Inequality, Poverty and Development

Montek S. Ahluwalia*
(World Bank, Washington, DC 20433, U.S.A.; May 1976, Revised Version August 1976)

1. Introduction
The relationship between the distribution of income and the process of development is one
of the oldest subjects of economic enquiry. Classical economic theory accorded it a central
position in analysing the dynamics of economic systems, and while this pre-eminence was
somewhat obscured in the heyday of neoclassical theory, in recent years it has again come
to occupy the center stage of development economics. The purpose of this paper is to
explore the nature of this relationship on the basis of cross country data on income
inequality. The use of cross country data for the analysis of what are essentially dynamic
processes raises a number of familiar problems. Ideally, such processes should be
examined in an explicitly historical context for particular countries. Unfortunately, time series
data on the distribution of income, over any substantial period, are simply not available for
most developing countries. For the present, therefore, empirical investigation in this field
must perforce draw heavily on cross country experience.

The results presented in this paper are based on a sample of 60 countries including 40
developing countries, 14 developed countries and 6 socialist countries.1 In the established
tradition of cross country analysis, the approach adopted is essentially exploratory. We have
used multivariate regression analysis to estimate cross country relationships between the
income shares of different percentile groups and selected variables reflecting aspects of the
development process which are likely to influence income inequality.2 The estimated
equations are then used as a basis for broad generalisations about the relationship between
income distribution and development. The difficulties inherent in this methodology are well
known, although all too often ignored. It is self-evident that the relationships thus identified
are primarily associational. They do not necessarily establish the nature of the underlying
causal mechanism at work for the simple reason that quite different causal mechanisms
might generate the same observed relationship between selected variables. Such alternative
mechanisms (or hypotheses) are observationally equivalent in the sense that our estimated
equations do not always permit us to choose between them.

The cross country relationships presented in this paper must be viewed in this perspective.
We should treat them as 'stylised facts' which can be observed, but which still need to be
explained, by an appropriate theory. The documentation of such 'stylised facts' is obviously
not the same thing as the development of a tried and tested theory, but it may contribute to
the development of such a theory in two ways. Firstly, the observed relationships may
suggest hypotheses about the nature of the underlying causal mechanisms at work, which
then need to be further tested and fashioned into a broader theory. Secondly, they provide

*
The author is on the staff of the Development Research Center of the World Bank.
The views presented are those of the author and not of the World Bank.
Acknowledgements are due to Bela Balassa, John Duloy, Ben King, Michael Lipton,
Graham Pyatt, Suresh Tendulkar and a referee for helpful comments. I am indebted
to Mariene Lehwing for computational assistance.
1
See Appendix for data sources and a brief discussion of data problems.
2
We have used income shares as the dependent variables instead of summary
indices of inequality such as the Gini coefficient because this permits us to focus
on the impact of the development process over different ranges of the income
distribution. The Gini coefficient is also a relatively insensitive measure and its
limited variation across countries makes it difficult to identify statistically significant
relationships. See, for example, Papanek (1976).
1
yardsticks for verifying theories of distribution and development by defining the observed
'behaviour' that such theories must explain.

What, then, do we know of the 'stylised facts' about income distribution and development. A
logical point of departure for our investigation is the hypothesis, originally advanced by
Kuznets (1955, 1963), that the secular behaviour of inequality follows an inverted 'U-shaped'
pattern with inequality first increasing and then decreasing with development. Following
Kuznets, the proposition that the distribution of income worsens with development, at least in
the early stages, has received considerable attention.3 More recently it has been advanced
in a much stronger formulation, which states that the process of development may lead not
just to increasing relative inequality, but also to the absolute impoverishment of the lower
income groups.4

These issues, together with several others, are examined in this paper. Section 2 presents
the cross section evidence on the relationship between inequality and the level of
development, showing a distinct U-shaped pattern as hypothesised by Kuznets. Section 3
extends our search for stylised facts by including a number of other correlates of inequality in
an expanded regression equation. Section 4 reviews the implications of the expanded
regression equations for the U-shaped curve and the hypothesis of absolute
impoverishment. Section 5 examines the relationship between inequality and the rate of
growth as a short-term relationship and attempts to draw some implications for policy.

2. Kuznets’ hypothesis: The ‘U-shaped curve’

We begin by documenting the evidence for Kuznets’ hypothesis that inequality tends to
widen in the early stages of development, with a reversal of this tendency in the later stages.
Following convention, we have tested this hypothesis by taking the per capita GNP of each
country (in USS at 1965-1971 prices) as a summary measure of its level of development and
including it as an explanatory variable in regression equations in which the income share of
different percentile groups is taken as the dependent variable.

The cross country regressions provide a substantial measure of support for the hypothesis
that there is a U-shaped pattern in the secular behaviour of inequality. Table 1 reports the
estimated equations describing the relationship between income shares of five different
percentile groups (the top 20 percent, the next or ‘middle’ 40 percent, the lowest 60 percent,
the lowest 40 percent and the lowest 20 percent) and the logarithm of per capita GNP.5 Two
equations are reported for each income share, one estimated from the full sample of 60
countries and the other estimated from the restricted sample of 40 developing countries only.
For the full sample, we have included a dummy variable for the socialist countries in order to
take account of the much higher degree of equality observed in these countries.6 The results
obtained can be summarised as follows:

(i) Taking the results from the full sample to begin with, there is clear evidence of a
nonmonotonic relationship between inequality and the level of development. The estimated
equations test for a quadratic relationship with the logarithm of per capita GNP. We find that

3
See, for example, Kravis (1960), Oshima (1962), Adelman and Morris (1973) and
Chenery and Syrquin (1975). For a recent review of this literature see Paukert (1973)
and Cline (1975).
4
This is argued in Adelman and Morris (1973. pp. 179-183).
5
'The logarithmic transformation gives equal weight to equal proportional differences
in GNP in measuring 'levels of development.' This has an intuitive appeal since
growth occurs at a compound rate over time.
6
Our results on this score are discussed in detail in section 3.
2
in all cases, both terms of the quadratic are significant7 and the coefficients have the
appropriate opposite signs to generate the U-shaped pattern hypothesised by Kuznets.
Income shares of all percentile groups except the top 20 percent first decline and then
increase as per capita GNP rises. Income shares of the top 20 percent display a
corresponding opposite pattern.

(ii) The turning point for income shares implied by the estimated equations are also reported
in table I. It is interesting to note that this turning point occurs at different levels of per capita
GNP for different income groups. In the case of the full sample, the turning point for the
income share of the top 20 percent occurs at per capita GNP levels of USS 364 (for the
economy as a whole) after which the income share of this group begins to decline. However,
the income shares of the middle 40 percent appear to improve after a per capita GNP level
of USS 291 is reached. As shown in table I, the turning point of income shares shifts
systematically further out as we go down the percentile groups, with the lowest 20 percent
having to wait until per capita GNP levels of about 5600 are reached. Taking these estimates
at face value (i.e. ignoring the question whether these estimated differences are significant),
the cross section evidence suggests that the reversal of the 'deteriorating phase' of relative
inequality begins fairly early, first for the middle income group and much later for the lower
income groups. It appears that if there is a ‘trickle down’ process, then it takes substantially
longer to reach the bottom!

(iii) The basic pattern described in (i) and (ii) above can also be discerned in the equations
estimated from the restricted sample of 40 developing countries, with slight differences.
Except for the equation explaining income shares of the middle 40 percent, the coefficients
on both terms in the quadratic are again significant and have the same sign patterns as in
the full sample, indicating a U- shaped pattern in the income shares of the lower income
groups offset by an opposite pattern in the income shares of the top 20 percent. The
absolute magnitudes of the estimated coefficients in these equations arc, however, some-
what different, implying that there are differences in the shape of the Kuznets’ curve
obtained in the two cases. The exclusion of the developed countries and socialist countries
from the sample has the effect of (a) shifting the turning point of the Kuznets’ curve slightly
inwards and (b) increasing the steepness of the observed U-shape in both phases. The
extent of the shift can be seen in fig. I, which compares the curve of estimated income
shares for the lowest 40 percent at different levels of per capita GNP from eq. (A.4), with the
curve of estimated shares from eq. (B.4), which includes developing countries only. We note
that the improvement in relative income shares in the later phases of development appears
markedly more modest if we look at the full sample than at the reduced sample of
developing countries only. This is because the quadratic formulation forces a symmetry of
shape with respect to the logarithm of per capita GNP and this allows the equations for the
restricted sample to fit a steeper U curve in conformity with the steepness observed over the
$75-500 range, which is where most of the developing countries are concentrated (see
Appendix, table 8). By contrast, the equation for the full sample is forced to fit a somewhat
less steep curve to reflect the relatively modest improvement in equality observed between
the middle income countries at the bottom of the U and the developed countries of today.

(iv) A major problem in interpreting the U shape revealed in cross country data is the
possibility that it may be generated solely by the fact that the middle income range is
dominated by countries with particular characteristics which generate high inequality. If so,
the U shape has little relevance to the long-term prospect facing the low income countries of
today, unless these countries share the same characteristics. Thus it is sometimes argued
that the U shape simply reflects the concentration in the middle income range of Latin
American countries, which display greater inequality because of particular historical and
structural characteristics not applicable to others. We have tested for this 'Latin America

7
Throughout this paper the term significant will be used to indicate that the estimated coefficients are
significantly different from zero with the sign indicated at least at the 10 percent level for a two-tailed
test. The critical value of / for this level of significance is 1.68 for our sample size.
3
effect' by including a dummy variable for the Latin American countries in each of the
equations (A.IMA.4). We find that the coefficient on this dummy variable is insignificant in all
cases and its inclusion leaves both the sign pattern and the significance of the coefficients
on the income quadratic largely unaffected.
Table 1 : The Kuznets curve.
Dependent variable Estimated coefficients on
explanatory variables a
Income shares of: Constant Log [Log Socialist R2 F D.W.b Turning point
per per Dummy per capita
capita capita GNP USS
GNP GNP] 2 (1965-71)
(A) Full sample
(1) Top 20 percent -57.58 89.95 -17.56 -20.15 0.58 27.9 2.05 364
(2.11) (4.48) (4.88) (6.83)
(2) Middle 40 87.03 -45.59 9.25 8.21 0.47 18.6 2.08 291
percent
(4.81) (3.43) (3.88) (4.20)
(3) Lowest 60 119.4 -73.52 14.06 17.52 0.61 31.4 1.97 412
percent
(5.85) (4.90) (5.23) (7-95)
(4) Lowest 40 70.57 -44.38 8.31 11.95 0.59 29.8 2.04 468
percent
(5.38) (4.61) (4.82) (8.45)
(5) Lowest 20 27.31 -16.97 3.06 5.54 0.54 24.3 1.93 593
percent
(4.93) (3.71) (3.74) (8.28)
(B) Developing countries only
(1) Top 20 percent -99.74 123.80 -24.18 0.12 3.6 2.24 363
(1.56) (2.35) O26)
(2) Middle 40 92.93 -49.13 9.65 0.01 1.4 2.19 351
percent
(2.12) (1.36) (1.32)
(3) Lowest 60 171.50 - 22.72 0.22 6.5 2.20 364
percent 116.40
(3.79) (3.12) (2.99)
(4) Lowest 40 106.80 -74.69 14.53 0.24 7.2 2.20 371
percent
(3.83) (3.25) (3.10)
(5) Lowest 20% 44.15 -31.33 6.07 0.22 6.3 1.98 381
(3.43) (2.%) (2.81)
Similarly, Papanek (1976) has argued that the observed U shape disappears when the
'strongly dualistic' countries are excluded from the sample. We note that our sample
excludes South Africa. Rhodesia, Libya, Niger and Trinidad, four of the seven dualistic cases
identified by Papanek, but this docs not affect the significance of the U shape.

a
statistics in parentheses
b
In estimating these equations the observations were entered in ascending order of per
capita GNP. The Durbin-Watson statistic therefore gives some idea of the pattern of
residuals with this ordering. The lack of serial correlation of residuals in the above equations
provides some reassurance that the quadratic formulation captures the underlying
nonlinearity reasonably well.
4
These results are broadly in line with the findings of Adelman and Morris (1973) and
Chenery and Syrquin (1975). They point to a marked decline in the relative income shares of
the lower income groups in the early stages of development and they also suggest that the
decline is most prolonged for the poorest groups. We will examine the welfare implications of
this pattern in section 4.

For the moment, we continue our examination of the U-shaped curve in order to determine
how far it tells the whole story. Judging this by goodness of fit, there is little doubt that the
equations in table 1 leave much to be desired. In the sample of developing countries, the
estimated equations explain just over a quarter of the observed variation in income shares of
the lower income groups and much less than this for the top 20 percent. The ⎯R2’s⎯R for the
equations estimated from the full sample are much higher, but this is to some extent due to
the fact that the inclusion of socialist countries in this sample adds substantially to the
intercountry variance in income shares and the dummy variable for these countries also
'explains' most of this added variance.

The relatively limited explanatory power of these estimated equations is hardly surprising.
The true relationship between inequality and development must be fairly complex, reflecting
the impact of a number of processes of structural change occurring with development. Such
a complex relationship obviously cannot be 'reduced' into a relationship with a single
explanatory variable. Per capita GNP is a useful summary measure of the level of
development8 in the sense that it is correlated with most of the processes occurring with
development, and as such, it may capture the net effect of these processes as observed in
cross country experience. The resulting estimated relationship is undoubtedly of some
interest as a possible indicator of the secular behaviour of inequality, but it tells us nothing
about the specific mechanisms through which development affects the degree of inequality.
Since it is precisely these mechanisms that are of interest, from the analytical as well as the
policy point of view, we need to extend our search for stylised facts to take account of them
to the extent possible. The next section is devoted to an exploration of the cross country
data along these lines.
3. Income distribution and economic structure

8
There are, however, major problems of comparability across countries in using per capita GNP as a measure of
the level of development. The use of official exchange rates to convert GNP measured in domestic currency to
GNP measured in USJ introduces obvious errors since exchange rates typically do not reflect purchasing power
parity. Kravis et al. (1975) investigating this problem found that the use of official exchange rates understates
GNP in developing countries compared to developed countries. Furthermore, within the developing countries
studied, the degree of understatement varies across countries sufficiently to create a switch in per capita GNP
rankings. These results indicate that there must be substantial measurement error in our explanatory variables.
5
Our search for the specific mechanisms through which development affects the degree of
inequality is necessarily limited by the availability of cross country data on explanatory
variables. The explanatory variables we were able to use are reported in table 2, where they
are grouped according to various aspects of the development process which are measured
(albeit somewhat crudely) by these variables. The relationship between income inequality
and these different aspects of development was explored by experimenting with alternative
combinations of these explanatory variables in explaining cross country variation in the
income shares of the different percentile groups.

Table 2 Explanatory variables used in cross country regressions.

Variables Correlation coefficient

with log of per capita GNP
1 Level and pace of development Logarithm of per capita 1.00 0.73 0.03
GNP (constant 1971 USS) Dummy variable for all
developed countries Rate of growth of GNP over the past 5-
10 years (percentage)
2 Education and human resources Literacy rate (percentage) 0.81 0.62 0.85
Primary school enrollment rate (percentage) Secondary
school enrollment rate (percentage)
3 Structure of production Share of agriculture in GDP -0.88 0.83
(percentage) Share of urban population (percentage)
4 Demographic characteristics Total Population Rate of -0.03 -0.55
growth of population (percentage)
5 Government activity Share of government revenue in GDP -0.24
(percentage) a
6 Other common variables reflecting structural commonality 0.01 0.12 -0.02
Dummy variable for Latin American countries Dummy
variable for Socialist countries Gini coefficient for land
distribution*
a
This variable is not available for all countries in the full sample. The correlation coefficient
slates to the sub sample for which observations are available.

In general, we find that three aspects of the development process appear to be

systematically related to the degree of inequality. These are:

(1) Intersectoral shifts involving a relative decline of the traditional agricultural sector and
a parallel shift of population to the urban sector.

(2) Expansion in the educational and skill characteristics of the population.

(3) The 'demographic transition' involving a reduction in the rate of growth of population.

Inclusion of explanatory variables reflecting these processes in our regression equations

substantially improves the goodness of fit obtained and the estimated relationships also
conform to a priori expectations about the impact of these recesses upon income inequality.

Table 3 provides an overview of the results obtained by the expanded regression equations
which include additional explanatory variables reflecting the impact of these aspects of
development. Equation (a) is a general form applied to all income shares and includes
variables with coefficients that are not always significant. Equation (b) for each income share
contains only those variables that are found to be significant in equation (a). Clearly, the
expanded equations have greater explanatory power than the equations reported in table 1,
in which development was measured solely in terms of per capita GNP. They explain

6
between two-thirds and three-fourths of the observed variation in income shares compared
to just over half in the case of the simpler formulation.9

A general problem in exploring these relationships is that the explanatory variables used to
reflect different aspects of development are all highly correlated with per capita GNP (sec
table 2) and this makes it difficult to attribute observed associations in a particular equation
to the impact of one or the other variable. This is an example of the problem of
'observational equivalence' of hypotheses mentioned in the introduction (section 1) and there
is no satisfactory solution to this problem. The approach we have adopted is to examine the
relationship with each variable under alternative specifications of the regression equation in
order to determine which relationships appear more stable in the face of inclusion and
exclusion of other explanatory variables.10

3.1. Intersectoral shifts

Intersectoral shifts occurring with development have long been recognised as a possible
mechanism through which the process of development affects inequality. This was first
pointed out by Ku/nets (1963), who argued that development typically involves accelerated
growth in the high income nonagricultural sectors, which slowly absorb population from the
low income, relatively stagnant, agricultural sector. Kuzncts showed that such a process
would lead to an increase in relative inequality in the early stages of development and, under
certain conditions, would generate precisely the U-shaped behaviour discussed above.

This effect of intersectoral shifts on overall inequality can be easily illustrated by a simplified
two-sector example.11 Taking the variance of the logarithms of income as a measure of
inequality, the variance V for the population as a whole can be decomposed into component
terms as follows,

where su is the population share of the urban (or nonagricultural) sector, Ku and Kr are the
variances of logarithms of income within the urban and rural sectors respectively, and Fu, Fr,
Fare mean logarithms of income for the two sectors and the whole economy respectively.
Using F = Fusu + Fr(l— sj, eq. (1) can be rewritten,

9
The increase in the variation explained is statistically significant at the 5 percent level in all cases.
10
It should be emphasised that this approach does not provide a rigorous basis for choosing between
alternative hypotheses. It is essentially a heuristic exploration of alternative pattern*.
11
For a fuller discussion see Robinson (1976).
7
 Table 3 Inequality and Development
Dependent variable Estimated coefficients on explanatory variables a
Income shares of: Constant Log per [Log per Share of Share of Literacy Secondar Populatio Socialist R2 F D.W.b Turning
capita capita Agricultur Urban rate y school n growth dummy point US$
GNP GNP]2 e in GDP population enrollmen rate
in total t
(1) Top 20 percent
(a) -8.711 49.620 -7.975 -0.258 -0.090 -0.094 -0.146 3.611 -9.443 0.75 23.4 1.90 1291
(0.26) (2.24) (2.10) (2.15) (1.58) (2.29) (2.63) (4.28) (3.27)
(b) -1.592 43.580 -7.157 -0.225 -0.107 -0.160 3.48 -9.287 0.75 25.7 1.95 1108
(0.05) (1.97) (1.87) (1.87) (2.60) (2.86) (4.09) (3.17)
(2) Middle 40 percent
(a) 34.27 -5.819 0.977 0.226 0.035 0.045 0.115 -2.448 0.751 0.70 18.0 1.82
(1.57) (0.40) (0.39) (2.86) (0.93) (1.68) (3.14) (4.43) (0.40)
(b) 30.46 0.172 0.154 -2.51 0.70 45.8 1.88
(14.63) (4.08) (6.49) (5.48)
(3) Lowest 60 percent
(a) 105.3 - 57.70 9.212 0.115 0.078 0.080 0.084 -2.50 10.43 0.74 22.47 1.96 1355
(4.02) (3.30) (3.08) (1.22) (1.72) (2.45) (1.91) (3.77) (4.59)
(b) 126.3 -68.470 10.68 0.068 0.076 0.089 -2.425 11.05 0.74 25.2 1.92 1605
(6.38) (4.53) (3.88) (1.53) (2.35) (2.03) (3.65) (4.96)
(4) Lowest 40 percent
(a) 74.500 -43.850 7.009 0.032 0.055 0.049 0.031 -1.161 8.702 0.68 16.48 2.06 1343
(4.01) (3.54) (3.30) (0.48) (1.73) (2.13) (1.0) (2.47) (5.40)
(b) 85.660 -51.440 8.41 0.057 0.056 -1.155 9.184 0.68 22.0 2.04 1144
(6.59) (5.29) (4.%) (1.85) (2.61) (2.48) (5.98)
(5) Lowest 20 percent
(a) 35.110 -21.24 3.337 -0.004 0.027 0.021 0.001 -0.333 4.697 0.58 11.37 1.83 1522
(3.73) (3.38) (3.10) (0.12) (1.67) (1.82) (0.08) (1.40) (5.75)
(b) 34.520 -21.00 3.315 0.027 0.022 -0.336 4.685 0.60 15.7 1.83 1454
(5.31) (4.32) (3.91) (1.77) (2.01) (1.44) (6.10)
a
t-statistics in parentheses.
b
In estimating these equations the observations were entered in ascending order of per capita GNP. The Durbin-Watson statistic therefore
gives some idea of the pattern of residuals with this ordering. The lack of serial correlation of residuals in the above equations provides some
reassurance that the quadratic formulation captures the underlying nonlinemrity reasonably well.

8
It is easily shown that on certain restrictive assumptions, the intersectoral shifts described above
generate a U-shaped pattern in inequality. Suppose, for example, that inequality within sectors
remains constant and incomes in the two sectors grow at the same rate so that ( F,,- F,)
remains constant. In this case, if there are no population shifts, our inequality measure V
remains constant. However, if development involves a progressive shift of population into the
high income sector then cq. (2) becomes a quadratic in .vu (the terms in square brackets being
constants). This function reaches a maximum for V at an urban population share given by

Clearly, development generates the familiar U shape in inequality provided the maximum occurs
within the relevant range 0 < su < 1. A sufficient condition for this is that | Vu - Vr | < [ Yu - Yr]2, i.e.
the squared difference between sectoral means is greater than the differences between the
within sector variances.12

The assumptions of constant inequality within sectors and equal growth rates in sectoral
incomes are obviously unrealistic. In fact we would expect both the degree of inequality within
sectors and mean income differences between sectors to change systematically with
development. Interestingly, there are plausible reasons for supposing that these changes might
reinforce the U-shaped pattern in overall inequality. For one thing, inequality in the urban sector
may itself follow a U-shaped pattern. It may increase initially, as accelerated economic growth
creates a strong demand for skilled labour, in the face of acute scarcities, with the result that
skill differentials expand. In the later stages, we can expect urban income to become more
equal as labour skills improve and become more widely dispersed in the population leading to
both an increase in wage share, as well as greater equality in the distribution of wage income
(sec subsection 3.2 below).

The ratio of mean incomes between sectors may also follow a U-shaped pattern with
intersectoral differences widening in the early stages, as scarce capital and other resources are
pre-empted by the modern (and typically privileged) urban sector, to the detriment of
productivity and income levels in the traditional sector. These differentials can be expected to
narrow in the later stages of development for two reasons. Firstly, as capital becomes less
scarce, more resources are likely to be made available to improve productivity in the low income
sectors. Secondly, as the size of the modern sector increases in the later stages of
development, its continued expansion has a proportionately larger impact on reducing the
pressure of population in the low income sector. Both factors lead to an accelerated increase in
productivity in the traditional sector in the later stages of development and can be expected to
reduce income differentials between sectors.

A systematic exploration of the impact of intersectoral shifts on inequality obviously calls for
data on inequality within each sector, sectoral mean incomes and sectoral population shares. In
the absence of such data we have used two explanatory variables which capture some aspects
of the process. These are the share of agriculture in GDP, which declines with development as
the non-agricultural sector grows at an accelerated rate, and the share of the urban population,

12
As pointed out by Robinson (1976), it is not essential for urban inequality to be greater
than rural inequality for the U shape to exist, provided the intersectoral mean
differences are large enough. As eq. (3) makes clear, Vu > Vu, only ensures that the
turning point occurs when more than half the population is in urban areas.

9
which can be expected to rise as population shifts away from the traditional agricultural sector.
These two variables are obviously closely related but they do reflect somewhat different aspects
of the same process. The share agriculture in GDP reflects the extent to which the income
generating capacity the economy has shifted into nonagricultural activity while the share of the
urban population reflects the extent to which this shift has been accompanied by increased
absorption of population into the nonagricultural sectors.

Our first step was to consider if either of the structural shift variables are associated with the U
shape in inequality. Table 4 summarises the results on this count. We note that either of the
intersectoral shift variables generate a shaped behaviour in inequality when entered in quadratic
form in the regression equations. It is obviously difficult to determine if the estimated
relationships reflect the effect of these variables or the effect of per capita GNP with which both
variables are highly correlated (see table 2). However it is worth nothing that these results are
slightly worse than those reported in table 1, in which the quadratic is defined in terms of the
logarithm of per capita GNP. We also find that the U shape defined in terms of per capita GNP
is more stable. When the quadratic formulation in either of the intersectoral shift variables is
included, together with the quadratic in the logarithm of per capita GNP, the coefficients on the
structural shift variables are no longer significant while the quadratic formulation in the per
capita GNP remains significant. Table 4 reports one equation in which the quadratic
formulations in all three variables included in the same regression equation. Again we find that
the quadratic per capita GNP remains significant.

10
 Table 4 Inter sectoral shifts and inequality.
Dependent variable Estimated coefficients on explanatory variables*
Income shares of: Constant Log per [Log per Share of [Share of Share of [Share of Socialist R* F
capita capita agricultur agriculture urban urban dummy
GNP GNP] 2 e in GDP in GDP]3 populatio population)
1
n
(1) Top 20 percent (a) 39.54 1.155 -0.018 -20.110 0.49 19.8
(13.73) (5.01) (4.63) (6.33)
(b) 52.08 0.28 -0.005 -18.140 0.43 15.8
(15.25) (1.62) (2.53) (5.39)
(c) 36.06 80.42 -15.64 -0.279 0.002 0.068 -0.005 -19.66 0.59 13.3
(0.73) (279) (2.64) (1.17) (0.66) (070) (0.92) (6.59)
(2) Middle 40 percent (a) 41.89 -0.671 0.010 8.453 0.34 10.9
(21.55) (4.31) (3.77) (3.94)
(b) 31.860 -0.050 0.002 7.066 0.28 8.5
(14.00) (0.43) (1.36) (3.15)
(c) 70.140 -41.88 8.870 0.093 0.002 0.249 -0.002 0.50 9.4
(2.16) (1.81) (2.27) (0.40) (0.55) (1.59) (1.25)
(3) Lowest 60 percent (a) 36.58 -0.867 0.014 17.24 0.55 24.5
(17.38) (5.15) (4.91) (7.42)
(b) 29.49 -0.310 0.004 15.99 0.50 20.4
(11.99) (2.46) (3.78) (6.54)
(c) 109.0 -65.67 12.12 -0.171 0.004 0.123 -0.000 17.50 0.62 14.8
(2.95) (2.49) (2.73) (0.65) (1.08) (0.69) (7.84)
(4) Lowest 40 percent (a) 18.56 -0.484 0.008 11.660 0.56 26.3
(14.25) (4.64) (4.61) (8.11)
(b) 16.060 -0.232 0.003 11.070 0.54 24.1
(10.71) (3.04) (3.69) (7.50)
(c) 66.010 -38.60 6.780 -0.162 0.003 0.031 0.000 12.090 0.61 14.4
(2.80) (2.30) (2.39) (1.0) (1.17) (0.27) (8.49)
(5) Lowest 20 percent (a) 6.02 -0.159 0.003 5.370 0.54 24.2
(10.11) (3.35) (3.6«) (8.18)
(b) 6.10 0.107 0.001 5.221 0.53 23.0
(9.00) (3.12) (3.39) (7.82)
(c) 27.570 -15.02 2.39 -0.112 0.002 0.011 0.000 5.712 0.57 12.1
(2.48) (1.90) (1.79) (1.41) (1.40) (0.21) (8.51)
a
t-statistics in parentheses.

11
These results suggest that the U-shaped relationship is better explained by per capita GNP than
the structural shift variables. We have, therefore, retained the quadratic formulation in per capita
GNP and entered the two structural shift variables as additional explanatory variables, but not in
quadratic form. When this is done we find that these variables are significantly related to income
inequality. The estimated relationships obtained are broadly comparable to those reported in our
expanded 'best' equations in table 3 (which include some additional explanatory variables). Our
general results on the impact of structural shifts on inequality can, therefore, be summarised
with reference to the results reported in table 3.

(i) We find that the share of agriculture in GDP and the urban share of total population are both
significantly related to the pattern of income inequality, but their effects on income shares of
different groups are not identical. The share of agriculture in GDP is not significantly related to
the income shares of the lowest groups, but it is positively related to the income shares of the
middle groups and negatively related to the income share of the top 20 percent (equa- tions (2b)
and (Ib) in table 3). By contrast, the share of the urban population in the total has no significant
effect on the income share of the middle group, but is significantly positively associated with the
income shares of the lowest groups and negatively associated with the income shares of the top
20 percent (equations (3b), (4b), (5b) and (Ib) in table 3).

(ii) These results point to an interesting asymmetry in the distributional impact of the
intersectoral shifts that occur with development. As the share of agriculture in GDP declines with
development, there appears to be a relative shift of income away from the middle group and
towards the upper groups. Alongside this 'disequalising' process, however, development also
generates a shift of population to the modern or urban sectors. According to the cross section
results, this latter process appears to favour the lowest income groups at the expense of the
rich.

The proposition that increasing urbanisation may raise the income shares of the lowest income
groups is consistent with a priori expectation. Given the dualistic nature of the development
process, a higher rate of urbanisation, other things being the same, reflects a wider access to
productive employment opportunities in the expanding nontraditional sector and a
correspondingly lower pressure of population in the rural areas. Both forces can be expected to
operate in favour of the lower income groups.13

The observed 'disequalising' impact of the decline in the share of agriculture in GDP in terms of
a shift from the middle income groups to the top is less easily explained. One possibility is that
as the relative size of agricultural activity diminishes, compared to nonagricultural activity, there
is a shift towards greater concentration of income and wealth because the nonagricultural sector
typically promotes larger size production units for both institutional and technological reasons. A
shift from small to large scale production can be expected to generate an increased
concentration of income in the upper income groups at the expense of the middle. We can also
speculate that the observed adverse effect on the middle groups is due to the fact that the
decline in the relative importance of agriculture probably has its strongest impact on small and
middle sized land holding cultivators who dominate the middle income groups. The slower
growth rate in agriculture implies a slower growth in income for these groups, which is in turn
reflected in a declining income share of the middle income group in the economy. The difficulty

13
Note that this is not to deny that in many circumstances the extent of migration may be excessive
compared to available employment possibilities. The urban population share variable measures the total
absorption of population in the urban areas which largely reflects expansion in employment opportunities.
These data provide no basis for measuring the degree of excess migration.

12
with this argument is that it implies that the poorest rural groups (e.g. landless labourers,
artisans, etc.) may also be adversely affected by the slower growth of agriculture and this
should be reflected in a positive relationship between the share of agriculture in GDP and
income shares of the lowest 40 percent or lowest 20 percent. Our results provide no evidence of
such a relationship. The equations in table 3 suggest that the income share of the poorest
groups are affected by the level of development as measured by per capita GNP, but not by the
relative importance of agriculture at a given level of development. One possible explanation is
that the poorer, landless groups arc relatively more mobile than landowning cultivators and are
more willing to shift from agriculture into unskilled employment in urban areas so that their
income share is not affected by the relative size of agricultural production. Note that since our
equations indicate that income shares of the lower groups fall as per capita GNP rises in the
early stages of development, the above result suggests only that the extent of the fall is
independent of the relative size of the agricultural sector.

3.2. Education and labour skills

Improvements in the educational characteristics and skill endowments of the labour force
provide another mechanism through which development affects inequality. The usual argument
is that this mechanism operates to promote income equality in the long run. The reasons
underlying this optimism are worth reviewing before considering our empirical results.

The central assumption underlying this view is that there is substantial scope for substituting
skilled for unskilled labour in the production process (particularly in a dynamic context) without a
decline in the marginal productivity in the former. This view of production and technological
change combined with, the conventional marginal productivity theory of factor rewards, implies
that a more skilled labour force will produce a shift from low paid, unskilled employment to high
paid, skilled employment. This shift, it is argued, produces higher labour incomes, a reduction in
skill differentials, and an increase in the share of wages in total output.14 This mechanism,
combined with economic policies that do not discriminate against labour-using and skill-
intensive production sectors, is often described as the key to the success of countries such as
Taiwan and Korea in achieving a rapid rate of development together with high growth rates of
employment and relatively equal income distribution.

In addition to the technological assumptions about factor productivity, there is also the argument
that skill intensive development patterns are less prone to income concentration than capital
intensive patterns. This is because of the peculiar characteristic of human capital-unlike physical
capital-that expansion in the stock of human capital in the economy necessarily involves
dispersion across a wider population. There is a limit beyond which human capital cannot be
accumulated in a single person, and at any rate it cannot be bequeathed across generations in
the same manner as physical capital. Both factors, it is argued, combine to generate strong
pressures towards equality in income distribution as the human resource endowment expands
with development.

14
Note that the assumption that a greater supply of skilled labour will not produce a sharp decline in its
marginal product is obviously crucial. Otherwise, skill upgrading will not have much effect upon total
output, and while it may reduce relative wage differentials it may also reduce the share of labour in total
income thus contributing to an increase in overall inequality. Indeed in these circumstances it is even
likely that skill differentials will not narrow because of the resistance of organised labour, and the result
will be cither open unemployment of skilled labour or displacement of unskilled labour by skilled labour in
the unorganised sector of the market. The existence of 'over-educated' manpower in many developing
countries may reflect just such a phenomenon. Proponents of the importance of education are usually
undismayed by this phenomenon and explain it away as an expansion in the 'wrong kind of education.'

13
The available data permit us to examine the education-inequality relationship in terms of three
explanatory variables which provide crude approximations to the level of human resource
development—the literacy rate, the primary school enrollment ratio and the secondary school
enrollment ratio. Because of the high correlation among these variables, it is necessary to be
somewhat selective in our choice of explanatory variables. We have chosen the literacy rate as
a measure of the basic education level of the stock of the population and the secondary school
enrollment rate as a measure of the degree of educational achievement beyond this basic
level.15 Two points about the exclusion of the primary school enrollment rate are worth noting.
First, the literacy rate provides us with a better measure of the basic educational level, being a
measure of the stock and not future additions to the stock. Secondly, we observe in our sample
that there is relatively little variation in the primary school enrollment rate across countries
beyond the middle level of development by when fairly high enrollment ratios are achieved in
most countries. The lack of variation obviously makes it less attractive as an explanatory
variable. The results obtained can be summarised as follows:
(i) There is clear evidence that education is significantly positively correlated with equality.
When the two education variables chosen —the literacy rate and the secondary school
enrollment rate-are entered in the regression equations without per capita GNP (equations (la),
(2a), (3a), (4a) and (5a) in table 5), we find that the secondary schooling variable is associated
with shifts in income from the top 20 percent to all other groups except the lowest 20 percent.
The literacy rate is not significant in any of the equations. When these variables arc included
together with the quadratic in per capita GNP the pattern changes in important respects. As
shown in table 5, the secondary schooling variable retains its positive impact on the middle
groups but the literacy rate variable now has a positive impact on the income shares of the three
lowest groups. It is tempting to conclude that the relationship thus revealed reflects the true
impact of literacy on inequality, which was previously swamped by the bias introduced by the
exclusion of per capita GNP.16 This basic pattern remains unchanged when the equations are
expanded to include other explanatory variables (see table 3).

(ii) The positive impact of education on relative equality appears to be quantitatively fairly
substantial. For example, an increase in the literacy rate from 10 percent to 60 percent is
associated with a 2.8 percentage point increase in the share of the lowest 40 percent. This
should be compared to an average share for this group of about 16 percent at low levels of
development. Similarly, an increase in secondary school enrollment from 10 percent to 40
percent is associated with an increase of 4.6 percentage points in the income share of the
middle 40 percent compared to an average share of 34 percent for this group at low levels of
development.

(iii) The fact that the secondary schooling variable benefits the middle groups while the literacy

15
Measuring the educational and skill levels of the labour force is no simple task. Ideally we would need
to develop appropriate definitions of skill levels and to measure (he skill structure of the labour force in
each country. Data of this sort relating to the stock of labour are not available for most of the countries in
our sample. The use of the school enrollment rate has obvious limitations since it is a flow variable and
not a stock variable.

16
Recalling that the decline in income shares with respect to per capita GNP was most
prolonged in the case of the lowest income groups (sec section 2), it is reasonable to
assume that the literacy rate, which is correlated with the per capita GNP variable, is
picking up the negative impact due to per capita GNP through most of the early stages
of development.

14
 rate benefits the lower group calls for some explanation. The most plausible explanation is in
terms of the likely beneficiaries of expansion in secondary schooling and expansion in literacy in
our sample. In general, the observed variation in secondary school enrollment (between 5 and
40 percent for most developing countries) is not such that the benefits of expansion in
secondary school enrollment are likely to have reached the poorer groups. Since access to
secondary schooling almost certainly expands from the top downwards, the observed range of
variation for most of the countries in our sample suggests that the lower income groups are
excluded from secondary schooling. By contrast, the variation in literacy rates observed in our
sample (between 10 to 80 percent for most developing countries) is such that the observed
expansion in literacy across different countries clearly does reach the lowest groups.

Table 5 Education and inequality.

Dependent Estimated coefficients on explanatory variables*

variable
Income shares Constant Log per [Log per Literacy Secondar Socialist R2 F
of: capita capita rate y school dummy
GNP GNP]1 enrollmen
t rate
(1) Top 20 (a) 58.76 0.022 -0.219 -14.85 0.56 25.9
percent
(28.60) (0.49) (4.32) (4.92)
(b) -57.340 85.150 - 14.420 -0.085 -0.172 -16.46 0.67 24.4
(2.11) (4.16) (3.84) (1.81) (2.70) (5.92)
(2) Middle 40 (a) 29.54 -0.014 0.152 5.16 0.52 22.3
percent
(23.27) (0.51) (4.86) (2.77)
(b) 81.080 -37.860 6.430 -0.033 0.130 5.900 0.57 16.7
(4.45) (2.76) (2.55) (1.05) (3.04) (3.16)
(3) Lowest 60 (a) 23.09 -0.015 0.143 13.50 0.54 24.5
percent
(14.25) (0.40) (3.58) (5.67)
(b) 124.40 -74.330 12.60 0.078 0.102 14.91 0.68 26.1
(6.54) (5.25) (4.44) (2.20) (2.10) (7.09)
(4) Lowest 40 (a) 11.70 -0.008 0.066 9.69 0.51 21.0
percent
(10.99) (0.11) (2.53) (6.05)
(b) 76.32 -47.35 8.00 0.052 0.042 10.57 0.65 22.5
(5.58) (4.60) (4.24) (2.21) (1.29) (7.55)
(5) Lowest 20 (a) 4.259 0.002 0.012 6.49 0.44 17.2
percent
(8.47) (0.18) (0.97) (6.46)
(b) 31.32 -19.67 3.27 0.024 0.005 5.10 0.57 16.8
(4.67) (3.90) (3.53) (2.12) (0.33) (7.44)
a
t-statistics in parentheses.

(iv) It is interesting to note that in the case of each of the education variables, the income share
of the relevant beneficiary group expands at the expense of the income share of the top 20
15
percent. This can be said to be an unambiguous improvement from the welfare point of view.

These results are broadly in line with earlier cross section studies-notably Adelman and Morris
(1973) and Chenery and Syrquin (1975 —in that they testify to the close relationship between
the process of educational and skill improvement, which occurs with development, and the
increase in relative equality. As we have seen, there are persuasive reasons for arguing that the
correlation reflects an important causal process, although even so the argument docs depend
upon some fairly crucial assumptions (see footnote 14) which may not hold in a number of
circumstances. Sceptics will also want to point out that the observed correlation may even
reflect the reverse direction of causation, i.e. educational levels may be determined by the
degree of inequality. We can only record that the cross section data are also consistent with this
hypothesis.17

3.3. Population growth and inequality

The reasons for expecting a particular relationship between income inequality and the rate of
growth of population are much less clear than the reasons for expecting the relationships
discussed above. For one thing, our understanding of the role of population growth in the
development process is fairly limited. In any case, most of the debate in this area has been
focussed on the relationship between population growth and the pace of development, while its
impact on the degree of inequality has received relatively less attention.18

In these circumstances, it is appropriate to begin with observed cross country experience - the
stylised facts-and then consider how far we have a theory to explain them. Our estimated
results unambiguously show high growth rates of population to be systematically associated
with greater income inequality. As shown in table 3, the rate of growth of population has a
significant positive impact on the income share of the top 20 percent, and a significantly
negative impact on the income shares of all other groups (except the lowest 20 percent for
which group the negative coefficient on the population variable is not significant). This general
pattern remains valid when the regressions are run with different combinations of other
explanatory variables and also with the reduced sample of developing countries only.

In interpreting this result, it is important to note that the relationship identified above holds after
controlling for other explanatory variables such as per capita GNP. In other words, we need to
ask ourselves why an economy with faster population growth would show greater inequality
when it reaches a given level of per capita GNP than another with a slower growth of population
observed at the same level of development. The literature suggests two possibilities, neither of
which has been fully explored.

(i) Perhaps the most important link between population growth and income inequality is provided
by the fact that different income groups grow at different rates, with the lower income groups
typically experiencing a faster natural rate of increase in population. Although we have not
allowed explicitly for intergroup differences in population growth (data on this subject are simply

17
A regression of the literacy rate and the secondary school enrollment rate on per capita GNP and the
income shares of the middle and lower groups yields significant positive coefficients on the latter.
18
The conventional wisdom in this area (which has been extensively aired in recent years) states that
there is a mutually reinforcing relationship between the pace of economic development and reductions in
the rate of growth of population. For an unconventional - and on the whole implausible-view that
population growth may actually provide the impetus to economic expansion in otherwise stagnant
societies, see Boserup (1965).

16
not available) it is arguable that high growth rates in total population reflect greater differentials
in population growth across groups, which in turn generate greater inequality. The argument can
be summarised as follows. It is well known that the process of development produces a
'demographic transition.' Suppose that this transition takes the form of a reduction in the natural
rate of population growth of each group as per capita income in the group rises, and there is a
flattening out of this response at high income levels. In that case, declines in the rate of growth
of total population will occur with development as the various groups slide down this curve, and
the flattening of the curve implies that this process will produce a narrowing of intergroup
differentials in population growth. The argument can be carried even further if we allow for the
fact that high population growth rates in some low income countries reflect an initial acceleration
in population growth as a result of lower mortality, after which they can be expected to go
through a 'demographic transition' as birth rates decline. It is reasonable to assume that this
type of acceleration in population growth occurs mainly in the low income groups where the
mortality rate is highest. These factors suggest that countries with high growth rates of
population probably suffer from larger intergroup differentials in population growth compared to
countries with low growth rates of population. It follows that in countries with high growth rates
of population, per capita income of the poorer groups will grow more slowly compared to per
capita income of the rich, leading to higher inequality at given levels of per capita GNP.

(ii) A second link between population growth and inequality is suggested by the fact that higher
growth rates of population imply greater pressure of labour supply on other productive factors
with a consequent deterioration in the share of labour in total output. This is especially so in the
presence of fixed factors such as land, which are likely to be particularly important in developing
countries. A higher population density generated by faster population growth is likely to produce
a higher rental share, which in turn generates greater inequality given the typically concentrated
pattern of land ownership. A similar argument can be advanced in the case of capital as a
productive factor. It is been argued that high growth rates of population lead to higher
dependency burdens which reduce the flow of private savings and also places a larger claim on
scarce public resources for 'nonproductive' public investment.19 Both factors tend to lower the
ratio of productive capital to labour. If we accept the conjecture that economics with a faster rate
of growth of population have lower 'equilibrium' capital-labour ratios at any level of development,
it is very likely that they will also have lower labour shares.20

3.4. Socialist countries

A consistent finding in all the equations estimated is that the six socialist countries in the sample
display substantially greater equality than is predicted by the cross country regression line. The
coefficient on the dummy variable for these countries is significantly negative in all the equations
explaining income shares of the top 20 percent and it is significantly positive for all other groups.
This is precisely what one would expect, given the absence of the disequalising effect of income
from property (i.e. land and capital), which is typically highly concentrated.

It is also of interest to consider how the absolute size of the coefficient on the dummy variable
changes as we allow explicitly for the impact of particular aspects of development. Taking the
various equations explaining the income share of the lowest 40 percent, for example, we find

19
See, for example. Left" (1969).
20
This relationship between labour shares and capital-labour ratios follows from
standard neoclassical assumptions, providing the elasticity of substitution is less than
unity.

17
that the coefficient on the dummy variable for socialist countries is about 12.0 in the case of the
basic equation reported in table 1. It drops to 9.7 when the educational variables are included in
table 5 and is further reduced to 8.7 by the inclusion of all the other explanatory variables in the
expanded equations reported in table 3. Thus a part of the much greater equality of the socialist
countries observed in the sample may be due to the fact that in these countries, the progress
made in expanding education levels and reducing the rate of growth of population is much
greater than for other countries at their income levels. However, it is clear that even after
allowing for these specific factors, the socialist countries display markedly greater equality.

In concluding this section, it is appropriate to remind the reader of some caveats. The results
discussed above serve to identify some of the correlates of inequality which provide a basis for
speculating about causal mechanisms. But it is important to note that our equations do not take
account of some of the crucial factors determining cross country differences in income
inequality. Two of these deserve special mention.

The most important omission from this point of view is our inability to examine the role of
differences in the concentration in ownership of productive assets, including land. This aspect of
economic structure is rightly regarded as a crucial determinant of income inequality and indeed
the greater equality observed in socialist countries testifies to this fact. Countries differ widely in
the degree of concentration in productive assets both in terms of the initial conditions in this
respect and the institutional and policy framework which determines the evolution of
concentration patterns over time. Unfortunately, for most countries, lack of data on patterns of
concentration of wealth make it impossible to quantify this relationship directly.21 We note that
we have experimented with the Gini coefficient of the distribution of agricultural land (which is
available for about thirty countries) as an explanatory variable, but no significant relationship
was identified between this variable and income inequality. This is not really surprising, given
that the direct impact on overall inequality will depend upon the size of the rural sector (in terms
of both population share and income share), which varies substantially across countries.

A second limitation of our exercise is the lack of explicit recognition of the role of the institutional
framework in which development takes place. The distribution of income generated by an
economic system is ultimately the result of a complex interaction between economic and socio-
political forces. Indeed, it can be argued that the very distinction between these forces results
from arbitrary classifications, which are adopted for 'analytical convenience,' at some risk of
over-simplification. For example, the distributional impact of an initial concentration of land
depends not merely upon the resulting concentration of 'equilibrium' factor incomes, but also on
its impact on the equilibrium itself, through the dominant position accorded to landowners in a
whole range of transactions. Similarly, improvement in labour incomes and labour shares in total
income, occurring with development, depends not only on the upgrading of skills and
technological 'factor productivities' as discussed above, but also on the growth of labour power
through social and political institutions. These socio-political factors are not easily quantified on
an ordinal scale, let alone the cardinal scale needed for regression analysis.22 Apart from the
introduction of a dummy variable for socialist countries, we have made no allowances for subtler
differences in socio-political structure. We note in passing that we have attempted to test for the

21
For a review of the importance of this aspect of economic structure in determining income inequality in
the economy and its overwhelming importance in the agricultural sector, see Chenery et al. (1974).
22
For an attempt at examining these relationships using analysis of variance techniques suitable for
ordinally measurable variables, see Adelman and Morris (1973). Measurement of this type inevitably
involves a high degree of subjective evaluation.

18
effect of the scale of government activity, as measured by the ratio of tax revenues to GNP, but
no significant relationship was discernible.

4. Relative inequality, absolute poverty and development

In the previous section, we showed that income inequality across countries is related not merely
to per capita GNP, but also to other variables measuring particular aspects of development. An
important feature of these results is that the inclusion of these additional explanatory variables
docs not swamp the observed relationship between inequality and per capita GNP. The
expanded equations reported in table 3 show a clear U-shaped pattern with respect to per
capita GNP, with the coefficients on the two terms of the quadratic being significant and of
opposite signs in the case of the top 20 percent, as well as the three lower income groups-the
lowest 60 percent, the lowest 40 percent, and the lowest 20 percent. In this section, we turn to
the implications of these results for the long-term path of inequality and the welfare implications
of this path.

4.1. The U-shaped curve revisited

The first point to note about the expanded equations reported in table 3 is that although they
confirm the U-shaped pattern, the absolute magnitudes of the coefficients on the quadratic
terms are substantially altered, implying a change in the curvature of the estimated U shape.

The nature of this change can be seen in fig. I, which compares the estimated U shape in the
income share of the lowest 40 percent as derived from the equations in table 1 with the U shape
as derived from the expanded equation in table 3.23 The U-shaped curve in the expanded
equation is substantially flattened out in the later phase when inequality declines. By
comparison, the earlier phase of sharply increasing inequality remains relatively unchanged.
This basic pattern of a flattening in the inequality reductions in the later phases of development
is repeated in the case of the income shares of the lowest 60 percent and the lowest 20 percent,
and there is a corresponding flattening of the inverted U shape for the top 20 percent also. In all
cases, the turning point is shifted substantially further out (compare results in table 3 and table
1).

The change in the estimated U shape suggests an interesting hypothesis for further study. It can
be argued that the two phases of inequality conventionally characterised by the U-shaped curve
arise from the operation of quite distinct forces. The reduction in inequality observed in the later
phases of development appears to be associated with a number of particular processes
occurring with development, such as the increasing absorption of labour in the relatively high
income modern sector, the expansion in education and the improvement in human resources,
and the reduction in population growth rates. Once we allow explicitly for the operation of these
processes (as we have done in the expanded equations) the ascending phase of the U-shaped
curve is almost completely damped.24 By contrast, we have not been able to isolate particular

23
In each case, the curve corresponds to the estimated share predicted by the regression equation for
different levels of per capita GNP, holding all other explanatory variables at their mean value for the
sample and holding the dummy variable for socialist countries at zero.
24
Note that the quadratic formulation is necessarily symmetric in logarithms of per capita GNP. Our
statement on the flattening of the curve in the ascending phase relates to the shape of the segment of the
quadratic over the relevant range. Attempts to estimate the exponent on the logarithm of income together
with the other parameters by using nonlinear estimation proved fruitless because of the flatness of the
maximum likelihood function. In fact, a wide range of values of α, β, γ and δ in the general form α + β log
Y + γ (log Y)δ can generate much the same U-shaped curve over a given range of Y.
19
processes occurring with development which underlie the initial phase in which relative
inequality deteriorates.

Can we explain this observed deterioration? Of the various relationships discussed in section 3,
only the process of a declining share of agriculture in GDP is associated with a deterioration in
inequality, and that too involves a shift from the middle group to the top, leaving the lower
income groups un- affected. If we focus on the income shares of the lowest 40 percent and the
lowest 20 percent, none of the additional explanatory variables with which we have
experimented help to explain this particular ‘stylised fact’ of cross country experience. We must
clearly look to country-specific studies for the answer.

4.2. The absolute impoverishment hypothesis

The prolonged decline in income shares of lower income groups in the early stages of
development has important welfare implications. There is clearly a very strong presumption that
the poor benefit from development much less than the rich. More seriously, it has been argued
by Adelman and Morris (1973) that developing countries face the prospect not just of increasing
relative inequality, but also of prolonged absolute impoverishment for the lower income groups.
This view needs to be examined in some detail.

The proposition that development may lead to absolute impoverishment of the poorer groups
obviously cannot be ruled out a priori. Such an outcome may result from the erosion of
traditional economic structures as a consequence of the expansion of the modern sector. An
aggressively expanding technologically advanced, modern sector, competing against the
traditional sector for markets and resources (and benefiting in this competition from an
entrenched position in the institutional and political context) may well generate both a relative
and absolute decline in incomes of the poor. Needless to say, history provides numerous
instances of the operation of such processes.25

Against this bleak view of the development process, there is another explanation of the
observed increase in relative inequality which is somewhat less pessimistic. On this view,
increasing relative inequality is not due to absolute impoverishment but to unequal benefits from
growth. Thus if economic expansion occurs in sectors and segments in which the initial benefits
accrue to the upper income groups, and if these groups have relatively weak income linkages
with the poorer income groups, we would expect income shares of the poorer groups to decline
without any decline in their absolute incomes. Interestingly, this disequalising impact is not
restricted to the case in which opportunities for economic expansion are limited to upper income
groups. It can be shown that if opportunities for economic expansion take a form in which
individuals in all income groups have prospects for an x percent increase in income (with x
constant across groups), but in all groups only a given percentage of the population (also
constant across groups) can actually achieve this, it may still be the case that income shares of
the poorer groups will decline.26

25
The European enclosure movements are often cited as an example in which the disruption of a
technologically inferior pre-existing system led to substantial absolute impoverishment. The decline of the
Indian cotton industry in the face of competition from Lancashire is another example in which an 'efficient'
system declined in competition with modern technology aided by a commercial policy designed to favour
the metropolitan power.
26
I am grateful to Ben King for this point which can be illustrated by a simple example. Consider an initial
distribution in which the poorest 10 people each have an income often units, the next ten each have an
income of 11 units and the richest 10 have incomes of 12 units each. The poorest 10 people have an
20
The differences between these two views of what lies behind the observed increase in relative
inequality are crucial. In the one case we assume that the disruption of low income traditional
economic activities is in some sense a necessary consequence (on some views it is even an
essential pre-condition!) for the growth of the modern sector. In the other case, the problem is
seen not as arising from a necessary contradiction.

Choosing between these views or, as is more likely, between an appropriate mix of these views
is ultimately an empirical issue to be resolved for each particular country experience. Cross
section analysis should not be used to derive general pronouncements to be applied to all
cases, but in keeping with our objective, it can help to document the 'stylised facts' of cross
country experience. The data permit two different tests of the absolute impoverishment
hypothesis. The first consists of using the estimated income shares of the lower income groups
from the regression equations described above to calculate the per capita absolute income of
these groups at different levels of per capita GNP.27 The second test consists of calculating the
average absolute income of the poorer groups for each country from the actual income shares,
and then estimating a cross country' relationship between this measure of average absolute
income of the poor and the level of development.

We find that both tests reject the hypothesis that the process of development produces
impoverishment in absolute terms.

(i) Table 6 presents estimates of average incomes of the lowest 60 percent, the lowest 40
percent and the lowest 20 percent at different levels of development based on income shares
for these groups predicted by eqs. (3b), (4b) and (5b), for the full sample of countries (table 3}
and from the same equations, estimated for the reduced sample of developing countries only. It
should be noted that the predicted income shares for this exercise are obtained by holding all
other explanatory variables constant at the mean value for the sample. This procedure can be
questioned on the grounds that these variables reflect different aspects of development and as
such it is not legitimate to hold them constant. An alternative would be to use income shares
predicted from the simpler formulations of table 1 in which per capita GNP is the sole variable
reflecting all aspects of development. We note for the record that the results obtained are
broadly in accord with table 6. In all cases, the estimated average absolute income of these
groups increases with per capita GNP.

(ii) These results are further supported by the attempt to estimate regression equations using
the logarithm of average absolute incomes of the three lower income groups as the dependent
variable. Table 7 reports the results of this exercise for the full sample of 60 countries as well as
for the reduced sample of developing countries only. The explanatory variables are the same as

income share of 30.3 percent. If half of each income group experience an increase in income of 20
percent, the resulting distribution will be one in which the income share of the poorest 10 will be 28.9
percent. It is easily seen that the decline in income shares arises because 'gainers' in a lower income
group overtake 'non-gainers' in a higher income group, and in the re-ordered distribution the lower
percentiles will always have a larger percentage of 'non-gainers.'
27
The per capita absolute income of the ith percentile is calculated as (silni) Y where si and ni are the
income share and population share of the ith percentile group and Y is the per capita GNP. Ideally we
should use average personal income instead of per capita GNP hut data on personal income are not
available for most of the countries in the sample. Since the ratio of personal income to GNP probably falls
as per capita GNP rises, our measure understates the income of the poor in the low income countries
relative to the poor in the high income countries. This exaggerates the extent to which absolute income of
the poor rises with development.

21
those used to explain income shares in the expanded equation in table 3 and include a
quadratic in the logarithm of per capita GNP in order to test for the existence of a phase of
absolute impoverishment. We find that although the coefficient on the first term in the quadratic
is negative in the equations for the lowest 40 percent, and the lowest 20 percent, it is not
significant in cither case. In any case, the coefficients imply no decline in absolute income over
the relevant range.

Table 6 Estimated average absolute income levels of low income groups. a

Lowest 60 percent Lowest 40 percent Lowest 20 percent

Per capita GNP Full Developing Hill Developing Full Developing
in US$ sample countries sample countries sample countries
75 51.6 45.5 42.o 37.8 32.6 27.7
100 63.2 55.4 50.2 42.0 38.3 31.7
200 103.8 89.0 77.3 63.7 56.3 44.9
300 140.3 118.2 101.2 84.6 70.8 57.5
400 175.2 145.5 124.0 106.4 84.1 71.1
500 209.4 171.8 146.7 129.6 96.8 86.3
600 243.4 197.6 169.6 154.4 109.4 103.1
700 277.4 223.1 193.0 180.7 122.1 121.6
800 311.6 248.5 217.0 209.1 135.0 141.6
900 346.0 273.9 241.6 238.9 148.2 163.3
1000 380.7 299.4 266.9 270.2 161.8 186.5
1500 560.0 403.6 235.7
3000 1158.9 909.1 519.1
a
The figures in the first column for each income group are based on income shares predicted
by eqs. (3b), (4b) and (4c) from table 3, and the figures in the second column arc based on
income shares predicted by the same equation estimated from the reduced sample of
developing countries only. In the latter case, absolute incomes for per capita GNP levels above
US$ 1000 are not shown because the predicting equation for income shares was estimated
from a sample which did not extend beyond this level.

The behaviour of absolute income of the poor over the relevant range is perhaps best seen in
terms of its elasticity with respect to per capita GNP. This can be directly obtained from the
estimated equations in table 7. Differentiating partially with respect to the logarithms of per
capita GNP gives 5 + 2$ [log per capita GNP] as the elasticity at that level of per capita GNP
(where & and $ are the coefficients on the two terms of the quadratic). As shown below, the
elasticity for all three groups obtained from eqs. (la), (2a) and (3a) from table 7 is positive, albeit
less than unity (reflecting the increase in relative inequality).28 Furthermore, it rises as per capita
GNP rises.

Elasticity of absolute income with respect to per capita GNP

at:
Per capita GNP US$ Per capita GNP Per capita GNP
100(1965-71) US$ 500(1965-71) US$750 (1965-71)
Lowest 60 percent 0.64 0.73 0.89
Lowest 40 percent 0.45 0.77 0.85

28
When the equations reported in table 7 are estimated without the squared term, the coefficient on the
logarithm of per capita GNP yields a constant elasticity for the absolute income of each group. In this
case, the elasticities for the three groups are 0.89, 0.85, and 0.76 respectively, with t-ratios of 13.9, 9.0
and 5.1 respectively.
22
 Lowest 20 percent 0.18 0.65 0.77
Table 7 Cross country regressions using absolute average income.
Dependent variable Estimated coefficients on explanatory variables a
Logarithms of Const Log per (Log Share Share Litera Second Popul Social R2 F D.W. b
average absolute ant capita per of of cy ary ation ist
income of: GNP capita agricu urban rate school growt dumm
GNP)2 lture pop. enrollm h rate y
in in ent
GDP total
0) Lowest 60
percent
(a) Full sample 0.959 0.070 0.143 0.002 0.001 0.001 0.001 - 0.106 0.98 444.8 2.03
(b) Developing (2.10) (0.23) (2.75) (1.05) (1.55) (2.60) (1.81) 0.052 (2.68) 0.94 87.8 2.02
countries 0.720 0.252 0.111 0.002 0.001 0.002 0.001 (4.47)
(0.79) (0.35) (0.79) (1-16) (0.86) (2.32) (1.21) -
0.061
(3.80)
(2) Lowest 40%
(a) Full sample 1.620 -0.475 0.231 0.001 0.001 0.002 0.001 - 0.189 0.97 214.7 2.02
(2.47) (1.09) (3.08) (0.28) (1.32) (2.36) (1.01) 0.047 (3.31)
(2.80)
(b) Developing 1.828 -0.697 0.289 0.001 0.001 0.002 0.001 - 0.88 41.2 2.04
countries (1.41) (0.69) (1.44) (0.50) (0.56) (2.16) (0.49) 0.055
(2.42)
(3) Lowest 20
percent
(a) Full sample 2.479 -1.183 0.340 -0.001 0.002 0.002 0.001 - 0.32 0.91 75.8 1.72
(2.31) (1.65) (2.77) (0.27) (I.II) (1-83) (0.39) 0.269 (3.46)
(1.00)
(b) Developing 3.216 -1.967 0.534 0.001 0.001 0.003 0.000 - 0.72 15.3 1.83
countries (1.55) (1.21) (1.66) (0.20) (0.32) (1.85) 0.036
(0.99)
a
t-statistics in parentheses. b In estimating these equations the observations were entered in
ascending order of per capita GNP.
It is worth noting, however, that the elasticity for the lowest 20 percent, although positive, is
much lower than for the other groups and takes much longer to rise to unity.

We conclude that while the cross country evidence points to unequal benefits from growth, it
does not support the hypothesis of a prolonged decline in absolute incomes for the poor as
development proceeds. There are, of course, a number of reasons why we should be careful in
interpreting this result. The percentile groups we have defined are very large, and increases in
'average incomes' for such broadly defined groups may hide significant declines in absolute
incomes for particular socio-economic groups which arc offset by growth in income for other
groups. Such offsetting changes may be an important aspect of the dynamics of income

23
distribution in developing countries but they cannot be captured by the traditional approach to
distributional problems which focusses on income shares of ordered percentile groups.29

A systematic treatment of the absolute impoverishment hypothesis obviously calls for an

examination of the trends in per capita income in particular socio- economic groups in particular
countries. It will be some time before reliable time series data suitable for such studies become
available, but until then it is worth noting that while cross section data strongly support the
hypothesis of worsening relative inequality in the early stages, they do not support the
proposition that absolute income levels of large sections of the population also decline.

5. Growth and inequality: The short-term relationship

Thus far, our analysis has focussed on what are essentially long-term relationships between
inequality and development. The worsening in relative inequality observed in our cross-country
data occurs over the phase of development presented by the transition from per capita GNP
levels of USS 75 to per capita GNP levels of around USS 750. For an economy experiencing
growth in per capita GNP at the rate of about 2.5 percent per year, this transition would take
about a hundred years.30 While long-term relationships are of considerable interest, it is
important to note that some of the debate in this area has focussed on much more short-term
impacts.

The importance of making a distinction between the secular relationship between inequality and
levels of development on the one hand and the short-term relationship between inequality and
growth on the other, is not always clearly recognised. For example, it has been suggested that
high growth rates in some of the developing countries observed over comparatively short
periods - e.g. Brazil between 1960 and 1970-have led to a marked increase in income
inequality. To some extent, the long-term relationship discussed above would produce precisely
such an outcome-a higher rate of growth observed over a given period raises the level of
development above what it would otherwise be, and this in turn affects inequality. However, the
debate on this issue has a somewhat different flavour. There is a definite suspicion that there
are short-run mechanisms which are quite distinct from any structural or long-term factors, and
which generate greater inequality as a direct consequence of faster growth. This raises the
question of whether the degree of inequality may be affected not only by the level of
development but also by the speed at which this level is achieved.

A plausible reason for the existence of such a relationship may be found in a number of short-
term pressures associated with high growth rates. For example, the existence of lags in factor
mobility across regions or sectors will ensure that as opportunities for accelerated growth arise
in particular regions or sectors, economic expansion creates factor market disequilibria which
may generate significant income differentials. Such disequilibria are likely to be more severe in

29
In a recent study, Adelman and Robinson (1976) have drawn attention to such offsetting changes in the
position of socio-economic groups as an important feature of the response of income distribution to policy
intervention.
30
This time frame is broadly consistent with that suggested by Kuznets (1955) on the basis of historical
evidence from some of the now developed countries. Kuznets observed that the time series evidence for
these countries indicated a steady reduction in inequality over the first half of this century, but he
contrasted this trend with the more probable trend of widening inequality in the early stages of their
industrialisation.

24
a high growth situation than in a more stable low growth situation.31 Income differentials
generated through this mechanism obviously need to be distinguished from those which reflect
long-term structural factors such as the endowment of labour skills of the economy as a whole.
They are perhaps better seen as being superimposed upon the structurally determined income
differentials, serving (at least in market economies) as necessary lubricants to overcome any
frictional resistance to a higher rate of expansion.

The cross country data permit only a crude test for the existence of a short- term relationship of
this type. We can include the rate of growth of GDP for each country over the ten years
preceding the point at which the distribution is measured as an additional explanatory variable in
our regression equations and consider if (other things being the same) faster growing countries
display greater inequality. The results obtained from this exercise reject the hypothesis that a
faster rate of growth leads to greater inequality. The coefficient on the growth rate variable
remained insignificant in a number of experiments with alternative combinations of explanatory
variables, including some in which the per capita GNP variable was not included.

The absence of any significant relationship between inequality and the rate of growth is a
potentially important aspect of observable cross country experience. One interpretation is that
while there may be a secular time path for inequality which developing countries must traverse,
and which contains a phase of increasing inequality, there is at least no evidence that countries
which traverse this path at a fast pace are worse off at the same level of development than
countries which traverse it at a slower pace. Such a proposition, if accepted, has dramatic policy
implications. It suggests that policymakers are perhaps best advised to think of the rate of
growth as determining essentially the speed of transition through the different phases of
development and inequality: higher growth rates accelerate this transition without necessarily
generating greater inequality than can be expected given the structural characteristics of the
economy at each level of development.

It would be naive to pretend that so important a conclusion can be firmly established on the
flimsy basis of the lack of an observable relationship between inequality and the rate of growth.
Precisely because of its importance, however, it is worth examining the results more carefully to
extract a legitimate interpretation. A major limitation of our methodology is that it does not permit
us to go beyond the simple measurement of rates of growth to examine differences in the type
of growth achieved in different situations. Since it is precisely these differences which will
determine the distributional impact of growth, we cannot hope to provide any insight into this
question by focussing solely on the levels of growth achieved. But if we cannot deny that certain
types of high growth processes lead to greater inequality than can be structurally expected, we
can at least assert that the cross section evidence docs not suggest that all fast growers
systematically display this pattern. If there are countries in which this is true, there are others
which display the opposite pattern, so that no systematic pattern emerges across countries.
Such time series evidence as is available tends to support this point of view.32 The experience
of Brazil, where the high growth was accompanied by worsening relative inequality, can be
contrasted with the experience of Taiwan, where substantial growth has taken place with an
actual reduction in income inequality.

31
A thorough examination of this issue would require us to distinguish between an increase in the growth
rate observed in transit to a high growth dynamic equilibrium and a high growth rate observed in an
established dynamic equilibrium. Clearly such distinctions go well beyond the capacity of our
methodology.
32
See, for example, Ahluwalia (1974).

25
Recognising this diversity of country experience is perhaps the most important lesson to be
learned from the data. At the very least, it shifts attention from an unquestioning suspicion of
high growth rates as such towards an examination of the particular nature of growth in different
countries and the implications of different types of growth for inequality. A systematic
investigation along these lines can only be conducted in the context of in-depth analysis of the
historical experience of particular countries.

6. Conclusions

It was stated at the outset that the objective of this paper is primarily exploratory. It could not be
otherwise, given the limitations of cross country analysis.

The results presented above therefore should not be viewed as definitive either in defining the
prospects facing particular developing countries or in providing unambiguous policy guidelines.
They are best viewed as a useful documentation of empirical regularities-the so-called 'stylised
facts' of cross country experience. More ambitiously, they can also be viewed as providing some
clues to the mechanisms through which the development process affects the degree of in-
equality. They can be no more than clues precisely because the essential com- plexity of a
dynamic process, and its great variety across countries, cannot be adequately captured in a
single equation.

Subject to these limitations, our cross country results can be summarised as follows.

(i) There is strong support for the proposition that relative inequality in- creases substantially in
the early stages of development, with a reversal of this tendency in the later stages. This
proposition holds whether we restrict the sample to developing countries or expand it to include
developed and socialist countries. Furthermore, it appears that the process is most prolonged
for the poorest group.

(ii) There are a number of processes occurring pan passu with development which are
correlated with income inequality and which can plausibly be interpreted as causal. These are
intersectoral shifts in the structure of production, expansion in educational attainment and skill
level of the labour force, and reduction in the rate of growth of population. The operation of
these processes appears to explain some of the improvement in income distribution observed in
the later stages of development, but they do not serve to explain the marked deterioration
observed in the earlier stages.

(iii) The cross section results do not support the stronger hypothesis that the deterioration in
relative inequality reflects a prolonged absolute impoverishment of large sections of the
population in the course of development. The cross country pattern shows average absolute
incomes of the lower percentile groups rising as per capita GNP rises, although slower than for
upper income groups.

(iv) Finally, the cross section results do not support the view that a faster rate of growth is
systematically associated with higher inequality than can be expected given the stage of
development achieved.

Appendix: Data sources and problems

The income distribution data used in this paper arc summarised in table 8. The sample of 60
countries was selected from a recent compilation of available cross country data undertaken in
the World Bank's Development Research Center and reported in Jain (1975). The source for

26
 each observation is given by the number in parentheses which corresponds to the source
number reported in Jain (1975).

The quintile shares given in table 8 differ slightly from those implied by the decile shares
reported in Jain (1975). This is because the latter are obtained from a Loren/ curve fitted to the
original data. Our figures are read off a Lorenz curve plotted through individual points recorded
in the original source. The main reason for preferring this method is that the fitted curve does
not necessarily pass through the observed points and since these points correspond to deciles
or quintiles in many cases, it is arguably more appropriate to use the exact figures. All equations
reported in the paper have also been estimated using the data reported in Jain (1975) and the
results are substantially unaffected.

The choice of 60 countries from a potential total of 71 countries was dictated by particular
judgments about the reliability of data in some cases. Although multiple observations were
available for some countries, we have restricted our data set to one observation per country on
the grounds that adding more than one observation for some countries would give too much
weight to particular country experience. Wherever available, household distributions have been
used, but in many cases we have had to use the distribution of individuals.

It is important to emphasise that our selection process does not necessarily ensure that the sixty
observations selected come up to some objective standard of quality. Income distribution data
are notoriously deficient and the many sources of error affecting them are well known. These
sources of error may be broadly grouped into three categories.

(i) There are a number of conceptual and definitional problems in measuring income inequality
and available surveys do not display any uniform practice in handling these problems. For one
thing, the concept of income that is relevant for the study of inequality is not easy to define
uniquely. It should obviously include subsistence income (valued appropriately) and the case
can even be made that it should refer in some sense to •permanent' income, smoothing out both
life cycle variations as well as purely stochastic variations around the life cycle profile. There are
also obvious problems associated with inequality measures using money incomes for groups
facing very different price levels.

(ii) There are well-known sampling problems which limit the reliability of measures of income
inequality based on survey data. These problems are aggravated by the fact that many of the
available surveys on which in- equality measures are based were not originally designed to
provide reliable measures of income inequality.

Table 8 Income distribution at different levels of development.

Income shares of quintiles

Country Per capita Bottom Second Third Fourth Top
GNP in 20% quintile quintile quintile 20%
US$ (1970
prices)
Developing countries
1. Chad 58(1) 79.3 7.3 10.3 17.0 22.0 43.0
2. Malawi 69(1) 80.0 3.8 9.1 13.3 18.6 33.2
3. Dahomey 59(1) 91.3 3.0 10.3 14.3 20.0 30.0

27
 Income shares of quintiles
Country Per capita Bottom Second Third Fourth Top
GNP in 20% quintile quintile quintile 20%
US$ (1970
prices)
Developing countries
4. Pakistan 63-64a 93.7 6.5 11.0 15.5 21.5 43.3
5. Tanzania 67 (1) 103.8 5.0 9.0 12.0 17.0 57.0
6. Sri Lanka 69-70 (3) 108.6 6.0 11.0 16.5 20.5 46.0
7. India 63-64 (4) 110.3 5.0 11.0 13.0 19.0 52.0
8. Malagasy 60(1) 138.7 5.5 8.0 9.5 16.0 61.0
9. Thailand 62 (1) 142.8 5,7 7.2 11.9 17.5 57.7
10. Uganda 70(1) 144.3 6.2 10.9 13.9 21.9 47.1
11. Kenya 69(1) 153.2 3.8 6.2 8.5 13.5 68.0
12. Botswana 71-72(1) 216.6 1.0 5.9 14.5 19.7 58.9
13. Philippines 65 (1) 224.4 3.9 7.9 12.5 20.3 55.4
14. Egypt 64-65 (1) 232.8 4.2 9.8 15.5 23.5 47.0
15. Iraq 56(1) 235.5 2.0 4.8 9.2 16.0 68.0
16. El Salvador 61 (1) 267.4 5.5 6.5 8.8 17.8 61.4
17. Korea 70 (5) 269.2 7.0 11.0 15.0 22.0 45.0
18. Senegal 60(1) 281.8 3.0 7.0 10.0 16.0 64.0
19. Honduras 67-68 (1) 301.0 2.0 4.4 9.1 19.5 65.0
20. Tunisia 70 (2) 306.1 4.1 7.3 12.0 21.6 55.0
21. Zambia 59(1) 308.2 5.6 9.0 11.9 16.5 57.0
22. Ecuador 70(1) 313.6 2.5 3.9 5.6 14.5 73.5
23. Turkey 68(1) 322.2 3.0 6.5 11.1 18.8 60.6
24. Ivory Coast 70 (2) 328.7 3.9 6.2 11.8 20.9 57.2
25. Guyana 55-56 (1) 350.8 4.0 10.0 16.8 23.5 45.7
26. Taiwan 68* 366.1 7.8 12.2 16.3 22.3 41.4
27. Colombia 70 (7) 388.2 3.5 5.9 12.1 19.1 59.4
28. Malaysia 70 (3) 401.4 3.4 8.0 12.6 20.1 55.9
29. Brazil 70 (4) 456.5 3.1 6.9 10.8 17.0 62.2
30. Jamaica 58(1) 515.6 2.2 6.0 10.8 19.5 61.5
31. Peru 70 (4) 546.1 1.5 5.0 12.0 21.5 60.0
32. Lebanon 55-60(1) 588.3 5.0 8.0 10.0 16.0 61.0
33. Gabon 68 (2) 608.1 3.3 5.5 7.9 15.8 67.5
34. Costa Rica 71 (2) 617.1 5.4 9.3 13.7 21.0 50.6
35. Mexico 69 (5) 696.9 4.0 6.5 9.5 16.0 64.0
36. Uruguay 67 (1) 720.8 4.3 10.0 15.1 23.2 47.4
37. Panama 69 (2) 773.4 2.9 6.5 12.5 18.8 59.3
38. Spain 64-65 (1) 852.1 6.0 11.0 15.7 22.1 45.2
39. Chile 68 (1) 903.5 4.5 8.5 12.7 17.5 56.8
40. Argentina 61 (1) 1004.6 7.0 10.3 13.1 17.6 52.0
41. Puerto Rico 63 (1) 1217.4 4.5 9.2 14.2 21.5 50.6
42. Japan 68 (2) 1712.8 4.6 11.3 16.8 23.4 43.8
43. Finland 62 (1) 1839.8 2.4 8.7 15.4 24.2 49.3
44. Netherlands 67 (2) 2297.0 3.1 10.5 16.4 21.5 48.5
45. France 62(1) 2303.1 1.9 7.6 14.0 22.8 53.7
46. Norway 63(1) 2361.9 4.5 12.1 18.5 24.4 40.5
47. United Kingdom 68 (2) 2414.3 6.0 12.8 18.2 23.8 39.2
48. New Zealand 70-71 (3) 2501.5 4.4 12.5 18.6 23.5 41.0
49. Denmark 63(1) 2563.9 50 10.8 16.8 24.2 43.2
50. Australia 67-68(1) 2632.4 6.6 13.5 17.8 23.4 38.7
28
 Income shares of quintiles
Country Per capita Bottom Second Third Fourth Top
GNP in 20% quintile quintile quintile 20%
US$ (1970
prices)
Developing countries
51. Germany, W. 70 (2) 3208.6 5.9 10.4 15.6 22.5 45.6
52. Canada 65 (1) 3509.6 6.4 13.6 16.5 23.3 40.2
53. Sweden 70 (2) 4452.2 5.4 9.9 17.6 24.6 42.5
54. United Stales 70 (3) 5244.1 6.7 13.0 17.4 24.1 38.8
Socialist countries
55. Bulgaria (1) 406.9 9.8 15.2 18.0 22.0 35.0
56. Yugoslavia 68 (2) 602.3 6.5 12.0 17.0 23.0 41.5
57. Poland 64(1) 660.8 9.8 13.6 18.1 22.5 36.0
58. Hungary 67 (1) 872.7 8.5 15.5 19.0 23.5 33.5
59. Czechoslovakia 64 (1) 887.7 12.0 15.6 19.0 22.4 31.0
60. Germany, E. 70 (1) 2046.3 10.4 15.8 19.8 23.3 30.7
a
The income distribution for Pakistan is for East and West Pakistan taken together as reported
in source 1 for Pakistan from Jain (1975).
b
The income distribution for Taiwan is from W. Kuo, 'Income distribution by size in Taiwan area-
Changes and causes,' in: Income Distribution Employment and Economic Development in
South East and East Asia, papers and proceedings of (he seminar sponsored by the Japan
Economic Research Center and the Council for Asian Manpower Studies, vol. 1.

(iii) Finally, quite apart from sampling errors, there are non-sampling errors that are particularly
serious in measuring income distribution. It is well known that response bias may lead to
intentional understatement of incomes at the upper end of the income range and there may also
be some overstatement at the lower end. More generally, it is widely recognised that surveys
which include only a few questions on income arc likely to elicit highly inaccurate statements
about actual incomes from most people.

The data we have used are undoubtedly subject to all these limitations and the result is that our
estimates of income distribution are subject to substantial measurement error. In defense of the
use of such data for cross section analysis we have only the familiar excuse: the presence of
random error in the data serves only to hide cross country patterns rather than to generate
spurious patterns. Such patterns as can be discerned despite these errors therefore deserve
serious, if critical, consideration.

29
References

Adelman, I. and C.T. Morris, 1973, Economic growth and social equity in developing countries
(Stanford University Press, Stanford, CA).
Adelman, I. and S. Robinson, 1976, Income distribution policy in developing countries: The case
of Korea (Stanford University Press, Stanford, CA) forthcoming.
Ahluwalia, M.S., 1974, Income inequality: Some dimensions of the problem, in: Chenery et al.
(1974).
Boserup, E., 1965, The conditions of agricultural growth: The economics of agrarian change
with population pressure (Chicago).
Chenery, H.B. and M. Syrquin, 1975, Patterns of development - 1950-70 (Oxford University
Press, Oxford).
Chenery, H.B., M.S. Ahluwalia, C.L.G. Bell, J.H. Duloy and R. Jolly, 1974, Redistribution with
growth: An approach to policy (Oxford University Press, Oxford).
Cline, W.R., 1975, Distribution and development. Journal of Development Economics 1, no. 4,
359-402.
Jain, S., 1975, The size distribution of income: A compilation of data (Johns Hopkins University
Press, Baltimore).
Kuznets, S., 1955, Economic growth and income inequality, American Economic Review,
March.
Kuznets, S., 1963, Quantitative aspects of the economic growth of nations: VIII distribution of
income by size. Economic Development and Cultural Change, January.
Kravis, B., Z. Kennessey, A. Heston and R. Summers, 1975, A system of international
comparisons of gross product and purchasing power (Johns Hopkins University Press,
Baltimore).
Kravis, B., 1960, International differences in the distribution of income, Review of Economics
and Statistics, November.
Leff, N.H., 1969, Dependency rates and savings rates, American Economic Review, December.
Oshima, H., 1962, The international comparison of size distribution of family incomes with
special reference to Asia, Review of Economics and Statistics, November.
Papanek, G.F., 1976, Economic growth, income distribution, and the political process in less
developed countries, Paper presented at a Symposium on Income Distribution and Economic
Inequality, June, Bad Homburg, West Germany.
Paukert, F., 1973, Income distribution at different levels of development, International Labour
Review, August-September.
Robinson, S., 1976, A note on the U hypothesis relating income inequality and economic
development, American Economic Review, June.

30