Journal of International Development: Vol. 5, No.

2, 183-192 (1993)

MEASURING DEVELOPMENT? THE

UNDP’S HUMAN DEVELOPMENT
INDEX
MARK McGILLIVRAY
Deakin University, Geelong, Australia
and
HOWARD WHITE
Institute of Social Studies, The Hague, The Netherlands

Abstract: This paper critically examines the UNDP’s Human Development Index
(HDI). We conclude that its treatment of income is inappropriate; the lack of year-to-
year comparability is undesirable; it is robust with respect to measurement error; and
that its contribution to the assessment of development levels differs markedly among
country groups.

1 INTRODUCTION

The UNDP’s composite Human Development Index (HDI), based on life expect-
ancy, educational attainment and material living standards, is an attempt to reorient
the assessment of development levels away from income to more broadly based
measures. Whilst we are sympathetic to this attempt, we believe there are a number
of shortcomings in the construction and presentation of the HDI (for critiques of a
more ideological nature see Desai (1991), Rao (1991) and Hopkins (1991)).
Section 2 presents the index and details of its construction, leading in Section 3 to
an analysis of the income and HDI disparities between countries. In Section 4 we
examine the UNDP’s claim that the HDI ranks countries very differently to per-
capita income and the possibility that such differences as there are, are simply the
result of measurement error. Section 5 examines a number of possible refinements
to the index and concluding remarks are made in Section 6 .

2 THE HUMAN DEVELOPMENT INDEX

The H D I is defined as:

09.54-1748/93/020183-10$10.00
1993 by John Wiley & Sons, Ltd.
184 M . McGillivray and H . White

where Ijj is the ith indicator of human deprivation in country j (i = 1, 2 and 3 and j
- 1, . . ., 1). In the 1990 Report deprivation was calculated from life expectancy
-
( X l j )in 1987, adult literacy (X2,) in 1985 and the logarithm of purchasing power
adjusted G D P per capita (X3,) in 1987. Each variable is scaled from zero to one by
the formula:

where XT is a subjectively chosen desired value of the ith variable, X u is the

observed value of the ith variable for j and X y is the observed minimum. If a
country attains the desired value of the ith variable there is no relative deprivation
(Ijj = 0). Conversely, the greater the gap between the three XTs and the corres-
ponding Xjs the lower is that country’s HDI.
The desired value for life expectancy used in 1990 was 78 years, which is that
achieved by Japan in 1987, while that for adult literacy was 100 per cent. For
purchasing power adjusted per-capita GDP, the corresponding value is the
logarithm of $4861, which was the average official poverty line income in
nine industrial countries. The 43 countries whose per-capita GDPs exceeded
this amount were ‘capped’ at $4861. This treatment of income is discussed
below.
The 1991 version of the HDI differed from its predecessor in two main re-
spects.’ Firstly, literacy was replaced with ‘educational attainment’, by redefining
X 2 as:
1 -l

X?, = LITj + -3 YRS,

L
(3)

where LIT is the adult literacy rate in 1985 and YRS mean years of schooling
received per person aged 25 years and over in 1980. The second refinement was in
the treatment of income. The observed range of GDPs per capita were divided in
multiples of the poverty line y * :

x3J = YJ for 0 < y , 5 y*

= y* + 2 (y, - y * y * for y* 5 y , 5 2y*
= y* + 2 (y*)”’ + 3 (y, - 2y*)’/’ for 2y* 5 y , 5 3y*
= y* + 2 (y*)”* + 3 ( y * ) l / 3 + 4 ( y , - 3y*)”4 for 3y* 5 y, 5 4y* (4)
and so on.

where Yj is country j’s real G D P per capita and y* is a poverty line real G D P per
capita of $US4829. The 1990 HDI formulation by comparison was:

X3, = log y for 0 < y 5 y*

X3, = l o g y * fory >y*.

’Our comments are, for the moment, confined to the ‘basic’ HDI only, and therefore ignore innovations
such as the gender-sensitive, distribution-adjusted and intertemporal HDIs (UNDP, 1991, Technical
Notes 3 to 5 ) .
 The UNDP’s Human Development Index 185

Unlike the 1990 formulation, the 1991 formulation therefore allows Xjli to rise
marginally after exceeding the poverty line.

3 WHY THE DISPARITY BETWEEN THE HDI AND INCOME PER

CAPITA?

The 1990 Report makes two claims in particular for the HDI: (1) that the disparity
between countries is less for the HDI than it is for income per capita; and (2) that it
ranks countries very differently to income measures. In this section we consider the
first of these and the treatment of income in constructing the HDI.
The 1990 Report contains a graph headed ‘Ranking of countries’ GNP per capita
and HDI’. This graph is reproduced as Figure 1.2It is simply a graph of HDI and
GDP per capita scores (indexed so that the highest value equals 100 in our
example) each sorted separately in ascending order. There is no information on
how country rankings compare by the two criteria. The graph is accompanied in the
Report by the following statement:
The chart shows two separate distributions of countries. The upper curve
shows their ranking according to the human development index while the
lower curve shows their ranking according to GNP per capita. The two curves
reveal that the disparity among countries is much greater in income than in
human development. There is no automatic link between the level of per
capita income in a country and the level of its development (UNDP, 1990:
15).
This last statement cannot be verified from the graph, since the two indices are
ranked separately. But we can see that there are many countries with relatively
low per-capita income levels compared to a few (the top 20 per cent or so) with
much higher income-by contrast the HDI rises steadily over the range of coun-
tries. Thus it is correct that disparities in the HDI are less than those in GDP per
capita.
What is the source of this disparity? There are two potential sources. First, the
inclusion of life expectation and literacy. Second, the treatment of income in con-
structing the HDI. Physiological limits on life expectancy and a 100 per cent ceiling
on literacy restrict the disparity that may be displayed by these variables, so that
their inclusion in the HDI will clearly reduce disparity (though whether or not they
affect ranking will depend on their correlation with per-capita income).
There are two adjustments made to income in the 1990 Report. The first is to
take logs. A log transformation of income is appealing to economists, since it
captures diminishing marginal utility of income-that is, successive increments in
income give successively smaller increments in utility. Such a transformation ‘closes
the gap’ at the top end of the income scale, so that the disparity between countries
in log income per capita is much less than that in income per capita-and indeed

* W e have used the purchasing power parity (PPP) GDP per capita for this figure. The HDR uses GNP
per capita-which increases the disparity. We prefer the former measure as both the use of domestic,
rather than national, product and applying PPP exchange rates give a more accurate income-based
measure of development.
186 M . McCillivray and H . White

COUNTRY

Figure 1. Rankings by 1990 HDI and 1987 GDP per capita.

less than that in the HDI.3 So it is not only the inclusion of health and education
indicators that explains the more equitable spread of the HDI compared to per-
capita income, it is also the use of the log transformation on income itself.
Yet the 1990 Report does not stop with the use of logs. It goes on to impose a
ceiling above which increments in income are assumed to have no value to develop-
ment whatsoever. U p to the cut-off of $4861 the log value of GDP per capita is used
in calculating the index. At any higher income level the log of $4861 is used. This
appears a difficult position to support-would the authors of the Report accept a
marginal rate of taxation of 100 per cent above an income level of $4861 since this
apparently will not damage their welfare?
The 1991 Report conceded that the zero weight on income above the poverty
threshold was ‘too drastic’, and proposed a new treatment in which income below
the threshold was taken at its actual value and above that successive roots (square
root, etc., as shown in equation 4) were taken for each multiple of the poverty
level. In practice this new treatment makes little difference: the cut-off is practically
the horizontal one of the previous year. The first $4829 is valued at its actual value
and the second $4829 at only $139-the 100 per cent tax rate has been replaced
with one of 97 per cent!
The 1991 HDR defends its methodology in Technical Notes 1: reporting the rank
correlation of the HDI with the HDI constructed with different treatments of
income. These correlations are all very high-the lowest being 0.994 if unlogged
income per capita is used with no cut-off. Such high correlations are to be expected:
if the different components of the index are highly correlated then excluding one of

’The Gini coefficient calculated using the log of PPP GDP per capita is only 0.07. compared 0.34 for the
HDI. That using (unlogged) PPP GDP per capita is 0.53 (still rather less than those usually reported that
do not use PPPs, which are typically 0.65-0.70).
 The UNDP’s Human Development Index 187

them (even if only over a range of countries, as the capping procedure does) will
not greatly alter the ran king^.^ Two conclusions may therefore be drawn from the
UNDP’s own analysis of the different treatment of income (but which are not
drawn by UNDP themselves): (1) they add support to the position that income is a
good proxy for the other variables; and (2) if a simple treatment yields the same
result as a more complex one then why not adopt the simple treatment?
We conclude that the capping of income should be abandoned-the use of log
GDP per capita is a theoretically appealing and practically transparent alternative.

4 THE HDI: AN INSIGHTFUL NEW INDICATOR?

The UNDP has claimed that the ‘human development index ranks countries very
differently from the way GNP per capita ranks them’ (UNDP, 1990: 14). A number
of previous studies have assessed predecessors of the HDI on this criterion. For
example, Larson and Wilford (1979), using simple correlation coefficients, argued
that the ‘Physical Quality of Life Index’ (a composite of life expectancy, adult literacy
and infant mortality) was redundant in that it ranked countries in a ‘statistically
indistinguishable fashion’ to GNP per capita (Larson and Wilford, 1979: 583).
Similar exercises have been conducted by Hicks and Streeten (1979), who con-
sidered a ‘Development Index’ (a composite of 18 social and economic indicator^)^,
and McGillivray (1991), who looked at the 1990 HDI.
A problem with these studies is ambiguity over the extent of statistical associa-
tion, as measured by a correlation coefficient, that actually determines one variable
redundant with respect to another.6 An arbitrary threshold must be specified which
differentiates redundancy from non-redundancy. In this paper we opt for two
threshold values: 0.90 and 0.70. If the rank correlation coefficient between the HDI
and an existing indicator (in this case GNP per capita) is not significantly less than
0.90 we call this ‘Level 1’ redundancy. If it is not significantly less than 0.70 we call
this ‘Level 2’ redundancy. Clearly, the second of these tests is the more difficult for
the HDI to pass in terms of not being redundant.
Why have we chosen these levels? We believe that 0.90 is sufficiently high to say
that if two variables have a correlation of this order of magnitude then it is difficult
to claim that one is imparting additional information to that given by the first. The
lower threshold of 0.70 is used by existing studies, which imply that a new index is
redundant if the majority of its variation is accounted for by an existing indicator.
This can be formally interpreted as a test that the coefficient of variation be not less
than 0.50, which we convert to a correlation coefficient of 0.70.

Such a high coefficient is also not surprising, since there are over 100 countries below the threshold,
and their ranks change very little (only the denominator of their income term changes). On the other
hand, for those countries whose income is more than three times the threshold the Spearman’s is only
0.74. This rises to 0.95 for those countries whose per-capita GDP is two to three times the threshold.

’For further details of this index. see McGranahan et al. (1972) and for the PQLI see Morris (1979).
Larson and Wilford, using a sample of 150 countries, concluded that the PQLI was redundant vis-u-vis
GNP per capita on the basis of a rank correlation coefficient of 0.776. McGillivray drew the same
conclusion concerning the 1990 HDI and GNP per capita for a sample of 119 countries on the basis of a
rank correlation coefficient of 0.889.
188 M . McGillivray and H . White

A related question concerns the redundancy of the HDI vis-h-vis its own
components. It would be overly demanding to correlate the HDI with each of its
components, since this correlates the HDI partially with itself. Therefore, we
calculated a 'restricted HDI' based on two of its three components only, and
correlated these values with the excluded component. This not only tests the
redundancy of the index vis-a-vis the excluded component, both with and with-
out this component, but also the composition of the HDI. If the component is
correlated at the chosen threshold(s), then its inclusion provides little additional
insight.
Correlation coefficients for the total sample of 160 countries for 1991 and for
subsamples of low, medium and high human development (LHD, MHD and HHD
respectively) are given in Table 1.'

Table 1. Rank correlation coefficients between HDI. its components and GNP per capita
(1991).

Full sample LHD MHD HHD

Variables (160 countries) (n=63) (n=64) (n=53)

HDI,,3 and LE 0.927' 0.826' 0.196 0.605**

HDI,,3 and EA 0.852* 0.670'* 0.195 0.595**
HDI,,3 and G D P 0.842' 0.381 0.376** 0.680**

HDI and GNP per capita 0.910' 0.511** 0.464** 0.796*

LE and GNP per capita 0.863* 0.477** 0.065 0.669**
EA and GNP per capita 0.767** 0.287 -0.217 0.619**
G D P and GNP per capita 0.954* 0.686** 0.799* 0.937*

* Level 1 redundancy; * * Level 2 redundancy. LE is life expectancy, EA is educational attainment and

G D P PPP is G D P per capita. All data from UNDP (1991).

First consider the estimated correlation coefficients between HDI components.

For example, HD12.3is a composite of educational attainment (X,) and adjusted
GDP per capita ( X 3 ) .Without exception in the full sample there is a high correla-
tion between the restricted HDI and the excluded component, indicating redundancy
of inclusion. Much less consistent results were obtained with the subsamples: there
is redundancy of composition for the HHD countries but this is not so for the other
two subsamples. Of the six reported coefficients, three indicate non-redundancy.
Now consider the rank-order correlation coefficients between the HDI and GNP
per capita. Again, there is Level 1 redundancy for the full and HHD country
samples, and Level 2 redundancy in the other two subsamples. We also report rank
correlation coefficients between GNP per capita and the HDI components. In the
majority of cases these correlation coefficients are insignificantly different from, or
greater than, either threshold level. This would seem to lie at the heart of the
indicated redundancy of the HDI vb-a-vis GNP.
The correlation between the HDI based on any two of its components and the
remaining component, and of the HDI with GNP per capita, is generally less when

'Results were also calculated for 1990 and using industrial and developing country subsamples. These
results, which support the conclusions here. may be found in McGillivray and White (1992).
 The U N D P s Human Development Index 189

considering subgroupings rather than all countries. This is unsurprising in that small
variations in rankings have a greater effect in smaller samples. But is it possible that
this result arises from variations in rank that may be explained solely by measure-
ment error?
To investigate this question we constructed ‘true’ values of the HDI by assuming
that the reported value of each of the different indicators differs from the true value
by some random error.8 We then compared the true values with the observed
values for the HDI and each of its components, to determine how much difference
in ranking could arise simply from measurement error.
The results shown in Table 2 assume a 10 per cent error on all variables for all
groups. The changes in rank are not very large-the largest is 18 places. The
overwhelming number of cases move less than five places-resulting in rank

Table 2 . Difference and rank correlation between ‘true’and observed values

of the HDI and its components.
LE EA GDP HDI

Ranking changes
Maximum 18 14 6 22
Minimum 18 12 -5 -24
Distribution
-16 and below 2 0 0 4
-15 to -6 25 9 0 18
-5 to 5 108 138 160 110
6 to 15 24 13 0 24
16 and above 1 0 0 24
Standard deviation 5.83 3.94 1.31 6.57
Rank correlation
between ‘true’ and
observed values 0.992 9.996 1 .ooo 0.990
~~ ~~~~

LE is life expectancy, EA is educational attainment, and GDP PPP is GDP per capita.

correlation coefficients between true and observed levels of over 0.99 for every
single indicator and the HDI. The HDI thus appears to be robust with respect to
measurement error, a finding that was reinforced by repeating the analysis with
errors of up to k 15 per cent and by calculating the coefficients at the subsample
level (see McGillivray and White, 1992). The greater differences in ranking
observed within subsamples do not therefore simply reflect the effects of measure-
ment error.

5 POSSIBLE CHANGES TO THE INDEX

The HDI is least redundant, and therefore makes its greatest contribution to assess-
ment of inter-country development levels, when used to compare broadly similar
groups of countries. By contrast, if used to compare all countries it adds little new
information to that provided by per capita income or any of the index’s components

‘ A more detailed account of the procedure may be found in McGillivray and White (1992).
190 M . McGillivray and H . White

taken individually. The response of the UNDP to this problem appears to be to

try to modify the index. Here we discuss these attempts and also suggest other
modifications that might usefully be made to the HDI.
High correlation between the HDI’s components in large samples is a difficult
problem to overcome. It is necessary to try to find additional variables which satisfy
the extremely demanding criterion of influencing the index to the extent that it is no
longer redundant with respect to its components and GNP per capita. The 1991
H D R does this by proposing distribution-adjusted and gender-sensitive HDIs.
Regrettably, however, due to data limitations these indices can be applied only to a
much narrower band of countries. For this option to become operational the issue,
as Desai (1991), McGillivray (1991) and the UNDP itself acknowledge, becomes
one of data collection. But there is no guarantee that the resulting index will
provide any more insights than the recent versions. Indeed, the distribution-
adjusted and gender-sensitive HDIs have rank correlation coefficients of 0.989
and 0.804 respectively with GNP per capita, so that the problem of redundancy
remains.
Two further courses of action could be pursued. The first would be to reduce the
weight given to G D P per capita in constructing the index (since it is most highly
correlated with GNP per capita). This would be consistent with the general philosophy
of the Report that income ‘is merely useful and is for the sake of something else’
(Aristotle, quoted in UNDP, 1990: 9). But even with a zero weight to G D P per
capita, the rank correlation coefficient between the 1991 index and GNP per capita
is 0.832 for the full country sample. Alternatively, rather than claiming the HDI is
what it is not (i.e. it does not give a very different ranking of countries) the UNDP
would be better to justify the index solely by its philosophical basis-it measures
the ends not the means.
Two further aspects of the HDI could benefit from refinement: the definition of
the educational attainment variable and year-on-year comparability of the index.
The 1990 Report used adult literacy as a measure of educational attainment and the
1991 Report uses a weighted average of adult literacy (two-thirds) and mean years
of schooling (one-third). The weights are, of course, arbitrary and we concur with
Hopkins’s (1991) view that not much is to be gained by suggesting alternative
weighting procedures. However, literacy and years of schooling are variables of
very different orders of magnitude-the former being typically 10-20 times the
latter and so will dominate any composite of the two. T o combine these two
measures with weights of two-thirds and one-third (as the wording of the Report
suggests is the objective) would require first indexing the variables (possibly so that
the maximum value equals 100).
The H D I from different years is not comparable for two reasons. One is the
redefinition of the education and income variables. The other is that the range used
to scale the variables (their desired and minimum observed levels) changes each
year. (The ranks are not comparable for the further reasons that the 1991 Report
contains 160 not 130 countries, and that it ranks top-down, whereas the 1990
Report was bottom-up.) The 1991 Report explicitly recognizes this problem by
producing a technical annex entitled ‘Measuring Progress in Human Development’
(see pp. 96-97), which calculates the HDI for 110 countries for the years 1970 and
1985 on a consistent basis. Whilst refinements in the HDI are desirable, so is
consistency and comparability across time. Also, the rationale for shifting the range
 The U N D P s Human Development Index 191

(a concept of relative poverty) does not, in our view, offset the advantages to be
gained from having a comparable time-series. We would recommend that maxima
and minima be set for each variable that may be used across a range of reports.’

6 CONCLUSION

We have reviewed the usefulness of the UNDP’s HDI from a number of perspect-
ives. In terms of assessing development levels, we find the HDI to be largely
redundant vis-h-vis GNP per capita at varying degrees for various country samples.
It tends to make its largest contribution when looking at country subsamples
classified according to observed human development levels. The HDI itself and
these conclusions are shown to be robust in the face of possible measurement error.
Attempts to differentiate across the whole range have proved unsuccessful, and we
suspect this will remain the case. Thus, we would suggest the index be given a more
limited application than some of the grandiose claims in the Human Development
Reports. We also make a number of suggestions on improving the index, regarding
weights, composition and comparability.

ACKNOWLEDGEMENTS

An earlier version of this paper was presented at the ESRC Development Eco-
nomics Study Group Annual Conference, Leicester, March 1992. The authors
would like to thank the conference participants, especially Graham Pyatt, for
helpful comments and suggestions. Thanks are also due to participants at a Queen
Elizabeth House, Oxford, seminar, especially Sudhir Anand, Robert Cassen and
Frances Stewart.

REFERENCES

Desai, M. (1991). ‘Human development: concept and measurement’, Europenn Economic

Review, 35, pp. 350-357.
Hicks, N. and Streeten, P . (1979). ‘Indicators of development: the search for a basic needs
yardstick’, World Development, 7 , pp. 567-580.
Hopkins, M. (1991). ‘Human development revisited: a new UNDP report’, World Develop-
ment, 19, pp. 1469-1473.
Larson, D. A. and Wilford, W. T. (1979). ‘The physical quality of life index: a useful social
indicator?’, World Development, 7 , pp. 581-584.
McGillivray, M. (1991). ‘The human development index: yet another redundant composite
dcvclopment indicator?’, World Development, 19, pp. 1461-1468.
McGillivray, M. and White, H. (1992). ‘Measuring development? A statistical critique of the
UNDP’s Human Development Index’, ISS Working Paper. The Hague.

For adult literacy the maximum of 100 per cent presents itself. For the other variables maxima should
be set above the conceivably attainable values over the next, say, 10 years (e.g. 85 for life expectancy).
This will mean that n o country may score the maximum of unity, but this secms no particular loss. The
setting of minima is also problematic-if the lowest current value is used a country may fall below that in
a subsequent period. One alternative is to set minima of zero, the other is to pick values below which it is
believed no country will fall.
192 M . McGillivray and H . White

McGranahan, D . V . , Richaud-Proust, C . , Sovani, N . V. and Sabramanian, M. (1972).

Contents and Memurernent of Socio-economic Development. New York: Praeger.
Morris, M. D. (1979). Measuring the Condition of the World’s Poor: The Physical Quality of
Life Index. New York: Pergamon.
Rao, V. V. B. (1991). ‘Human Development Report 1990: review and assessment’, World
Development, 19, pp. 1451-1460.
U N D P (1990). Human Development Report 1990. New York: Oxford University Press.
U N D P (1991). Human Developrnenj Report 1991. New York: Oxford University Press.