul and anthony Venables (1995), "Globalization and the inequality of nations",
Quarterly Journal of Economics, vol. 110, No. 4, pp. 857-880.
Maddison, Angus (1998), Chinese Economic Performance in the Long Run, Paris: OECD
Development Centre.
Maddison, Angus (2003), The World Economy: Historical Statistics, Paris: OECD Development
Centre Studies.
Maddison, Angus (2003), The World Economy: Historical Statistics, Paris: OECD Development
Centre Studies.
Milanovic, Branko (2002), "True world income distribution, 1988 and 1993: First calculation
based on household surveys alone", The Economic Journal, vol. 112, No. 476, January 2002, pp. 51-92.
40
Milanovic, Branko (2005), Worlds Apart: Global and International Inequality 1950-2000,
Princeton, N.J.: Princeton University Press.
Myrdal, Gunnar (1957), Economic theory and underdeveloped regions. London: Hutchinson.
Perroux, Francois (1970), "A note on the concept of growth pole" in D. McKee, R. Dean and W.
Leahy (eds.), Regional economics theory and practice, New York: Free Press, pp. 93-103.
Perroux, Francois (1988), "The Pole of Development's New Place in a General Theory of
Economic Activity", Regional economic development: Essays in honour of Francois Perroux, Boston;
London; Sydney and Wellington, New Zealand: Unwin Hyman, pp. 48-76
Petrakos, George, A. Rodriguez-Pose and A. Rovolis (2003), "Growth, integration and regional
inequality in Europe", paper presented at the 47th Congress of European Regional Science Association,
Jyvaskyla, Finland, august 27-30, 2003.
Available at http://www.ersa.org/ersaconfs/ersa03/cdrom/papers/46.pdf.
Parks, Richard (1978), "Inflation and relative price variability", Journal of Political Economy, vol.
86, pp. 79-95.
Ram, Rati (1992), "Interstate income inequality in the United States: measurement, modelling and
some characteristics", Review of Income and Wealth, vol. 38, pp. 39-49.
Rawski, Thomas (2001). "What's Happening to China's GDPs Statistics", China Economic
Review, No. 4, vol. 12., pp. 347-354.
Renard, Marie-Francoise (2002), "A Pessimistic View on the Impact of Regional Inequalities",
China Economic Review vol. 13, No.4, pp. 341-44
Rosenstein-Rodan, Paul (1943), "Problems of industrialisation of Eastern and Southeastern
Europe", Economic Journal, vol. 53, pp. 202-211.
Sharping, Thomas (2001), "Hide-and-seek: China's elusive population data", China Economic
Review, No. 4, vol. 12., pp. 323-332.
Shorrocks, Anthony and Guaghua Wan (2004), "Spatial decomposition of inequality", WIDER
Discussion Paper No. 2004/01.Available at http://www.wider.unu.edu/publications/publications.htm.
Venables, Anthony J. and Ravi Kanbur (2003), "Spatial Inequality and Develoment", Department
of Applied Economics and management working paper 2003-38, Cornell University, November.
(Republished as introduction to Venables and Kanbur (eds), Spatial Development and Inequality, Oxford:
Oxford University Press, 2005).
Wei, Shang-jin and Yi Wu (2001). "Globalization and Inequality: Evidence from Within China."
Center for Economic Policy Research Discussion Paper, December, No. 3088. Williamson, Jeffrey (1965),
"Regional inequality and the process of national development: a description of patterns", Economic
development and cultural change, vol. 13. No,. 4, pp. 3-45.
Xu, Xianchun (1999), "Evaluation and Adjustments of China's Official GDP by the World Bank
and Prof. Maddison", Journal of Econometric Study of Northeast Asia, vol. 1, No.2.
Yao, Shujie and Zongyi Zhang (2001), "On regional inequality and diverging clubs: A case study
of contemporary China", Journal of Comparative Economics, vol. 29, pp. 466-484.
Yemtsov, Ruslan (2002), "Quo Vadis: Inequality and Poverty Dynamics Across Russian
Regions", World Bank, mimeo.
Zhang, Zongyi, Aying Liu and Shujie Yao (2001), "Convergence of China's regional incomes
1952-1997", China Economic Review, vol. 12, pp. 243-258.
Zhang, Xiaobo and Kevin H. Zhang (2003), "How does globalisation affect regional inequality
within a developing country? Evidence from China", Journal of Development Studies, vol. 30, No.4, April,
pp. 47-67.
41
Annex 1. Administrative organization of the five countries and data coverage
China
China is officially administratively divided into 34 regional units: 23 provinces
(including Taiwan, China), 2 special administrative regions (Hong Kong and Macau), 5
autonomous regions (Guangxi, Inner Mongolia, Ningxia, Xinjiang and Tibet) and 4
municipalities (Beijing, Shanghai, Tianjin and Chongqing). The autonomous regions
differ from the rest because of the presence of significant non-Han minorities in them.
We are not including in the analysis Taiwan (China), Hong Kong and Macau
because the GDP data for them are not shown together with other regions. In addition,
Chongqing acquired the status of a separate unit (municipality) only in 1997 and in our
data is included together with Sichuan. This therefore gives us 30 regional units which
for simplicity we call "provinces." Out of these thirty, full data sets (both nominal and
real GDPs) are available for 27 provinces which represent 99 percent of Chinese
population (not counting Taiwan, China; Hong Kong; Macau). For two provinces
(Qinghai and Hainan), nominal GDP is available from 1985 onwards. 35Hence in
calculations with nominal values, 29 provinces are included. Finally, for Tibet neither
real nor nominal GDP data are available.
A note is in order regarding Chinese GDP per capita data.36 There are three types
of problems with these data. First, all-China GDP data are the subject of a long academic
dispute. Many authors argue that Chinese official statistics exaggerate both the level and
rate of growth of the economy (Maddison, 2003, p. 151; Maddison, 1998; Heston, 2001;
Xu 1999; Rawski, 2001). If this is true then provincial growth rates must also be
exaggerated. While we tend to believe that these authors do have a point and that the total
GDP correction does make sense, the only source of provincial Chinese statistics is the
official State Statistics Bureau. We thus use official regional and nation-wide statistics
35Hainan became a separate province in 1988. Previously it was part of Guangdong province.
42
for the entire 1978-2001 period. Second, until 1994, the sum of provincial GDPs was
approximately equal to the official value of the nation-wide GDP. The discrepancy was
within a very narrow range of 1-2 percentage points possibly due to the mistakes of
classification. But after 1994, the sum of provincial GDPs is systematically greater (up to
15 percent) than the nation-wide GDP (see Holz, 2004, pp. 388-9; Milanovic, 2005, pp.
97-8). It is unclear what the source of this large discrepancy is. Heston (2001, p. 3)
ascribes it to the "winds of falsification". Holz (2004) similarly blames data falsification
at lower-level administrative units. Third, the denominator in GDP per capita data is the
officially registered population (see Scharping, 2001). This tends to underestimate actual
population living in richer provinces which, due to their dynamism, attract the bulk of
"floating population" (people who do not have official permits to live in a given city).
GDP per capita for richer provinces is accordingly overestimated and GDP per capita for
poorer provinces underestimated, thus biasing our inequality statistics upward. We have
no way out of this conundrum. This is because the only existing data on provincial GDP
and provincial population are the official data.
India
India was until the year 2000 administratively divided into 25 states and 7 Union
territories. The Union territories are very small with the exception of the federal capital of
Delhi. In total they account for about 11 million people or a little over 1 percent of total
population (in the 1991 census). India is officially defined as a "union of states" (Article
2 of the 1949 Constitution). Out of 25 states, we have GDP per capita data for 14 states.
These are the largest states comprising 92 percent of total Indian population in 2000. Of
the states for which we do not have the data, the largest are Assam, and Jammu and
Kashmir (respectively 23 and 8 million according to the 1991 census). The disputed
parts of Kashmir are also omitted from the analysis (the data for them are included in
Chinese and Pakistani statistics).
36An entire issue of China Economic Review (vol. 12, No. 4, 2001) was dedicated to the data problems in
Chinese statistics.
43
In 2000, three new states were created. They are Uttaranchal or Uttarkhand
(population of 4 million), carved out of Uttar Pradesh (170 million before the split),
Jharkhand (population of 15 million) carved out of Bihar (107 million before the split),
and Chhattisgargh (population of 19 million) carved out of Madhya Pradesh (79 million
before the split). All three are based on ethnic distinctiveness from their "mother" states.
In our data for 2000, these states are included as part of their "mother" states.
United States
United States is administratively divided in 50 states and 1 federal district (the
federal capital of Washington). In the Constitution, the United States is not officially
defined as a federation although this is implicit from the text. United States also controls
some 12 Commonwealths and territories out of which Puerto Rico is the largest (3.7
million people). Our data cover the 50 states and the District of Columbia. However, the
latter, due to its peculiarities,37 is not included in the analysis which thus bears only on
the 50 federal units. Since District's population is only about 600,000 people, our
coverage of the United States is almost complete.
Brazil
Brazil is administratively divided in 26 states and 1 federal district (the federal
capital of Brasilia). It is officially a federation (hence the official name since 1988 of
"Federative Republic of Brazil"). Our data include all states with the exception of
Tocantins for the years 1985-88. The federal district of Brasilia (population 2 million) is
not included. Thus the population coverage is about 99 percent. 38
Indonesia
37For example, District's GDP per capita is unrealistically high, at more than $PPP 90,000. This is due to
the fact that many businesses are registered in the District while people who work or own them live just
outside the borders of the District.
38The state level value added data for Brazil (available at Instituto Braseleiro de Geografia e Estatistrica,
http://www.ibge.gov.br/english/estatistica/economia/contasregionais/default.shtm) account for between 95
and 98 percent of all-Brazil sum from the official statistics. Almost all of that difference is explained by the
fact that state data used here exclude the federal territory of Brasilia (its share of all-Brazil GDP varies
between 3 and 5 percent).
44
Indonesia is administratively divided in 26 provinces. Before East Timor's
independence in 1999, there were 27 provinces. The data for East Timor are available for
the entire period during which it was an Indonesian province, but for the reasons of
comparability, we have omitted it from our calculations throughout the entire period. The
population coverage of Indonesia is thus 100 percent.
45
Included countries and states/provinces
China India Indonesia Brazil USA
(Prices: 1978) (Prices: 1980-81) (Prices: 1983) (Prices: 1985) (Prices: 1995)
Beijing Andhra Pradesh Aceh Acre Alabama
Tianjin Bihar North Sumatra Alagaos Alaska
Heibei Gujarat West Sumatra Amapa Arizona
Shanxi Haryana Riau Amazonas Arkansas
Neimeng Karnataka Jambi Bahia California
Liaoning Kerala South Sumatra Ceara Colorado
Jilin Madhya Pradesh Bengkulu Distrito Federal Connecticut
Heilongjiang Maharashtra Lampung Espirito Santo Delaware
Shanghai Orissa Jakarta Goias Florida
Jiangsu Punjab West Java Maranhao Georgia
Zhejiang Rajasthan Central Java Mato Grasso Hawaii
Anhui Tamil Nadu Yogyakarta Mato Grosso do Sul Idaho
Fujian Uttar Pradesh East Java Minas Gerais Illinois
Jiangxi West Bengal Bali Para Indiana
Shandong West Kalimantan Paraiba Iowa
Henan Central Kalimantan Parana Kansas
Hubei South Kalimantan Pernambuco Kentucky
Hunan East Kalimantan Piaui Louisiana
Guangdong North Sulawesi Rio de Janeiro Maine
Guangxi Central Sulawesi Rio Grande do Norte Maryland
Hainan South Sulawesi Rio Grande do Sul Massachusetts
Sichuan Southeast Sulawesi Rondonia Michigan
Guizhou West Nusa Tenggara Roraima Minnesota
Yunnan East Nusa Tenggara Santa Catarina Mississippi
Shannxi Maluku Sao Paulo Missouri
Gansu Irian Jaya Sergipe Montana
Qinghai Timor Timur Tocantins Nebraska
Ningxia Nevada
Xinjang New Hampshire
New Jersey
New Mexico
New York
North Carolina
North Dakota
Ohio
Oklahoma
Oregon
Pennsylvania
Rhode Island
South Carolina
South Dakota
Tennessee
Texas
Utah
Vermont
Virginia
Washington
West Virginia
Wisconsin
Wyoming
46
Annex 2. Regional inequality in China
China at the end of the 20th century: different regional inequality break-downs
(Concept 2 inequality throughout)
Akira and Kawamura Kanbur and Zhang Modified (Milanovic) Bhalla, Yao and Zhang Milanovic (here)
(2002) (2003, Table 2) Kanbur and Zhang (2003)
(2003)
Year 1998 2000 2000 1995 2000
Welfare concept GDP per capita Mean real per capita Mean real per capita Mean real per capita GDP per capita
consumption 1/ consumption 1/ consumption 2/
Smallest unit with data District level GDP per Rural/urban Rural/urban Rural/urban Provincial GDP per
capita (=peasant/non-peasant) capita
Aggregation By province (26) and by By province (28) By province (28) By province (28) ---
regions (3)
Total number of observations (per 335 56 56 56 29
year)
Concept 2 Theil 3/ 24.9 24.8 24.8 16.54 (see note) 10.4
(which Theil?) Theil (1) Theil (0) Theil(0) Theil (1) Theil (1)
Between 3 large regions 6.6 (27)
Between provinces (within each 2.8 (11)
region)
Between 28 provinces alone 9.4 (38) 10.9 (44) 11.0 (67) 10.4 (100)
Within provinces (between districts) 15.6 (62)
Within provinces (between rural and 13.9 (56)
urban)
Between overall rural and overall 13.9 (56) 10.4 (75)
urban mean
Within rural areas and within urban 10.9 (44) 6.1 (25)
areas
Within rural areas 3.8 (21)
Within urban areas 2.3 (4)
Concept 2 Gini 3/ n.a. 37.2 37.2 n.a. 24.4
47
Sources: Takahiro and Kawamura (2002, Table 1, p. 26). Kanbur and Zhang (2003, Table 2, p. 27). Bhalla, Yao and Zhang (2003, Table 2, p. 947).
Modified Kanbur Zhang consists of taking Kanbur and Zhang data and applying somewhat different partitions (for explanation, see the text below).
1/ In 1983 prices. 2/ In 1990 prices, using provincial price deflators. 3/ Based on the most detailed partition used in the paper, e.g. 335 observations in
Akira and Kawamura paper etc.
Note: Theil (1) is Theil entropy index. Theil (0) is mean log deviation index. The definitions are:
n
Theil _1 = p yi yi
i log
i=1
n
Theil _ 0 = p
ilog
i=1 yi
Note: The Kanbur and Zhang (2003) and Bhalla, Yao and Zhang (2003) results should be very similar since their welfare concepts and the
aggregations/partitions are practically the same. And indeed, for the year 1995 (last year in Bhalla et al.) Kanbur and Zhang report a Concept 2 Theil of 17.7
while Bhalla et. al. have obtain a Theil of 16.5 (see the results in the Table here). However, inexplicably, in their paper Bhalla et. all (Table 2, p. 947) show the
total Concept 2 Theil of 13.87 while the sum of the three element into which the Theil is decomposed yields 16.54. I think that the latter amount is correct.
48
The decomposition rules can be also presented graphically. In the Akita and
Kawamura (2002) paper, total inequality is the sum of three inequalities given in
"bolded" boxes with their respective values in the year 2000 given underneath the boxes.
We can now easily see that provinces within each of the three large regions are fairly
homogeneous (in terms of per capita income) and that the bulk of inequality is
concentrated at the provincial level (differences in incomes between districts) and
between the three large regions (East, Central, West)
China =
Inequality between 3 large regions
+ Sum of inequaliies within each of 3
large regions
6.6
Within-provincial
Inequality between inequality
provinces (wihin each
of the 3 large regions)
2.8
Inequality between
districts within each
province
15.6
49
Kanbur and Zhang (2003) decomposition has three components:
Overall Overall
Urban Rural
mean mean
Province A
Ua Ra Province B
U R P
(1) Difference between the overall rural and urban mean, (2) inequality within the urban
means (area U), and (3) inequality within the rural means (area R). The component (1) is
equal to 13.9 Theil points, the sum of components (2) and (3) is equal to 10.9 Theil
points.
A modified Kanbur-Zhang (K-Zh) decomposition which we use here is to break
inequality by provinces. Then, the component (1) is inequality between all provincial
(total) means, and the component (2) is the sum of all inequalities between provincial
rural and urban means, that is inequality between Ua and Ra, plus inequality between Ub
and Rb and so forth. The component (1) amounts to 10.9 Theil points, 39the sum of rural-
to-urban inequalities within provinces amounts to 13.9 Theil points.
39Note that because of Theil's exact decomposition property, inequality among Ps must be equal to
inequality among U's plus inequality among R's, that is the sum of Kanbur and Zhang's components (2)
and (3) must be equal to our component (1).
50
The results of the modified Kanbur and Zhang (K-Zh) decomposition can then be
represented as
Inequality between 26 Sum of inequality within each
China = provinces + province
10.9
Sum of rural/urban
inequalities within each
province
13.9
These two decompositions suggest the following. First, note that the sum of two
shaded boxes in the Akura and Kawamura (AK) decomposition equals (by definition) the
shaded box in the modified Kanbur and Zhang decomposition.40 But while for Akira and
Kawamura, the importance of the two shaded boxes adds to 9.4 Theil points, the shaded
box in the (modified) Kanbur and Zhang decomposition amounts to 10.9 Theil points.
The difference must be due to the use GDP per capita in one study vs. expenditure per
capita from household surveys in another.
Second, from the Akira and Kawamura decomposition, we can conclude that
most of regional inequality in China is found between the three large regions and within
provinces (in other words, provinces contain districts with fairly unequal average
incomes). But, the Kanbur and Zhang decomposition suggests that these districts are not
random, that is they are not randomly poor or rich. The main line of differentiation goes
between rural and urban areas. For while AK decomposition shows that inequality
between districts (within all provinces) equals 15.6 Theil points, the K-Zh decomposition
40This must be true for Theil (which both authors use) because of its decomposability.
51
shows that 13.9 out of these 15.6 Theil points is due to the differences between mean
rural and mean urban incomes.41
Thus, in conclusion, we can say that if we break China into a very fine regional
partition, about 40 percent of thus calculated Concept 2 inequality is due to differences in
incomes between provinces. The bulk of the remaining 60 percent is due to the
differences in mean urban and rural income within provinces.
41Abstracting from the fact that the things are not fully comparable because A-K use GDP per capita while
K-Zh use household survey data.
52