NBER WORKING PAPER SERIES

DECOMPOSING THE TWIN-PEAKS IN THE

WORLD DISTRIBUTION OF OUTPUT-PER-WORKER

Paul Beaudry
Fabrice Collard
David A. Green

Working Paper 9240

http://www.nber.org/papers/w9240

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
October 2002

This paper was written while the second author was visiting the Department of Economics, University of
British Columbia. We have benefited from discussions with D. Acemoglu, C. Blackorby, A. Kotwal, T.
Lemieux, G. Verdier, A. Young and from presentation at the CIAR 2002 summer workshop. The views
expressed herein are those of the authors and not necessarily those of the National Bureau of Economic
Research.

© 2002 by Paul Beaudry, Fabrice Collard, and David A. Green. All rights reserved. Short sections of text,
not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including
© notice, is given to the source.
Decomposing the Twin-peaks in the World Distribution of Output-per-worker
Paul Beaudry, Fabrice Collard and David A. Green
NBER Working Paper No. 9240
October 2002
JEL No. O33, O41

ABSTRACT

This papers examines changes in the distribution of per-worker-output across countries over
the period 1960-98, with a particular focus on identifying the forces behind the hollowing out of the
middle of the distribution and the associated emergence of a twin-peaks phenomenon. The main
finding of the paper is that most of the change in shape of the world distribution of income between
1960-1998 can be accounted for by changes in the parameters driving the growth process. In
particular, we show that role of physical capital investment and population growth in affecting
output growth has increased substantially over the period and that this increase can account for all
the hollowing-out of the distribution. In contrast, we do not find that changes in the distribution of
variables played much of a role, nor do we find any significant effects coming through non-linear
convergence mechanisms or increased importance of education. Our results suggest that research
aimed at understanding changes in the world distribution of income should focus on explaining why
the social returns to physical capital accumulation where so high over the period 1978-98. The paper
ends by discussing elements that help understand this phenomena.

Paul Beaudry Fabrice Collard David A. Green

Department of Economics CNRS{GREMAQ and IDEI Dep. of Economics
Univ. of British Columbia Universite de Toulouse I Univ. of British Columbia
997-1873 East Mall Manufacture des tabacs, batiment F 997-1873 East Mall
Vancouver, B.C. 21 allee de Brienne Vancouver, B.C.
Canada, V6T 1Z1 31000 Toulouse Canada, V6T 1Z1
CIAR and NBER. France green@econ.ubc.ca
paulbe@interchange.ubc.ca fabrice.collard@univ-tlse1.fr
Introduction

Over the period 1960 to 1998, the shape of the distribution of output per worker (and output
per capita) across countries has changed considerably. As discussed in Quah [1997] and Jones
[1997] among others, in 1960 this distribution was clearly uni–modal and close to log–normal.
However, over time, the middle of distribution hollowed–out quite substantially, as mass moved
away from the mean of the distribution towards two new modes. This process — which Quah
named the twin–peaks phenomenon — is the object of study of this paper. The approach we
adopt in the paper is to decompose the observed changes in the world distribution of output per
worker1 over the period 1960–1998 into three main components: (i) changes in the distribution
of driving forces that affect growth across countries, (ii) changes in the importance of these
driving forces and (iii) changes in the distribution of residual or unobservable factors affecting
growth.2 We believe that providing such a decomposition will help orient research interested
in the world distribution of income in the same manner that within-country decompositions of
wage distributions have helped direct recent research in labour economics.

The first observation we make, which thereafter influences our approach to the data, is that the
twin–peaks phenomenon is especially noticeable among non–sub–Saharan African (NSSA here-
after) countries.3 This observation is of interest since it immediately eliminates the possibility
that the twin–peaks phenomenon is driven exclusively by the poor growth performances of the
sub–Saharan African countries. We therefore find it useful to first focus on understanding the
changes in the distribution of output per worker across NSSA countries before tackling the whole
distribution. The second important observation we make is that the hollowing–out of the mid-
dle of the output per worker distribution (at least for NSSA countries) essentially began around
1978. In effect, before 1978 this process was not at all apparent among NSSA countries. This
leads us to pose the question: what differences between the period 1960–1978 and 1978–1998 can
account for the observed changes in the shape of the world distribution of output per worker?

The main finding of the paper is that most of the change in shape of the distribution of output
per worker in NSSA countries between 1960–1978 versus 1978–1998 can be accounted for by
changes in the importance of two traditional growth factors, that is, increases in the impact
of the rate of investment in physical capital (or the savings rate) and the rate of population
growth on output per worker. Indeed, our estimates indicate that the importance of these
factors has increased by a magnitude of about 3 within the 1960–1998 period. In contrast, we
do not find that changes in the distribution of investment rates, rates of population growth,
school enrolment rates or residual factors have played much of a role. Similarly, we do not
1
Throughout the paper we will refer to the world distribution of output as the distribution of output across
countries, and not the distribution of income across individuals around the world. The later distribution has
recently been studied by Sala–i–Martin [2002].
2
Our approach builds on the labor literature that studies changes in the distribution of wages (see e.g. Juhn,
Murphy and Pierce [1993]).
3
Our use of the term sub Saharan actually refers to countries neither in northern nor southern regions of
Africa.

3
find any significant effects coming through a non–linear convergence mechanism, scale effects,
or increased importance of schooling. As we will show, this result extends to the case where we
include all African countries. The only additional factor that appears important when including
all African countries in our study is the poor residual draws experienced by sub Saharan African
countries in the early period of our study, that is the period 1960–78.

The remainder of the paper is organized as follows. In section 1, we document changes in the
distribution of output per worker across NSSA countries between 1960 and 1998. In particular,
we document when the twin peaks phenomena emerged. In section 2, we present our method
of decomposing change in the distribution of output per worker into changes due to observable
and unobservable factors. Section 3 focuses on the stability of the growth process over the
period since it is necessary step prior to implementing the decomposition. Section 4 implements
this decomposition on our sample of NSSA countries. Our main finding that the changes in
importance of two factors — the investment rate and population growth — can account for
most of the change in the shape of the distribution. In section 5, we introduce sub–saharan
Africa countries into our sample show that a very similar conclusion emerges. Finally, in section
6 we discuss a set of issues that help refine the set of possible explanations that lay behind this
process.

1 The emergence of the twin–peaks phenomenon

This section describes the changes in the distribution of output per worker in the set of NSSA
countries available in the World–Penn Table 6.0 over the period 1960 to 1998. The list of
countries composing our sample is given in appendix A.3.4 ’5 The measure of output used is
real GDP evaluated at constant world prices6 and the denominator in our output per worker
measure is the number of individuals aged between 15 and 64.7

Figure 1 reports the distribution of (log) output per worker across the set of NSSA countries
in 1960 and 1998. The plotted distributions are kernel density estimates based on a Gaussian
kernel.8 Both distributions are expressed as deviations from the given year’s mean in order
to emphasize changes in the shape of the distribution. It should be noted that the actual
4
We included in our sample all countries that did not have missing data for more that two years between 1960
and 1998. This gave us a sample of 75 countries, with 69 countries having no missing data, and 6 countries with
at most two years of missing data. For these 6 countries, the missing data was linearly interpolated. All our
results are robust to excluding these 6 countries from the sample.
5
See appendix A for greater details on the data.
6
There are several measures of real GDP in the World Penn Tables. Our analysis is robust to the use of these
different measures.
7
It would be preferable to have a more precise measure of the number of workers, but this is not available for
a sufficiently wide set of countries. For OECD countries, we have examined the potential biases induced by not
taking into account differences across countries in participation rates. We found that using a direct measure of
number of workers for OECD countries provides even stronger evidence in favor of the main claim of the paper
than that found using population data (see Beaudry and Collard [2002]).
8
Details can be found in appendix B.

4
distribution shifted substantially to the right from 1960 to 1998. In fact, the average output
per worker increased by 134% between 1960 and 1998 for this set of 75 countries, implying an
average annualized rate of growth of 2.27%. Figure 1 also reports the points in the distribution
associated with the interquartile ranges in 1960 and 1998.9 As noted in the introduction, we
begin our exploration by focussing on the set on NSSA countries and leave the analysis of the
whole sample until section 5.10 The advantage of adopting this approach will become clear later.

The distributions in figure 1 display several prominent features. First, the shape of the distribu-
tion changed considerably from 1960 to 1998. In particular, the distribution of 1960 (log–)output
per worker was uni–modal and, apart from thick tails, was close to normal. As time has passed,
mass shifted away from the centre towards two new modes, leading to a bi–modal distribu-
tion in the late nineties. This observation corresponds to the so–called twin–peaks phenomenon
documented in Quah [1993] and Jones [1997].

Figure 1: Across-Country Income Distribution: 1960–1998

0.5
1960
0.45 1998

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−3 −2 −1 0 1 2 3

Note: The dashed lines indicate the upper and lower bounds delimiting
the interquartile range.

As can also be seen from figure 1, the change in the distribution takes the form of a widening
of the interquartile range. Indeed, the interquartile range expanded quite substantially between
1960 and 1998. This movement is also documented in table 1 which reports the interquantile
range of the distribution in 1960, 1978 and 1998 and the changes that took place over each
sub–period. As can be seen in the table entries, the interquartile range increased by about
26.5% over the period. This widening is also found at other decile-ranges of the distribution,
9
The points of the interquartile range are calculated from the raw data and not from the kernel estimates of
the density function.
10
The expositional gains achieved by this approach will become clear in Section 4.

5
for example at the 20–80, 30–70 and 40–60 percentile differences. However, the 5–95 percentile
difference declined over the same period, indicating there was actually inward movement in the
tails at the same time as the widening of the interquartile range.11

Table 1: Changes in the interquantile ranges

Ranges Changes
1960 1978 1998 1960–1978 1978–1998 1960–1998
05–95 2.6340 2.5252 2.2923 -0.1087 -0.2329 -0.3416
10–90 2.2697 2.0666 1.9051 -0.2030 -0.1615 -0.3646
15–85 1.7623 1.3964 1.8404 -0.3659 0.4440 0.0782
20–80 1.3857 1.2781 1.5925 -0.1076 0.3144 0.2068
25–75 1.0570 0.9465 1.3223 -0.1104 0.3758 0.2653
30–70 0.8704 0.8513 1.0674 -0.0190 0.2161 0.1971
35–65 0.6043 0.5966 0.8204 -0.0077 0.2238 0.2161
40–60 0.3380 0.3835 0.4861 0.0456 0.1026 0.1481

Since a widening of an inter-quartile range can arise in many different forms, it is useful look
at how the fraction of countries lying within a given window around the mean changed over
time. For example, let us take a window of ± 50% (i.e. 0.5 log–points) around the mean and
examine the fraction of countries within that band in 1960 versus 1998. 12 Table 2 reports these
numbers. In 1960, 49% of countries fell into the ± 0.5 log–points window, while in 1998 only
39% of countries fell within the same window. The phenomenon is also apparent if we examine a
window of ± 0.7 log–points around the mean, as the mass of countries lying within this band was
61% in 1960 and only 49% in 1998. These numbers — when taken together with the percentile
differences — indicate a widening process that has taken place around the interquartile range
and corresponds to a hollowing out of the middle of the distribution with mass moving towards
two modes, but without a fattening of the tails.

Table 2: Mass around the mean

Mass round the mean Changes

∆ 60 78 98 60–78 78–98 60–98
0.50 49.333 46.667 38.667 -2.667 -8.000 -10.667
0.60 54.667 58.667 44.000 4.000 -14.667 -10.667
0.70 61.333 66.667 49.333 5.333 -17.333 -12.000
0.80 65.333 73.333 60.000 8.000 -13.333 -5.333
0.90 70.667 78.667 73.333 8.000 -5.333 2.667
1.00 78.667 81.333 78.667 2.667 -2.667 0.000

11
Accordingly, the variance of the distribution only increased marginally over the period.
12
This amounts to compute the mass of countries lying within the band [y t − ∆; y t + ∆] where ∆ determines
the width of the window.

6
The second issue we want to examine is the timing of this hollowing–out process. Panels (a) and

Figure 2: Distributional Dynamics

(a) Percentiles differences (b) Mass around the mean
80
1.6 20−80
25−75
30−70

% of the distribution
1.4 70

1.2 60

1 50
0.6
0.7
0.8 0.8
40
1960 1970 1980 1990 1960 1970 1980 1990 2000
Years Years

(b) of figure 2 plot, respectively, key movements in percentile differences and our mass around
the mean statistics for the entire period 1960 to 1998. From panel (a), one can see that the
20–80 percentile difference, the interquartile range and the 30–70 percentile differences were all
rather stable through the sixties and much of the seventies but that somewhere around 1978 all
three began to rise sharply. The interquartile range, for example, actually declined by about
11% from 1960 to 1978 but increased by almost 40% between 1978 and 1998. For easy reference,
table 1 reports the actual levels of these on other percentile ranges for the years 1960, 1978 and
1998. Thus, figure 2 suggests that the particular widening of the distribution observed in figure
1 has arisen only since 1978.

The notion that there was an important change starting around 1978 is reinforced in panel (b)
where we plot the time series for the fraction of countries lying within a given window around
the mean. We report information corresponding to windows of ±0.6, 0.7, and 0.8 log–point
around the mean. The actual values for these windows for the years 1960, 1978 and 1998 are
given in table 2. Just like the percentile differences in panel (a), the mass around the mean
was rather stable up until the late seventies (1978), then it fell sharply over the eighties and the
nineties. Overall, these figures indicate that the twin–peaks phenomenon observed in figure 1
has arisen mainly since 1978. This pattern is confirmed in 3, where we plot the 1978 output per
worker distribution along with the 1960 distribution. The figure shows the interquartile range
had actually reduced over this period and there was not yet any hollowing out of the middle of
the distribution.13
13
It should be noted that a twin–peaks phenomenon is even more pronounced when we look at output per capita
rather that output per worker. However, the distribution of output per capita differs from that of output per
worker in that it shows widening not only in the middle of the distribution but also in the tails. Consumption–
per–capita presents a similar pattern, but with less bi–modality. (graphs are available from the authors upon
request)

7
 Figure 3: output per worker Distribution:1960–1978
0.5
1960
0.45 1978

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−3 −2 −1 0 1 2 3

2 A framework for decomposing changes in the cross-country

distribution of output per worker

The object of this section is to present a simple framework for assessing the importance of
different factors in explaining the observed changes in the world distribution of output per
worker. Our strategy, which is quite standard14 , is to provide a decomposition of the changes in
the distribution that emphasizes three main factors: (i) changes in the distributions of observable
factors affecting growth, (ii) changes in the importance of these observable factors, and (iii)
changes due to residual or unobserved factors. To this end, let us start with the identity given
in equation (1), whereby a country’s log–level of output per worker in 1998, denoted y i98 is
expressed as the log–level of output in 1960 (yi60 ) plus the sum of its growth rate between
1960–1978 (G60−78
i ) and its growth rate between (G78−98
i ). We explicitly choose to decompose
the growth rate over the entire period into growth between 1960–1978 and 1978–1998 based on
our discussion in the previous section suggesting 1978 as the date after which the twin peak
phenomena started to emerge.

yi98 = yi60 + G60−78

i + G78−98
i = yi60 + 18 × gi60−78 + 20 × gi78−98 (1)

In equation (1), gij is the annual rate of growth of output in country i over the sample period
j={60–78,78–98}. The growth rates for each of the sub-periods can then be expressed, as in the
standard growth regression literature, as the sum of a force of convergence related to initial levels
14
Our approach is closely related to that used in the labour literature for decomposing changes in the distribution
of individual outcomes such as wages. For an example see Juhn et al. [1993]. However, because of the dynamic
aspect of our problem, certain new issues arise.

8
of income–per–worker plus the effects of a set of country specific variables X i which potentially
explain growth.

gi60−78 = β060−78 + βy60−78 yi60 + Xi60−78 βx60−78 + ε60−78

i (2)
gi78−98 = β078−98 + βy78−98 yi78 + Xi78−98 βx78−98 + ε78−98
i (3)

In the above two equations, εji denotes the component of growth that is accounted for by
unobservable forces, with the distribution of these residuals being allowed to change over time.

In such a framework, changes in the distribution of y 98 relative to y 60 come through the two
growth rate equations. Given the reference date of 1978, an interesting question to pose is
what changed in the period 1978-1998 versus 1960-1978 to bring about a distribution which
is bi–modal and which has a wider inter-quartile range? This change between period can be
allocated to four sources: (i) changes in the distribution of the driving variables (i.e. changes
in the distribution of the X’s and y0 between 1960-78 versus 1978-98), (ii) changes in the
magnitudes of the β’s, (iii) changes in the distribution of unobservable forces and (iv) a residual
category which corresponds to that continuation of the (minor) dynamic process observed over
the 1960–1978 period. In light of this decomposition, we can easily build counterfactual income
distributions for 1998 which control for one or more of these different effects.

As an example of this approach, consider evaluating the effects due to changes in the distribution
of the X’s. To do this, we can construct a counterfactual growth rate for the second period, say
giX , which uses the distribution of X’s observed over the first period (1960–78) to replace the
actual X’s observed in the second period. This counterfactual growth rate is then given by

giX = β078−98 + βy78−98 yi78 + Xi60−78 βx78−98 + ε78−98

i (4)

Using the definition of the rate of growth, we can build a counterfactual distribution of y i98 ,
denoted yiX as follows

yiX = yi78 + 20 × giX

¡ ¢
= yi78 + 20 × β078−98 + βy78−98 yi78 + Xi60−78 βx78−98 + ε78−98
i
¡ ¢
= yi78 + 20 × β078−98 + βy78−98 yi78 + Xi78−98 βx78−98 + ε78−98
i + 20 × (Xi60−78 − Xi78−98 )βx78−98
= yi98 + 20 × (Xi60−78 − Xi78−98 )βx78−98 (5)

Likewise, the effects of changes in the magnitude of the βx on the distribution of income can be
measured by constructing the rate of growth, giβx , that controls for changes in βx

giβx = β078−98 + βy78−98 yi78 + Xi78−98 βx60−78 + ε78−98

i (6)

similar calculations as in system (5) yields

yiβx = yi98 + 20 × Xi78−98 (βx60−78 − βx78−98 ) (7)

9
In the same fashion, changes in either the distribution of the y0 ’s, the speed of convergence or
in the distribution of unobservables can be accounted for by computing the appropriate rates of
β
growth (giy0 , gi y and giε respectively) and the implied counterfactual income distributions (yiy0 ,
β
yi y and yiε ) as follows:1516

giy0 = β078−98 + β178−98 yi60 + Xi78−98 βx78−98 + ε78−98

i (9)
β
gi y = β078−98 + βy60−78 yi78 + Xi78−98 βx78−98 + ε78−98
i (10)
giε = β078−98 + βy78−98 yi78 + Xi78−98 βx78−98 + εi60−78 (11)
yiy0 = yi98 + 20 × (yi60 − yi78 )βy78−98 (12)
β
yi y = yi98 + 20 × (βy60−78 − βy78−98 )yi78 (13)
yiε = yi98 + 20 × (εi60−78 − ε78−98
i ) (14)

Each of the counterfactual y distributions constructed according to these equations can be used
to give an account for what the distribution of income would have been (in a partial equilibrium
sense) if one of the components had remained constant over the two periods. It should be
noted that — because of the dynamic aspect of our problem — even if we simultaneously take
into account the effects of changes in the distribution of Xs, changes in the βs and changes
in the distribution of the ε, the resulting counterfactual will not mechanically reproduce the
distribution of either 1960 or 1978. Instead, the distribution that arises when we consider all
three of these effects simultaneously, which we will denote as the distribution of y R , corresponds
to an outcome resulting from projecting to 1998 the dynamic process observed between 1960
and 1978, that is, yiR = yi78 + 20 × gi60−78 . Hence, when we evaluate our counterfactuals, it is
interesting to compare them with the 1960 and 1978 distribution as well as the distribution of
y R .17

Obviously, in order to implement the above decomposition, it is necessary to choose an appropri-

ate set of Xs. Our approach to this issue will be to focus on the forces emphasized in neoclassical
growth theory (see Solow [1956]) as well as other possibilities such as effects of education, scale
effects and non–linear dependence to initial conditions.
15
Note that if we want to account for both changes in the distribution and the coefficients at the same time,
we do not simply have to add up terms, as there exists an interaction between changes in the coefficients and
changes in the distribution. For instance, a measure of the effect of an overall change in the driving forces of the
growth process (other than convergence) is obtained by building the variable

yiXβ = yi98 + 20 × (Xi60−78 βx60−78 − Xi78−98 β 78−98 ) (8)

16
Our approach to decomposing a distribution implicitly assumes that the data satisfy the requirement of a pure
location model, that is, a mean shift in an X variable is assumed to shift the distribution but not change its shape.
Alternatively, we could adopt a less restrictive framework as made possible by quantile regression techniques.
However, a preliminary exploration along this front did not reveal any additional insights and therefore we did
not pursue this line.
17
If we wanted to follow the labor literature more closely, it would be reasonable to build all our counterfactuals
starting from the distribution of y R . However, we have chosen not to do so since it render our analysis less
transparent. Nonetheless, we have explored this alternative base for building counterfactuals and it produces
identical insights.

10
3 Understanding the growth of output per worker

In order to implement the above decomposition, the first step is to examine the stability of the
growth process over the period 1960-78 versus 1978-98. In particular, if the process is stable,
then changes in the βs cannot account for any of the change in the distribution. To this end,
in this section we report a set of cross–country regressions relating growth in output per worker
to the initial level of output per worker and other X variables for both the period 1960-78 and
1978-1998. As discussed above, we begin by focusing on the case with only two X variables ( (i)
the investment rate and (ii) the rate of population growth) since this will allow us to highlight
what we believe to be the major change in the growth process over this period. 18 Afterwards,
we will show how the change in importance of these two factors is robust to the inclusion of
other variables, and especially to controlling for human capital accumulation in different forms.

3.1 The rise in importance of accumulation forces

Table 3 reports a series of estimates for the regression of the average yearly growth in output per
worker on the initial log-level of output per worker, the average yearly growth in labor force and
the log of the average rate of investment. Estimates are reported for both the periods 1960-78
and 1978-98, and are estimated using ordinary least squares (OLS) and weighted least squares
(WLS). Averages for both the regressors and regressants are taken over the respective sub–
samples. In the case of WLS, two weighting schemes were used; the first one (WLS1) weights
each country according to the logarithm of its average population over the whole period, and
the second one (WLS2) uses the logarithm of average GDP. (Averages for both the regressors
and regressants are taken over the respective sub–periods).

Let us first focus on the OLS results. The convergence parameter — the coefficient associated to
y0 — is close to that reported by other studies (see e.g. Mankiw, Romer and Weil [1992]). More-
over, the estimates of this effect have remained quite stable over the two sub–periods. Indeed,
the stability test of this parameter, which we denote by Q(y0 ) in the table, does not indicate
a rejection of stability at conventional levels (p–value=0.855).19 In contrast, the estimates re-
ported in the two first columns of table 3 indicate that the magnitude of the coefficients related
to the rate of growth of the labor force and the investment/output ratio have increased quite
substantially between the two sub–periods. In particular, in the first sub-period (1960-1978), the
coefficient measuring the impact of population growth on the growth process was not significant
at the conventional 5% level and amounted to -0.137, implying that a negative 1% differential
in population growth would have translated into a positive differential in output per worker of
about 2.5% over the 18 year period. In the second sub–period, this coefficient rose to -0.606,
implying the same negative 1% differential in the rate of growth would have translated in a pos-
18
The data used are those described in section 1 and in appendix A.
19
Note that all our stability tests are performed allowing for residuals to be correlated within countries over
the two samples and the variances in the two samples are allowed to differ.

11
 Table 3: Growth Regressions

OLS WLS1 WLS2

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.185 0.208 0.161 0.213 0.180 0.211
( 0.025) ( 0.026) ( 0.023) ( 0.021) ( 0.024) ( 0.025)
y0 -0.013 -0.014 -0.010 -0.015 -0.013 -0.014
( 0.002) ( 0.002) ( 0.002) ( 0.002) ( 0.002) ( 0.002)
n -0.137 -0.606 -0.118 -0.623 -0.133 -0.618
( 0.164) ( 0.163) ( 0.173) ( 0.161) ( 0.162) ( 0.161)
i/y 0.016 0.027 0.018 0.025 0.017 0.026
( 0.003) ( 0.004) ( 0.003) ( 0.004) ( 0.003) ( 0.004)
R2 0.36 0.52 0.43 0.62 0.40 0.56
Q(Total) 14.642 [0.002] 8.590 [0.035] 10.994 [0.012]
Q(y0 ) 0.033 [0.855] 2.692 [0.101] 0.424 [0.515]
Q(n, i/y) 12.021 [0.002] 8.212 [0.016] 10.263 [0.006]

Note: Standard errors in parenthesis, p–values in brackets. WLS1: Weights

correspond to the log of the country’s population. WLS2: Weights correspond
to the log of the country’s total real GDP. 75 observations.

itive differential in output per worker of about 12% over the second sub–period. A somewhat
similar pattern is found in the behavior of the coefficient affecting the investment/output ratio.
Over the first sub–period, the coefficient affecting investment is 0.016 and significant. Over
the second sub–period, this coefficient rose to 0.027. The test statistic, Q(n, i/y), which tests
the null hypothesis of stability in these two parameters across sub-periods, decisively rejects
stability (p–value=0.002). The restriction of the overall stability of all three coefficients (not
including the constant), which is given by Q(Total), is also clearly rejected, which thereby opens
the door to the possibility that changes in coefficients affecting the growth process is a potential
candidate for understanding changes in the distribution documented in Section 2.

At this point, it is relevant to ask whether OLS estimates may be dominated by the experiences
of a subset of small countries. In order to address this issue, we report in table 3 results
from estimation using weighted least squares (WLS) where countries are respectively weighted
according to either their total population (log) or their total GDP (log). The first striking feature
that emerges from the table is that the OLS results remain largely unchanged by moving toward
WLS estimation, both in terms of magnitude and stability of the coefficients. For example, the
size of the effects of population growth on output growth process is very close to that obtained
using OLS. The coefficient is not significant over the first sub–period and becomes significantly
negative and large over the second sub–period. Likewise the effects of the investment rate almost
doubled between the two sub–periods and the estimates are very close to those obtained by OLS.
Furthermore, the stability test for these two variables also leads to strong rejection. The only
noticeable exception is found in the stability test of convergence in WLS1, which now indicates
that the convergence process may have increased over the two sub–periods. This is not too

12
surprising given that this procedure weights the Chinese and Indian experience more heavily,
both of which caught up rapidly over this period.

Table 4: Growth Regressions (Instrumental Variables)

IV1 IV2 IV3

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.147 0.203 0.155 0.202 0.149 0.204
( 0.027) ( 0.029) ( 0.035) ( 0.032) ( 0.027) ( 0.028)
y0 -0.010 -0.014 -0.011 -0.013 -0.011 -0.014
( 0.002) ( 0.002) ( 0.003) ( 0.003) ( 0.002) ( 0.002)
n -0.181 -0.614 -0.172 -0.616 -0.203 -0.638
( 0.171) ( 0.164) ( 0.170) ( 0.166) ( 0.172) ( 0.166)
i/y 0.009 0.026 0.010 0.025 0.009 0.026
( 0.004) ( 0.005) ( 0.006) ( 0.006) ( 0.004) ( 0.004)
Q(Total) 16.444 [0.001] 11.813 [0.008] 16.988 [0.001]
Q(y0 ) 0.923 [0.337] 0.439 [0.508] 0.911 [0.340]
Q(n, i/y) 14.936 [0.001] 9.841 [0.007] 15.406 [0.000]

Note: Standard errors in parenthesis, p–values in brackets. IV1: the variable

used for instrumenting i/y is its initial level in each sample period, IV2: the
variable used for instrumenting i/y is the average of (c/y) over each the sub–
period, IV3: the variables used to instrument i/y and n are the initial level of
(i/y), the average (c/y) over the sub–sample, and the average growth rate of
population over the 15 first periods of the sub–sample. 75 observations.

A second potential concern is the possibility of endogeneity of regressors. To address this issue,
we report in table 4 estimation results using different instrumental variable strategies. Three sets
of instruments are considered. The first two, IV1 and IV2, address the potential endogeneity of
the investment rate with respect to good growth outcomes in each sub-period. In the first case,
IV1, we use as an instrument for the average investment rate its initial (log) level in each sub–
period. In the second case, IV2, we instrument the investment rate using the country’s average
saving rate over each period, given by the average level of the consumption to output ratio (in log
form) over the sub–period. In the last case, IV3, we use both previously mentioned instruments
to instrument the investment rate, moreover we also instrument the rate of population growth
using the average growth rate of population 15–64 over only the 15 first years of each sub–period.
This latter instrument is aimed at correcting for the potential endogeneity of fertility over each
sub–period. As it takes at least 15 years for fertility to affect the labor force, this instrumental
variable strategy eliminates any possibility of within period endogeneity of fertility. Hereafter,
this particular set of instruments will be referred to as IV? . As can be seen from the table, the
pattern of results are barely affected by the IV procedures. Convergence still displays stability
and the coefficients are extremely close to those obtained using OLS. Furthermore, the stability
tests indicate that the impact of investment rates and population growth changed significantly
between the two sub–periods, as the p–value associated with Q(n, i/y) is essentially 0 for all

13
 three IV procedures. The only substantial change associated with these IV procedures is a
reduction in the coefficient on the investment–output ratio in the 1960–78 period (from 0.016
with OLS to 0.009 with IV). This implies a larger increase in the impact of the investment rate
across periods using IV.

Up to now, we have split our sample in 1978 based on our investigation of the dynamics of the
distribution of output per worker, and in particular this choice was motivated by the observa-
tion that the widening of the interquartile range seemed to have occurred in the late seventies.
Nevertheless, one may question whether or results are robust to changes in the splitting date.
To address this issue, we estimated the two growth processes using alternative split dates. In
particular, we explored all split dates between 1975 and 1983 and in all cases found significant
evidence of a change in the latter sample relative to the earlier sample. In order to get a sense
of change in the growth process taking place over time, we report in figure 4 the estimated coef-
ficients on population growth and investment rate as well as the speed of convergence obtained
from rolling regressions, with each variable measured in relation to the given window. We chose
a size of window of 19 years since it corresponds to half of our sample length. 20 Each rolling

Figure 4: Rolling regressions

Population Investment ratio Convergence
0 0.035 −0.005

0.03
−0.2
0.025 −0.01

−0.4 0.02

0.015 −0.015
−0.6 OLS
WLS 0.01
IV
−0.8 0.005 −0.02
1960 1965 1970 1975 1980 1960 1965 1970 1975 1980 1960 1965 1970 1975 1980
Years Years Years

regression was carried out using OLS, WLS and IV. When using WLS, the weight was given
by the (log) level of population. In the IV procedure, we used the set of instruments IV ? . As
can be seen from the figure, the main pattern highlighted in our simple regressions can also be
observed in rolling regressions. In effect, there has been a continuous increase in the magnitude
of the population growth effect as well as in the investment effect, even though the latter effect
decreases slightly at the end of the period.21 Furthermore, the size of the coefficients over the
20
In rolling regressions, we estimate a series of regressions with the first regression using observations 1 through
19, the second regression using observations 2 through 20, etc. . . We also explored window sizes between 15 to 25
years, all of which led to similar conclusions. In Beaudry and Collard (2002) this approach was used to examine
changes in the growth process among OECD country’s.
21
In effect, this later decline suggests that the period of especially high gains to physical capital accumulation

14
different windows is relatively robust to the estimation method. Hence, these results suggest that
the main pattern we observed in our simple OLS regression is robust both across the estimation
method and the splitting year of the sample.

An alternative means of examining the role of accumulation forces, one which is commonly used
in the literature, is to gather the effects of the investment rate and of population growth into
one summary variable which captures a country’s long-run capital intensity. In effect, from the
standard law of motion of capital, one can readily derive a country’s long-run capital-output
ratio. The log of this ratio, which we will denote by K, is given by the following relationship
where δ is the rate of depreciation of capital and γ is the rate of growth of labor augmenting
technological progress.
µ ¶ Ã i
!
k y
K ≡ log = log
y (1 + n)(1 + γ) − (1 − δ)

To construct K we followed Mankiw et al. [1992] and set the annual depreciation rate of capital at
3% (δ = 0.03) and the rate of technological growth of 2% (γ = 0.02). We checked the robustness
of our results against alternative rates of depreciation and found no major differences.

In the first two columns of table 5, we report IV estimates using K as regressor. The instruments
used are those corresponding to the set IV? . Two main comments are in order. First, the
convergence parameter shares the same properties as those presented earlier. If we now focus on
our accumulation variable K, it shares the properties of our previous estimates of the investment
rate and population growth. In particular, its size has more than tripled between the two sub–
samples shifting from 0.009 to 0.029, and the stability test strongly indicates rejection.

By reducing the number of covariates in this way, we make it easier to consider results within
sub-samples. In particular, we can now explore whether the observed instability of coefficients
is being driven by primarily by rich or poor countries. To this end, we report in table 5 results
based on splitting the sample of countries along the median level of incomes in 1960. 22 As can
be seen in Table 5, among the richer countries — those with an initial output per worker above
the median of the distribution — convergence has actually declined slightly between periods,
although not in a significant manner. Conversely, the coefficient on the accumulation forces is
significantly different across the two sub–periods, with a change of the same order as the whole
sample. If we now look at the poorest countries, we basically observe two phenomena. The first
one is related to the convergence process. Convergence seems to have risen slightly between the
two sub–samples, from no significant convergence between 1960 and 1978 to a convergence rate
of -0.012 over the 1978–1998 period. More interestingly, a look at the accumulation parameter
reveals that there has been a significant change in the importance of accumulation forces of a
magnitude similar or higher than that observed in the whole sample. Therefore, the observed
increase in the importance of traditional accumulation forces appears to have occurred rather
may have been a temporary phenomena that was concentrated in the 1978-96 period.
22
The set of instrument is again IV? .

15
 Table 5: Growth Regressions (Accumulation, IV)

Total y1960 <med(y1960 ) y1960 >med(y1960 )

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.121 0.121 0.060 0.114 0.176 0.075
( 0.021) ( 0.022) ( 0.046) ( 0.037) ( 0.039) ( 0.075)
y0 -0.011 -0.014 -0.003 -0.012 -0.017 -0.010
( 0.002) ( 0.002) ( 0.006) ( 0.004) ( 0.004) ( 0.008)
K 0.009 0.029 0.008 0.026 0.012 0.034
( 0.003) ( 0.004) ( 0.005) ( 0.005) ( 0.004) ( 0.008)
Wald 0.154 1.981 0.483 0.124 7.190 3.286
[0.694] [0.159] [0.487] [0.724] [0.007] [0.069]
Q(Total) 17.846 [0.000] 6.808 [0.033] 23.874 [0.000]
Q(y0 ) 0.834 [0.361] 1.929 [0.165] 0.648 [0.421]
Q(K) 15.647 [0.000] 6.561 [0.010] 7.580 [0.006]

Note: Standard errors in parenthesis, p–values in brackets. The variables used to instrument
K are the initial level of (i/y), the average (c/y) over the sub–sample, and the average
growth rate of population over the 15 first periods of the sub–sample. 75 observations.
The line denote Wald reports the Wald test of the restriction implicit in the accumulation
variable K that the coefficient of (i/y) is the same as that of log((1 + n)(1 + γ) − (1 − δ)).

uniformly across rich and poor countries.

Together, the results reported in tables 3 through 5 suggest that there may have been a major
change in the determinants of growth across NSSA countries, whereby the importance of the two
traditional growth factors emphasized in Solow [1956] — (i) the rate of investment in physical
capital (or the savings rate) and (ii) population growth — having increased by a factor of almost
3 between the pre and post 1978 periods. In contrast, we find that the speed of convergence
has remained rather stable over the period. Before turning to examining the relevance of these
effects in terms of their role in changing distribution of output per worker, we first need to
establish their robustness.

3.2 Beyond Capital Accumulation: Checking Robustness

In this section we report regression results that explore the robustness of our previous observation
regarding the apparent increased importance of physical capital accumulation over the period
1978-98 versus 1960-78 period. More precisely, we first consider the effects of a possible mis–
specification error in the convergence process stemming from the omission of non–linear terms.
We then address in substantial detail the important question of the role of education.

Nonlinear Convergence: We now address a potential source of bias that may explain our
previous results: a mis–specification error underlying the overall convergence process. Implicit
in our growth regression is the assumption that the transition dynamics are linear. However, it

16
may be the case that basins of divergence and convergence have emerged over this period.

Table 6: Non–linear convergence

OLS IV
60–78 78–98 60–78 78–98
Const. 0.044 0.059 0.033 0.058
( 0.007) ( 0.007) ( 0.008) ( 0.008)
n -0.227 -0.581 -0.291 -0.616
( 0.167) ( 0.178) ( 0.176) ( 0.181)
i/y 0.016 0.027 0.008 0.025
( 0.003) ( 0.004) ( 0.004) ( 0.004)
y0 -0.022 -0.014 -0.023 -0.014
( 0.015) ( 0.015) ( 0.016) ( 0.015)
y02 -0.001 -0.001 -0.004 -0.002
( 0.012) ( 0.012) ( 0.013) ( 0.012)
y03 0.001 -0.001 0.000 -0.001
( 0.003) ( 0.003) ( 0.003) ( 0.003)
R2 0.39 0.52 – –
Q(Total) 19.233 [0.002] 23.191 [0.000]
Q({y0j }3j=1 ) 3.084 [0.379] 4.310 [0.230]
Q (n, i/y) 9.239 [0.010] 13.595 [0.001]

Note: Standard errors in parenthesis, p–values in brackets. WLS: Weights are

given by the log of the population. IV: the variables used as instruments for
i/y and n are the initial level of (i/y), the average (c/y) over the sub–sample,
and the average growth rate of population over the 15 first periods of the sub–
sample. 75 observations.

One simple way of addressing this possibility is to introduce higher order terms in the initial level
of output in our growth equation. This allows the growth process to be different for countries
starting from different initial levels. We introduce square and cubic terms in the initial (log) level
of output to explore the relevance of this effect.23 In parallel to our previous results, we report
in table 6 estimates using OLS, and IV. The set of instruments is IV? . As can be seen from the
table, non–linear terms are not found to be significant whatever the sub–period and whatever
the estimation method. Accordingly, convergence does not exhibit any significant instability
between the two sub–periods, while the accumulation forces continue to exhibit this instability.
Hence, nonlinear convergence forces do not appear particularly relevant in our data. 24

Education: Up to now, we have focused on whether the importance of physical capital accu-
mulation may have changed over time. However, education, and the broader concept of human
23
Note that for presentation purposes, the initial level of output is normalized by the 1960 US level for this
exercise.
24
In the absence of the investment rate and the rate of population growth in the regression, we do find some
minor evidence of non–linearity.

17
capital, is thought by most to be important contributing factor to growth. For instance, in the
theory of endogenous growth, sustained growth is often the result of the accumulation of human
capital over time (see Uzawa [1965] or Lucas [1988]). Furthermore, in their influential paper,
Mankiw et al. [1992] propose an augmented version of the Solow growth model in which human
capital — measured essentially by the proportion of the working age population in secondary
school — enters the production function. Their results indicate that including both type of cap-
ital — physical and human — enables the neoclassical growth model to fit the data relatively
better. Such a finding opened the route to a large number of studies that deliver more am-
biguous conclusions concerning the exact role of education in the growth process (see Durlauf
and Quah [1999] or Klenow and Rodrı̀guez-Clare [1997] for discussions). Therefore, without
wanting to take a strong position on the overall importance of education in the growth process,
we nevertheless want to examine (i) whether the effects of education have changed over time
and (ii) whether controlling for educational investment (or human capital) affects our results
regarding the change in importance of traditional growth factors. To this end, tables 7 through
9 report a set of regression where we introduced different measures of educational investment
or human capital into our growth equations.25 Our first measure of educational investment is a
weighted average of school enrollment for primary, secondary and tertiary degrees of education,
as documented in Barro and Lee [1993].26 This measure departs from the original Mankiw et al.
[1992] specification of educational investment and rather follows Klenow and Rodrı̀guez-Clare
[1997] who argue that focusing on this wider education class is more appropriate. In table 7
we report estimates for both OLS and IV estimation. Two sets of instruments are investigated.
The first one consists of our IV? set and is used to instrument both the investment rate and the
rate of population growth. The results from this exercise are given under the heading IV1 in
table 7. The second set augments IV? , as to allow us to instrument our educational investment
variable, by including the number of years of schooling at the beginning of the period. The
latter is aimed at mitigating potential measurement error in the school enrollment rate data.

Two main conclusions can be drawn from table 7. First of all, the change in importance of
the traditional growth factors over the two samples is not at all diminished when we include
educational investment into our equation. In effect, rather than diminishing the change in
coefficients of the investment rate and on population growth, the introduction of education
actually magnifies it slightly. Second, in none of the three cases do we find the importance
of educational investment to have increased over the two sub–samples. In fact, the estimates
suggest that there may have been a decrease. Although the results in this table do not support a
25
The education data are discussed more extensively in appendix A. Note that because of data availability
problems, we loose 7 countries from our sample.
26
The weights for calculating this measure of educational investment correspond to the average number of years
spent in each degree µ ¶
6 × EP + 6 × ES + 4 × ET
Educ ≡ log
16
where E P , E S and E T denote respectively the enrollment rate in the primary, secondary and tertiary sector.

18
 Table 7: Education (I)

OLS IV1 IV2

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.218 0.216 0.197 0.218 0.197 0.249
( 0.032) ( 0.036) ( 0.033) ( 0.036) ( 0.036) ( 0.046)
y0 -0.017 -0.014 -0.015 -0.014 -0.015 -0.017
( 0.003) ( 0.003) ( 0.003) ( 0.003) ( 0.003) ( 0.004)
n -0.082 -0.710 -0.117 -0.736 -0.117 -0.709
( 0.164) ( 0.159) ( 0.171) ( 0.161) ( 0.171) ( 0.164)
i/y 0.014 0.029 0.006 0.030 0.006 0.028
( 0.003) ( 0.004) ( 0.004) ( 0.005) ( 0.004) ( 0.005)
Educ. 0.010 -0.002 0.015 -0.004 0.015 0.008
( 0.006) ( 0.009) ( 0.006) ( 0.009) ( 0.008) ( 0.014)
Q(Total) 24.275 [0.000] 28.970 [0.000] 28.844 [0.000]
Q(y0 ) 0.342 [0.559] 0.053 [0.818] 0.152 [0.697]
Q(n, i/y) 18.934 [0.000] 24.660 [0.000] 19.879 [0.000]
Q(Educ.) 1.488 [0.223] 3.100 [0.078] 0.220 [0.639]

Note: Standard errors in parenthesis, p–values in brackets. IV1: the variables

used to instrument i/y and n are the initial level of (i/y), the average (c/y) over
the sub–sample, and the average growth rate of population over the 15 first
periods of the sub–sample. IV2: the set of instruments of IV1 is completed by
the years of schooling to also instrument the education variable. 68 observations.

strong effect of education investment on growth — at least not in the frequency of two decades 27
— this may have arisen because of measurement error. In effect, as we see in the last column,
the negative effect of educational investment observed in the 1978-98 period in the first columns
disappears when we instrument educational investment with the number of years of education.

To explore the appropriateness of our measure of educational investment, in table 8 we report

a set of regression using two alternative measures to control for educational differences. The
first one, P&S, still refers to educational investment as it is a weighted average of primary and
secondary enrollment rates, where the weights are the duration (DURP and DURS) in years
of each of these two education stages in the different countries as provided in Barro and Lee
[1993].28 The second variable, Years, is a direct measure of the stock of human capital per capita
as it corresponds to the average schooling in the total population over 25 also from Barro and
Lee [1993]. The results in table 8 are essentially identical to those found in table 7, that is,
27
This result is in the lines of that obtained by Benhabib and Spiegel [1994] who also find a insignificant and
wrong signed effect of education investment on GDP growth. However it contrasts with the findings of Krueger
and Lindhal [2001] who point to a positive effect of education on output growth. It should nevertheless be noted
that the strong positive effect of education on growth obtained in Krueger and Lindhal [2001] only holds when
they abstract from physical capital (see table 1 p. 1112 in Krueger and Lindhal [2001]). As soon as physical
capital is brought back into the regression, education changes are not significant.
28
The measure is µ ¶
DU RP × E P + DU RS × E S
P&S ≡ log
DU RP + DU RS

19
the importance of traditional growth factors appear to have increased while the importance of
education has stayed the same or possibly decreased.

Table 8: Education II

P&S (IV1) P&S (IV2) Years (IV1)

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.192 0.218 0.191 0.245 0.174 0.234
( 0.032) ( 0.034) ( 0.040) ( 0.040) ( 0.031) ( 0.031)
y0 -0.015 -0.014 -0.015 -0.017 -0.014 -0.016
( 0.003) ( 0.003) ( 0.004) ( 0.004) ( 0.003) ( 0.003)
n -0.146 -0.732 -0.146 -0.721 -0.129 -0.710
( 0.170) ( 0.159) ( 0.170) ( 0.161) ( 0.171) ( 0.161)
i/y 0.007 0.030 0.007 0.028 0.008 0.029
( 0.004) ( 0.005) ( 0.004) ( 0.005) ( 0.004) ( 0.004)
Educ. 0.015 -0.005 0.015 0.009 0.006 0.002
( 0.006) ( 0.010) ( 0.008) ( 0.015) ( 0.003) ( 0.004)
Q(Total) 28.554 [0.000] 28.489 [0.000] 28.072 [0.000]
Q(y0 ) 0.041 [0.839] 0.232 [0.630] 0.241 [0.624]
Q(n, i/y) 24.182 [0.000] 18.653 [0.000] 22.829 [0.000]
Q(Educ.) 3.347 [0.067] 0.112 [0.737] 0.593 [0.441]

Note: Standard errors in parenthesis, p–values in brackets. All regressions are estimated
by instrumental variables. IV1: the variables used to instrument i/y and n instruments
are the initial level of (i/y), the average (c/y) over the sub–sample, and the average growth
rate of population over the 15 first periods of the sub–sample. IV2: the set of instruments
are those of IV1 completed by the years of schooling to instrument the education variable.
68 observations.

As an alternative way of examining the role of education, we construct a long run human capital
intensity variable as we did for physical capital. This variable, which we denote by H, is built
using educational investment rate data (E) as follows.29
µ ¶
E
H ≡ log
(1 + n)(1 + γ) − (1 − δ)

As for the physical capital accumulation variable, we set γ + δ = 0.05. Table 9 reports a set of
OLS and IV regressions using both K and H as regressors. The sets of instruments are those
associated with IV? and then augmented in include the average years of education to instrument
H. Once again we find that the importance of education has not increased while the importance
of physical investment appears to have increase by at least an order of 3.

One last way of assessing the role of education in the growth process is two split the sample
of countries in two depending on the average level of education in a country in 1960. Then by
running regressions for each of the sub-samples, we can gauge whether the apparent effects of the
increased importance of capital accumulation may be limited to higher educated countries. Table
29
The measure of educational investment we use is the same weighted average of enrolment rates as in table 8.

20
 Table 9: Education (III)

OLS IV1 IV2

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.149 0.128 0.131 0.130 0.135 0.142
( 0.022) ( 0.023) ( 0.023) ( 0.023) ( 0.024) ( 0.024)
y0 -0.016 -0.016 -0.014 -0.016 -0.014 -0.019
( 0.003) ( 0.003) ( 0.003) ( 0.003) ( 0.003) ( 0.004)
K 0.013 0.030 0.006 0.031 0.006 0.028
( 0.004) ( 0.004) ( 0.004) ( 0.005) ( 0.004) ( 0.005)
H 0.007 0.005 0.013 0.003 0.012 0.016
( 0.006) ( 0.008) ( 0.006) ( 0.008) ( 0.007) ( 0.010)
R2 0.36 0.56 – – – –
Q(Total) 19.549 [0.000] 25.378 [0.000] 28.605 [0.000]
Q(y0 ) 0.000 [0.991] 0.087 [0.768] 1.217 [0.260]
Q(K) 10.320 [0.001] 16.824 [0.000] 12.412 [0.000]
Q(H) 0.062 [0.803] 1.050 [0.305] 0.133 [0.715]

Note: Standard errors in parenthesis, p–values in brackets. IV1: the variables

used in instrument K are the initial level of (i/y), the average (c/y) over the
sub–sample, and the average growth rate of population over the 15 first periods
of the sub–sample. IV2: the set of instruments is that of IV1 completed by the
years of schooling to instrument H. 68 observations.

10 reports the results for such an experiment. We re-adopt the parsimonious specification that
only involves the initial level of (log) output per worker and our physical capital accumulation
variable, K, for this exercise since the degrees of freedom are more limited. We split the sample
into two groups of countries: those for which the number of years of schooling is higher than the
median of the distribution in 1960, and the rest of the countries. The estimation is carried out
using IV where the set of instrumental variables corresponds to IV? . Two main results stand
out when we look at the table. First, stability tests reveal that the convergence process has
remained significantly unchanged in highly educated countries (with a p–value of about 0.6)
and at a level comparable to that obtained in our earlier estimations. Conversely, low educated
countries have experienced a large and significant (p–value of 0.03) increase in the convergence
rate. Second and more interestingly, the importance of physical capital accumulation seems to
have undergone a significant change in both sets of countries, as stability tests indicate the null
of a stable coefficient is strongly rejected by the data. It is interesting to note that the change
appears to have been slightly greater for the lower educated countries.

As can be seen from all the preceding experiments, our main findings — (i) a steady convergence
and (ii) an increase in the importance of traditional forces of accumulation — are robust to many
specification changes and estimation methods. Accordingly, we will use the estimated growth
process from our benchmark OLS regression in table 3 in what follows to decompose the changes

21
 Table 10: Accumulation and Education (IV)

Total H1960 <med(H1960 ) H1960 >med(H1960 )

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.125 0.131 0.077 0.148 0.170 0.130
( 0.024) ( 0.023) ( 0.041) ( 0.037) ( 0.033) ( 0.035)
y0 -0.011 -0.015 -0.005 -0.017 -0.017 -0.015
( 0.003) ( 0.003) ( 0.005) ( 0.004) ( 0.004) ( 0.004)
K 0.010 0.033 0.005 0.032 0.018 0.030
( 0.004) ( 0.004) ( 0.005) ( 0.006) ( 0.005) ( 0.005)
Wald 0.000 2.934 0.178 0.876 0.285 1.718
[0.997] [0.086] [0.673] [0.349] [0.593] [0.189]
Q(Total) 26.816 [0.000] 16.186 [0.000] 5.339 [0.069]
Q(y0 ) 1.349 [0.245] 4.517 [0.034] 0.233 [0.629]
Q(K) 22.478 [0.000] 15.806 [0.000] 3.203 [0.074]

Note: Standard errors in parenthesis, p–values in brackets. All regression estimated by IV.
The variables used to instrument K are the initial level of (i/y), the average (c/y) over
the sub–sample, and the average growth rate of population over the 15 first periods of the
sub–sample. The lines denote Wald reports the Wald test of the restriction implicit in the
accumulation variable K that the coefficient infront of (i/y) is the same as that affecting
log((1 + n)(1 + γ) − (1 − δ)).

in the distribution of output per worker. Results related to using other estimates are presented
in the appendix.

4 Implementing the decomposition

In this section we decompose the observed changes in the distribution of output per worker over
the period 1960–78 using the framework presented in section 2. As suggested in the results of
the last section, part of the explanation for the observed changes in the shape of the distribution
over this period may be due to the observed change in the importance of the traditional forces
of the accumulation process — i.e. changes in the βs associated with the rate of investment
and the effect of population growth. Figure 5 reports the decomposition of changes in the
distribution for different potential effects using our benchmark OLS regression as our estimate
of the growth process (results using IV estimates are reported in the appendix, figure 9). Each
panel reports three pieces of informations. The grey line corresponds to the initial distribution
of output per worker as observed in 1960, the plain dark line corresponds to the distribution
of 1998 output per worker, and the dashed dark line is the distribution of the 1998 output per
worker “corrected” for the effect that we examine. More precisely, the last distribution is based
on each counterfactual experiment described in section 2. For example, when we look at the
effects of a change in the coefficients of population growth and the investment/output ratio,
the distribution we report is that obtained using equation (7). This counterfactual experiment

22
actually furnishes an answer to the question: what would the output per worker distribution
in 1998 have been had the coefficients in the growth process remained at their values from the
earlier period? This graphical analysis is complemented with a quantitative assessment of the
observed changes in tables 11 and 12 which report, respectively, the same decomposition in terms
of the (20–80, 25–75 and 30–70) percentile differences and the mass-around-the-mean statistic
at ±0.6, 0.7 and 0.8 log–points around the mean.

Figure 5: Decomposing the observed changes in the distributions

78−98
(a) Observables (y ) (b) Coefficients (βy =0)
0
0.5 0.5
1960
0.4 1998 0.4
Corr.
0.3 0.3

0.2 0.2

0.1 0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4
(c) Distribution (n,i/y) (d) Coefficients (n,i/y)
0.5 0.8

0.4
0.6
0.3
0.4
0.2
0.2
0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4

(e) Observables (n,i/y) (f) Unobservables

0.8 0.5

0.4
0.6
0.3
0.4
0.2
0.2
0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4

Panel (a) of figure 5 reports the decomposition obtained when we control for the effects due
to a change in convergence forces (this amounts to building the counterfactual 1998 output
β y 60
per worker yi y = yi98 + 20 × (yi60 βy60−78 − yi78 βy78−98 )). As clearly indicated by the figure,

23
 Table 11: Percentiles

perc. 98–60 98–78 98-y R βn,i/y ∆Xn,i/y Obs.n,i/y Obs.y0 ε

20–80 0.207 0.314 0.361 0.194 0.130 0.148 -0.049 0.145
25–75 0.265 0.376 0.373 0.264 0.035 0.305 0.065 0.224
30–70 0.197 0.216 0.320 0.370 -0.053 0.226 -0.035 0.060
Average 0.223 0.302 0.351 0.276 0.037 0.226 -0.006 0.143

Table 12: Mass around the mean

∆ 98–60 98–78 98-y R βn,i/y ∆Xn,i/y Obs.n,i/y Obs.y0 ε

0.60 -10.667 -14.667 -12.000 -6.667 -4.000 -4.000 0.000 -2.667
0.70 -12.000 -17.333 -14.667 -10.667 -4.000 -14.667 1.333 -6.667
0.80 -5.333 -13.333 -12.000 -16.000 1.333 -10.667 2.667 -4.000
Average -9.333 -15.111 -12.889 -11.111 -2.222 -9.778 1.333 -4.445

changes in the convergence forces do not explain much of the changes in the decomposition, as
the counterfactual distribution remains very close to the 1998 distribution. This should have
been expected from our preliminary econometric analysis and the results in terms of observed
distribution of output per worker. Indeed, the stability tests together with the rolling regressions
from the previous section all suggested that the convergence parameter remained steady across
the two sub–periods. In addition, the 1978 distribution output per worker was found to be close
to that for 1960. This graphical appraisal is confirmed by the quantitative results reported in
tables 11 and 12. For instance, the changes in the interquartile range between 1960 and 1998
that are accounted for by changes in convergence forces is 0.065, to be compared to the 0.265
in the data. Likewise, convergence does not account for changes in the mass lying between 0.6
and 0.8 away from the mean of the distribution. However, this does not mean that convergence
played no role in designing the shape of the distribution. In order to illustrate this, we report
in panel (b) of figure 5 the distribution what would have been obtained had convergence had
totally vanished over the second sub–period (βy78−98 = 0). As can be seen from the figure, had
convergence disappeared, the distribution would have widened substantially with most of the
phenomenon taking place in the tails of the distribution. Therefore, we are not arguing that the
role of convergence was irrelevant over this period, but instead that changes in this force does
not help to account for the twin–peaks phenomenon.

We now turn our attention to the accumulation forces which we found to be unstable between
the two periods. Let us first focus on the effects of the observed changes in the distribution of
the population rate of growth and the investment/output ratio. As can be seen from panel (c)
of figure 5, changes in the distribution of the accumulation forces do not explain much of the
changes in the distribution as the counterfactual distribution built using equation (5) is close to

24
the actual 1998 distribution. This result is unsurprising given figure 6 which reports the 1960-78
and 1978-98 distributions of the investment/output ratio and the population rate of growth. The
distribution of these two variables remained rather stable across the two sub–periods, with the
distribution of the investment rate becoming – if anything – more peaked in the second period.
This explains why we should not expect a large impact from this component. This graphical

Figure 6: Distribution of Accumulation Forces

Investment to output ratio Population growth
1 40
1960 1960
0.8 1978 1978
30
0.6
20
0.4
10
0.2

0 0
−4 −2 0 2 −0.04 −0.02 0 0.02 0.04

appraisal is confirmed by the inspection of columns ∆Xn,i/y of tables 11 and 12, which indicate
for instance that out of the 0.265 change in the interquartile range between 1960 and 1998 only
0.035 are accounted for by a change in the distribution of the traditional accumulation forces.

Hence if the explanation of the change in the distributions is to be found in the forces associated
with physical capital accumulation, it should be found in a variation in the coefficients. This
case is examined in panel (d) of figure 5 with the counterfactual experiment based on equation
(7). The figure shows that changes in the coefficients affecting population growth and the
investment/output ratio contribute a great deal to the phenomenon. The first feature that
jumps out of panel (d) is that correcting for the instability in the accumulation coefficients
totally undoes the twin–peaks phenomenon. In other words, the twin–peaks phenomenon does
not seem to be the by–product of a change in the convergence forces, but rather by a change
in importance of physical capital accumulation. Beyond the effects in terms of bi-modality,
changes in these coefficients clearly had an effect on the spread of the distribution. Setting
the coefficients to their 1960-78 values generates a more compressed counterfactual distribution,
indicating that the large change in these coefficients between periods contributed to the widening
of the distribution.

The qualitative appraisal based on the figures is confirmed by the results in column β n,i/y of
tables 11 and 12. For example, out of the 0.265 change in the interquartile range that took place
between 1960 and 1998, the change in the coefficient of the two accumulation forces accounts
for 0.264. The same pattern is found for the 20–80 percentile difference, as a change of 0.194
is predicted by the instability of the coefficients to be compared to the 0.207 found in the
data. There actually is tendency to over-explain the phenomenon when we look at the 30–70
percentile difference, as this coefficient effect is almost twice that observed in the data. Very

25
similar results are obtained when we look at the mass around the mean. For example, when
we deviate by 0.6 log-points from the mean, the change in the βn,i/y predicts a decrease in
mass of 11 percent between 1960 and 1998, while the mass in the corresponding window of the
data changed by 7 percentage points. Likewise, for a deviation of ±0.7, the instability in the
coefficients accounts for a drop of about 11 percentage points compared to the almost 12 found in
the actual distribution. In other words, the changes in the magnitude of the accumulation forces
appear to have contributed enormously to the observed changes in the shape of the distribution.

Overall changes in the accumulation forces, that is, effects due to both changes in the coefficients
and changes in the distributions of i/y and n, actually slightly over-explain the observed modifi-
cation in the shape of the distribution of output per worker. This is shown in Obs. n,i/y in tables
11 and 12 and in panel (e) of figure 5. However, when we examine results in terms of changes
between 1978 and 1998, the accumulation forces explain the observed change almost exactly,
with little over- explanation. For example out of the 0.376 observed change in the interquartile
range between 1978 and 1998, the changes due to the total of accumulation forces (both β and
the distribution of Xs) is 0.305. Furthermore, on average the changes in the percentile differ-
ences predicted by the variation of the accumulation forces was 0.312 to be compared to the
0.226 observed in the data for the period 1978–1998. Likewise, on average the mass around the
mean has dropped by 15 percentage points, out of which 10 are predicted by the changes in the
traditional growth factor.

Our point is not to deny any role to other potential factors in affecting growth, but rather to
emphasize the potentially very important role played by traditional growth factors in reshaping
the world distribution of output per worker since 1978. Our framework suggested, as seen in
panel (f) of figure 5, that omitted or residual factors did not play a very decisive role in this
process. Indeed, as can be seen from the figure, controlling for unobservable components leaves
the distribution of 1998 output per worker almost unchanged. Such a finding is confirmed
by the last columns of tables 11 and 12 which indicate that the contribution of unobservable
components to the explanation of the changed in the percentile differences or the mass-around-
the-mean was much lower than the contribution due to changes in the importance of population
growth and the investment rate. This observation is especially surprising given that, by adopting
a very parsimonious specification of the growth process, we could have expected the residual
component to play a large role.

Our approach in tables 11 and 12 has been to present the effects of different forces (βs, Xs,
unobservables. . . ) one at a time – altering only that component while leaving the others at
their 1978-98 values. This approach has the attractive property of treating each force symmetri-
cally. However, a more common means of reporting results of decompositions is to sequentially
combine effects such that the end results mechanically add up to explain the relevant change in
distribution. Thus, most decompositions exercises typically change first, say, X to its alternative
value, then X and β, etc. . . In table 13 and 14 we report results in a manner consistent with this

26
adding–up reasoning and in a manner that recognizes that such results are not independent of
the order of sequencing. Since our claim is that changes in the coefficients on population growth
and investment rate explain most of the change in distribution, we choose the following two
sequences to back this claim. First, we put the change in the coefficients on n and i/y first in
the ordering (ordering I). This effectively means that all the other forces — which we denote by
R — act as a residual, capturing whatever is left of the change between y 98 and y R (where y R
controls for all effects) after accounting for the effects of n and i/y. Note that the set of R forces
correspond to the sum of effects due to unobservable components, changes in convergence forces
and changes in the distributions of variables. Second, we implement the reverse ordering, where
we begin be assessing the role of all the R forces, and leave the effects of the coefficients on n
and i/y to be determined residually (ordering II). As can be seen in tables 13 and 14, regardless
of the ordering the result is quite clear: changes in the coefficients on n and i/y explain all (or
too much) of the hollowing–out process as captured by our reference percentile differences and
the mass–around–the–mean statistics.

Table 13: Percentiles II

Ordering I Ordering II
perc. 98-y R βn,i/y R R βn,i/y
20–80 0.361 0.194 0.167 -0.048 0.409
25–75 0.373 0.264 0.109 -0.082 0.455
30–70 0.320 0.370 -0.050 -0.098 0.418
Average 0.351 0.276 0.075 -0.076 0.427

Table 14: Mass around the mean

Ordering I Ordering II
∆ 98-y R βn,i/y R R βn,i/y
0.60 -12.000 -6.667 -5.333 0.000 -12.000
0.70 -14.667 -10.667 -4.000 -4.000 -10.667
0.80 -12.000 -16.000 4.000 -4.333 -7.667
Average -12.889 -11.111 -1.778 -2.777 -10.111

In light of the all preceding results, we infer that most of the observed changes in the distri-
bution of output per worker between 1960 (or 1978) and 1998 — in particular its widening
in the interquartile range area and the twin–peaks phenomenon — is related to an increase in
importance of traditional accumulation forces. More particularly, this change resulted from a
change in the form of the process (coefficients) rather than a modification in the distribution of
these forces.

27
5 Bringing back sub Saharan Africa

In the previous section, we focussed exclusively on NSSA countries. We now want to extend our
sample to include the sub-Saharan African countries and examine how this affects our interpre-
tation of the role of different observable and unobservable factors is explaining the changes in
the distribution.

5.1 Still the twins?

Figure 7 reports the empirical distribution of (log) output per worker in years 1960, 1978 and
1998 for all countries in the Penn World Tables 6.0 (106 countries) for which data is available. 30
We will refer to this larger set of countries as the “world”. As in section 1, the distributions

Figure 7: Distribution of output per worker

0.4
1960
1978
0.35 1998

0.3

0.25

0.2

0.15

0.1

0.05

0
−4 −3 −2 −1 0 1 2 3 4

are estimated using a Gaussian kernel density estimator and are centered around their mean.
Figure 7 also reports the points in the distribution associated with the interquartile ranges in
1960, 1978 and 1998.31 Several comments are in order. First and foremost, it appears that,
as in our restricted sample, the shape of the distribution has changed considerably over time.
As before, the distribution of output per worker is uni–modal and almost normal in 1960. As
time has passed, mass shifted away from the center, creating two new modes. While the shape
in 1998 is bi modal, it is also quite different relative to the NSSA sample. In effect, instead of
characterizing the 1998 distribution for this wider set of countries as twin–peaked, it may be
30
We include all countries with at most two years of missing data between 1960 and 1998.
31
The points of the interquartile range are calculated from the raw data and not from the kernel estimates of
the density function.

28
more appropriately called hump–shaped. Moreover, there is another major difference relative
to our earlier observation, which is that the hump–shape is already present by 1978 while the
twin peaks in the NSSA sample only emerged after 1978. In what follows, we will argue that
there were two distinct phenomena going on in the world distribution. First, in the 1960–1978
period, sub–Saharan African countries experienced particularly bad draws in residual factors —
that is factors not associated with either human or physical capital accumulation — and these
bad draws created the hump–shape observed in 1978. Secondly, after 1978, the growth process
changed in a manner that favoured countries which favoured physical capital deepening and this
again explains most of the change in distribution observed from 1978 to 1998 (without a need
to refer to particular bad draws for sub-Saharan Africa in the later period). In what follows, we
substantiate this claim.

5.2 Reexamining changes in the growth process

As in the previous section, our goal is to document the forces underlying the observed changes in
the world distribution of the output per worker over the period 1960–98. A preliminary step in
our argument is therefore to examine whether the instability in the growth process documented
for our restricted sample holds when sub–Saharan African countries are included. Table 15
reports a series of estimates for the growth process over the period 1960–78 versus 1978–98
using ordinary least squares (OLS), weighted least squares (WLS) and instrumental variables
(IV). When WLS are performed, each country is weighted according to the logarithm of its
average population. When an IV procedure is used, we instrument the investment rate and
population growth using the IV? set of instruments. Note that we also include a sub Saharan
Africa dummy in order to capture potential region specific effects. However, it is worth noting
that the exclusion of this dummy variables does not significantly affect the other estimates in
the table.

We focus first on the OLS results. Not surprisingly, the sub Saharan Africa dummy is signif-
icantly negative in both sub-periods. The convergence parameter — the coefficient associated
to y0 — is very close to that obtained in section 3.1. Furthermore, just like in our restricted
sample, the estimates of convergence remained rather stable over the two sub–periods, as the
stability test statistic (Q(y0 ) in the table) does not permit rejection of the stability hypothesis
at conventional significance levels (p–value=0.543). Once again, the main change that emerges
from the data relates to the traditional forces of growth. Indeed, the results reported in the
two first columns of table 15 suggest that the magnitude of the coefficients affecting the rate
of growth of the labour force and the investment/output ratio increased in a manner similar
to that observed in our restricted sample. The coefficient measuring the impact of population
growth on the growth process was not significant at the conventional 5% level and amounted
to -0.211 in the period 1960–1978, and became strongly significant and large (-0.772 to be com-
pared to -0.606 in our restricted sample) in the period 1978–1998. A very similar pattern can
be found analyzing the behavior of the coefficient affecting the investment/output ratio. Over

29
 Table 15: Growth Regressions (WSSA)

OLS WLS IV IV
60–78 78–98 60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.143 0.175 0.136 0.187 0.126 0.165 0.182 0.170
( 0.024) ( 0.024) ( 0.021) ( 0.020) ( 0.025) ( 0.026) ( 0.026) ( 0.035)
y0 -0.010 -0.011 -0.009 -0.013 -0.008 -0.011 -0.014 -0.011
( 0.002) ( 0.002) ( 0.002) ( 0.002) ( 0.002) ( 0.002) ( 0.002) ( 0.003)
n -0.211 -0.772 -0.174 -0.737 -0.266 -0.798 -0.127 -0.743
( 0.180) ( 0.148) ( 0.173) ( 0.146) ( 0.183) ( 0.151) ( 0.157) ( 0.170)
i/y 0.011 0.021 0.013 0.018 0.007 0.018 0.005 0.016
( 0.003) ( 0.003) ( 0.003) ( 0.003) ( 0.003) ( 0.004) ( 0.003) ( 0.005)
Educ. – – – – – – 0.014 0.003
( 0.004) (0.007)
SSA-Dum. -0.018 -0.012 -0.017 -0.019 -0.019 -0.013 -0.020 -0.016
( 0.004) ( 0.004) ( 0.004) ( 0.004) ( 0.004) ( 0.005) ( 0.005) ( 0.006)
R2 0.32 0.52 0.44 0.60 – – – –
Q(Total) 6.335 [0.042] 8.616 [0.013] 5.619 [0.060] 8.944 [0.011]
Q(y0 ) 0.371 [0.543] 3.158 [0.076] 0.627 [0.429] 0.451 [0.502]
Q(n, i/y) 14.466 [0.001] 9.147 [0.010] 12.255 [0.002] 13.464 [0.001]
Q(Educ.) – – – – – – 1.921 [0.166]

Note: Standard errors into parenthesis, p–values into brackets. WLS: Weights are given by the log
of the population. IV: the instruments are the initial level of (i/y), the average (c/y) over the sub–
sample, and the average growth rate of population over the 15 first periods of the sub–sample. 106
observations except for the last column where only 86 observations are available.)

30
the first sub–period, this coefficient was significant and amounted to 0.011. Over the second
sub–period, it remained significant and rose to 0.021 — therefore doubling between the two
sub–periods. The stability test accordingly indicates a significant structural break between the
two periods (p–value=0). The results are once again robust to the adoption of a WLS procedure
and instrumental variable estimation.

Beyond our benchmark regression, table 15 also reports an IV estimates augmented to include an
education variable. The education variable we include is a weighted average of school enrollment
for primary, secondary and tertiary degrees of education.32 Data availability problems imply
that 20 countries are lost in that experiment. Interestingly, the results indicate that over the
1960–1978 sub–period education contributed significantly and positively to growth, but over the
1978–1998 sub–period this effect vanished, which once again suggests that the importance of
education did not increase over time. Furthermore, the inclusion of education in the growth
equation has basically two effects: (i) it makes convergence parameter very stable over the two
sub–periods and (ii) it reinforces the increasing effect of accumulation factors. For instance,
on average the effects of population growth are multiplied by a factor of 5 and the effects of
investment are multiplied by almost 3. This is actually slightly stronger than the effects we
obtained previously.

Hence, it appears that the introduction of Sub Saharan African countries does not affect our
observation that the importance of traditional growth factors increased substantially from the
period 1960-1978 to the period 1978-1998.

5.3 Implementing the decomposition

We now undertake a decomposition of the observed changes in the world distribution of output–
per worker for our wider sample of countries. We want to argue that once we take care of a
specific effect characterizing sub Saharan African countries in the early period (1960–78), then
the explanation we presented in section 4 — i.e. the increase importance of tradition factors of
accumulation over the later period — remains capable of explaining the bulk of changes observed
between 1978 and 1998.

Figure 8 reports the decomposition of changes in the distribution controlling for different effects.
In performing this decomposition, we use our estimates from the OLS regression in table 15. The
first observation we want to highlight concerns the change in the distribution from 1960 to 1978.
To this end, in panel (a) of figure 8 we report the observed distribution of output per worker in
1960 (grey line) and in 1978 (dark plain line), and the distribution of counterfactual 1978 output
per worker (dark dashed line) where we control for all effects specific to sub Saharan African
countries. More specifically, we built a counterfactual output per worker series for 1978 taking
into account (i) the sub Saharan dummy and (ii) the sub Saharan unobservable component.
32
See footnote 26 for details on this variate.

31
 Figure 8: Decomposing the Observed Changes in the Distributions
(a) Sub−Saharan effect (60−78) (b) Sub−Saharan effect (78−98)
0.4 0.4
1960 1978
1978 1998
0.3 Fitted 0.3 Fitted

0.2 0.2

0.1 0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4

(c) Distributions (n,i/y) (d) Coefficients (n,i/y)

0.4 0.4
1978 1978
1998 1998
0.3 Fitted 0.3 Fitted

0.2 0.2

0.1 0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4
(e) Observables (n,i/y) (f) Obs. (n,i/y) + Sub−Saharan effect (60−78)
0.4 0.4
1978 1960
1998 1998
0.3 Fitted 0.3 Fitted

0.2 0.2

0.1 0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4

32
That is, we build the counterfactual as follows:

60−78 SSA
ybi78 = yi78 − 18 × (βD Di + DiSSA × ε60−78
i )

60−78
where DiSSA = 1 for sub–saharan African countries and 0 otherwise, βD is the associated
coefficient. In this counterfactual, the sub-Saharan African countries effectively only contribute
their fitted mean values based on the convergence, population growth and investment variables
and parameters. As can be seen from the figure, once we control for all effects specific sub
Saharan African countries, the hump–shape observed in 1978 disappears: the counterfactual
1978 distribution is almost the same as the 1960 distribution. We perform the same exercise for
the period 1978 to 1998 to see if this effect is repeated in the second period. The results from
this exercise is reported in panel (b) of figure 8. In this case, the effects specific to sub Saharan
effect are not sufficient to reproduce the 1978 distribution. Indeed, there remains substantial
movement in the distribution which needs to be explained. Therefore, we now turn to examining
whether changes in the growth process can explain the movements observed between 1978 and
1998.

In panel (c) of figure 8, we examine the effects induced by changes in the distribution of tradi-
tional accumulation factors, that is, population growth and investment rate. This counterfactual
experiment is based on equation (5). As in our restricted sample, changes in the distribution of
accumulation forces accounts for little of the observed change in output per worker distribution.
Conversely, as can be seen from panel (d) of figure 8, and just as in our restricted sample, the
modification in the shape of the distribution seems to originate from a change in the coefficients
of accumulation forces. In fact, again as in our restricted sample, this change in the coefficients
tends to over-explain the observed changes in the distribution — though to a lesser extent than
we observed with the restricted sample. If, as in panel (e) of the figure, we combine the effects
of changes in the distribution of accumulation factors and changes in their related coefficients —
the counterfactual experiment being then based on equation (8) — the 1998 distribution (dark
plain line) is brought close to the 1978 distribution (grey line). In other words, the observed
changes in the shape of the world output per worker distribution between 1978 and 1998 — just
as in our restricted sample — appears to be the reflection of a change affecting the accumulation
process that drives output growth.

As a final exercise, we correct both for the change in the observable component of accumulation
and for the specific sub Saharan Africa effect observed in the period 1960-78. In this case — as
illustrated by panel (f) of the figure — this counterfactual brings the 1998 distribution almost
inline with the 1960 distribution, implying that two factors are important for explaining the
world distribution of output per worker over the period 1960-98: (i) a dramatic change in the
importance of traditional factors affecting growth after 1978 and (ii) the bad residual draws
experienced by sub Saharan African countries during the earlier period of the sample.

33
6 Discussion

The results we have presented up to this point invite the question of what type of model could
explain the observed increase in importance of traditional growth factors. Although a complete
answer to this question is beyond the scope of the current paper, we nevertheless want to
highlight three elements that are potentially relevant for answering such a question. First,
following Bernanke and Gürkaynak [2001], we want to examine whether the mechanisms at play
most likely reflects traditional neo–classical forces or if, alternatively, some element of endogenous
technological change is likely part of the story. Second, we want to examine whether international
trade is potential a relevant element to be considered.

6.1 Could it all be due to Neoclassical forces?

In recent work, Bernanke and Gürkaynak [2001] have examined whether total factor produc-
tivity (TFP) growth across countries appears mainly exogenous with respect to a country’s
accumulation path, or if, conversely, a country’s accumulation affects TFP growth and thereby
suggests some form of externality or endogenous technological change. In this section, we pursue
this idea one step further by comparing the behaviour of TFP over the periods 1960–78 and
1978–98. Our goal is to assess whether the behaviour of TFP has changed over this period, as
has been the case for GDP growth, or whether the observed increase in importance of traditional
growth factors is restricted to the output growth process. To perform this exercise, we build
TFP series for each country by subtracting from the rate of growth of output per worker the
share weighted contribution of the rate of growth of physical capital per worker. Since we do
not have good share data for all these countries, we build two TFP series using a share of capital
of 0.35 — which is close to the average share of capital across–countries reported in Bernanke
and Gürkaynak [2001] ( we verified that our results are robust to using a share of capital of 0.5
instead of 0.35). Capital series were constructed using the perpetual inventory method, with
the details of this construction given in appendix A.

Table 16 reports estimation results for TFP regressions. In the first four columns of the table, we
report estimation results for OLS and IV regressions relating the growth in TFP to the traditional
accumulation factors (population growth and investment rate), and an additional term that aims
at capturing a potential technological catchup phenomenon — i.e. the initial (log) level of TFP
for the given sub–period. Note that the results presented in this table correspond to NSSA
sample of countries.33 The set of instrument we use is that corresponding to IV? . The main
feature that emerges from this table is that the results found for output growth appear to apply
for TFP growth. In particular, the magnitude of the coefficients affecting the two traditional
accumulation forces increased significantly between the two sub–periods. The stability tests,
both for our OLS and IV estimations, clearly lead to a strong rejection of the stability of these
33
Results are robust to including sub-Saharan African countries in the sample.

34
 Table 16: TFP regressions

OLS IV1 IV1 IV2

60–78 78–98 60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.118 0.112 0.111 0.114 0.030 0.038 0.142 0.125
( 0.016) ( 0.021) ( 0.017) ( 0.022) ( 0.005) ( 0.006) ( 0.023) ( 0.034)
n -0.073 -0.481 -0.122 -0.484 -0.201 -0.451 -0.010 -0.548
( 0.125) ( 0.130) ( 0.128) ( 0.133) ( 0.151) ( 0.144) ( 0.130) ( 0.138)
i/y 0.009 0.014 0.005 0.015 0.003 0.013 0.001 0.017
( 0.002) ( 0.003) ( 0.002) ( 0.003) ( 0.003) ( 0.003) ( 0.004) ( 0.005)
T F P0 -0.014 -0.012 -0.014 -0.012 – – -0.019 -0.013
( 0.003) ( 0.003) ( 0.003) ( 0.003) ( 0.004) ( 0.006)
Educ. – – – – – – 0.013 -0.000
( 0.006) ( 0.011)
R2 0.38 0.43 – – – – – –
Q(Total) 9.734 [0.021] 12.830 [0.005] 8.050 [0.018] 22.467 [0.000]
Q(T F P0 ) 0.276 [0.599] 0.157 [0.692] – – 0.899 [0.343]
Q(n, i/y) 9.071 [0.011] 12.160 [0.002] – – 14.198 [0.001]
Q(Educ) – – – – – – 1.028 [0.311]

Note: Standard errors into parenthesis, p–values into brackets. IV1: the instruments are the initial
level of (i/y), the average (c/y) over the sub–sample, and the average growth rate of population over the
15 first periods of the sub–sample. (75 observations) IV2: the set of instruments of IV1 is completed
by the years of schooling to instrument the education variable. (68 observations)

coefficients. In particular, if we consider the coefficient measuring the impact of investment on

TFP growth, we find that it was multiplied by a factor of 3 between the two sub–periods when
estimated with IV. Similarly the magnitude of the population growth effect was multiplied by
a factor of 4. Furthermore, those two coefficients are strongly significant over the second sub–
period, indicating that treating TFP as exogenous over this period would be unreasonable. If we
now focus on the coefficient on the initial level of TFP, we find that — whatever the estimation
method — it is strongly significant but remains stable over the two sub–periods. Hence, these
data also suggest that there was significant technological catchup taking place over the entire
period between 1960 and 1998. Even if we decide to abstract from this catchup phenomenon
(third set of results in table 16), accumulation forces are still found to be unstable between the
two sub–periods, but the change in population growth is less pronounced.

To complete the picture regarding TFP growth, in the last two columns of the table we introduce
educational investment34 in the regression, and we add the years of schooling in the list of
instruments. The results are once again very close to those found for the output growth process.
In particular, there is no indication that educational investment has played more of a role in
the second period. Furthermore, the inclusion of educational investment does not reduce the
significance of our previous observations.
34
This variable corresponds to the weighted sum of primary, secondary and tertiary enrolments rates.

35
To sum up, the results of table 16 suggests that explanations of the observed changes in the
world distribution of income between 1960 and 1998 should include an endogenous growth type
mechanism that explains the observed increase in importance of traditional growth factors on
TFP growth.

6.2 Is trade relevant?

We now want to explore whether international trade may also be a relevant component for
explaining the observed increase in the role of traditional growth factors. To this end, in table
17, we report a set of regressions where we split the sample into two groups depending on their
ratio of total exports plus imports over GDP, which is a measure of openness available in the
Penn World Tables. More precisely, we split the sample according to the median value of the
average of this openness variable over the period 1960–78.35 We report results based on our
parsimonious specification where K is used to capture the effects of traditional growth forces.
The set of instruments used is again IV? . The results reported in the first two columns closely
parallels those in table 5. The main feature we want to highlight from this table is the difference

Table 17: Growth Regressions and Openness

Total O 60−78 <med(O 60−78 ) O60−78 >med(O 60−78 )

60–78 78–98 60–78 78–98 60–78 78–98
Const. 0.121 0.121 0.096 0.125 0.169 0.126
( 0.021) ( 0.022) ( 0.023) ( 0.024) ( 0.035) ( 0.041)
y0 -0.011 -0.014 -0.009 -0.013 -0.015 -0.015
( 0.002) ( 0.002) ( 0.003) ( 0.003) ( 0.004) ( 0.004)
K 0.009 0.029 0.017 0.018 0.006 0.041
( 0.003) ( 0.004) ( 0.004) ( 0.004) ( 0.005) ( 0.006)
Stab.(Total) 17.846 [0.000] 1.565 [0.457] 48.966 [0.000]
Stab.(y0 ) 0.834 [0.361] 1.453 [0.228] 0.001 [0.971]
Stab.(K) 15.647 [0.000] 0.048 [0.827] 27.795 [0.000]

Note: Standard errors into parenthesis, p–values into brackets. the instruments are the
initial level of (i/y), the average (c/y) over the sub–sample, and the average growth rate
of population over the 15 first periods of the sub–sample. O denotes the degree of openess
of the economies as measured by the ratio of total exports plus imports to gdp.

in the behaviour of estimates between the more open countries (O 60−78 >med(O 60−78 )) and the
less open countries (O 60−78 <med(O 60−78 )). In particular, for the more closed economies, we
do not observed a significant change in importance of traditional growth factors; while for the
more open economies we see a very significant and substantial change in the importance of this
factor. Moreover, the difference in behaviour is quantitatively quite large. In the case of the
35
We experimented different assumption on the splitting of countries, making it depend on either the degree of
openness in 1960, or the average degree of openness in the second sub–period, and found no significant differences
with respect to the results we are currently reporting.

36
more closed economies, the increase in the coefficient on K was only 0.001, while it is 0.035 (35
times bigger) for the sample of more open economies. It is immediately worth noting that this
pattern contrasts sharply with previous splits, whereby the instability result was found to be
robust to dividing the sample along either income or schooling lines. Hence the results in table
17 suggest that the observed rise in importance of traditional growth factors, and the associated
twin–peaks phenomenon, may reflect the process of globalization or, alternatively, may reflect
changes in the prices of traded goods.36

7 Concluding remarks

The object of the paper has been to shed light on the relative importance of different factors
reshaping the world distribution of output-per-worker between 1960 and 1998. To this end, we
have broken down the problem in several steps. Firstly, we discuss why it may be useful to ex-
amine sub-Saharan Afican countries apart from the rest of the sample. Secondly, we showed why
it is relevant to separate the time period into two sub–periods: 1960-78 and 1978-98. Thirdly,
we present a simple decomposition methodology which allows to separate observed changes in
the distribution output per worker into three components: (i) changes in the distribution of
observable factors affecting growth, (ii) changes in the importance of the observable factors
affecting growth, and (iii) changes in the distribution of residual (or un–measured) factors af-
fecting growth. Our main finding is that most of the change in distribution observed since
1978 appears driven by changes in the importance of the two traditional factors reflecting the
gains to physical capital accumualtion — that is the investment rate and the rate of population
growth. In contrast, we found that factors other than that associated with capital deepening
were relevant only for understanding the sub–Saharan African countries in the early period.

We take the results of our exercise as indicating new directions for research aimed at under-
standing global trends and the associated changes in the world distribution of income since 1960.
In particular, we think that such research should investigate the reasons for the substantive in-
crease in the social returns to capital accumulation we document in the period 1978–98 relative
to 1960–78. We believe that it is only by explaining these changes that one will explain the
observed change in the distribution of output-per-worker. As a preliminary step in this direc-
tion, we have argued that an explanation for this change will likely involve (i) some element of
endogenous technical change associated with capital deepening externalities, (ii) some element
of international transmission through trade.

36
It is interesting to note that, over the 1978–98 period, the estimated speed of convergence in the more opened
economies is identical to that in the more closed. If this speed of convergence is interpreted as a reflection of
technological diffusion, which would be consistent with our results from the TFP regressions, it suggests that the
differences between the more open and more closed economies is not due to differences in technological diffusion.

37
References

Barro, R. and J.W. Lee, International Comparisons of Educational Attainment, Journal of

Monetary Economics, 1993, 32 (3), 363–394.

Beaudry, P. and F. Collard, Why has the Employment–Productivity Tradeoff among Industrial-
ized Countries been so Strong?, Working Paper 8754, NBER 2002.

Benhabib, J. and M.M. Spiegel, The Role of Human Capital in Economic Development: Evi-
dence from Aggregate Cross–Country Data, Journal of Monetary Economics, 1994, 34 (2),
143–174.

Bernanke, B.S. and R.S. Gürkaynak, Is Growth Exogenous? Taking Mankiw, Romer and Weil
Seriously, in B.S. Bernanke and K. Rogoff, editors, NBER Macroeconomics Annual, Vol. 16,
Cambridge MA: MIT Press, 2001, chapter 1.

Durlauf, S.N. and D.T. Quah, The New Empirics of Economic Growth, in J.B. Taylor and
Woodford M., editors, Handbook of Macroeconomics, Vol. 1, Amsterdam: Esevier Science,
1999.

Jones, C.I., On the Evolution of the World Income Distribution, Journal of Economic Perspec-
tives, 1997, 11 (3), 19–36.

Juhn, C., K.M. Murphy, and B. Pierce, Wage Inequality and the Rise in Returns to Skill,
Journal of Political Economy, 1993, 101 (3), 410–442.

Klenow, P.J. and A. Rodrı̀guez-Clare, The Neoclassical Revival in Growth Economics: Has
it Gone too Far?, in B.S. Bernanke and J. Rotemberg, editors, NBER Macroeconomics
Annual, Vol. 12, Cambridge MA: MIT Press, 1997, chapter 2.

Krueger, A.B. and M. Lindhal, Education for Growth: Why and for Whom?, Journal of
Economic Literature, 2001, 39 (4), 1101–1136.

Lucas, R., On the Mechanisms of Economic Development, Journal of Monetary Economics,

1988, 22, 3–42.

Mankiw, N.G., D. Romer, and D.N. Weil, A Contribution to the Empirics of Economic Growth,
Quarterly Journal of Economics, 1992, 107 (2), 407–437.

Quah, D., Empirical Cross–Section Dynamics in Economic Growth, European Economic Review,
1993, 37 (2/3), 426–434.

, Empirics for Growth and Distribution: Polarization, Stratification, and Convergence

Clubs, Journal of Economic Growth, 1997, 2 (1), 27–59.

Sala–i–Martin, X., The World Distribution of Income (estimated from Individual Country Dis-
tributions), miméo, Columbia University 2002.

38
Solow, R.M., A Contribution to the Theory of Economic Growth, Quarterly Journal of Eco-
nomics, 1956, 70 (1), 65–94.

Uzawa, H., Optimum Tehcnical Change in an Aggregative Model of Economic Growth, Inter-
national Economic Review, 1965, 6 (1), 18–31.

39
A Data

Two datasets were used in this study. Most of the data are taken from the latest version of the
Penn World Table 6.0 downloadable from http://webhost.bridgew.edu/baten/.

Education data are taken from the Barro and Lee [1993] dataset, which is downloadable from
http://www.nuff.ox.ac.uk/Economics/Growth/datasets.htm.

A.1 Main data

Our measure of income, y, is the logarithm of real GDP chain per worker (RGDPW in PWT 6.0),
where the definition of a worker is based of economically active population.

Population is POP in PWT 6.0. Workers are computed as the population from 15 to 64 obtained
from
real GDP chain per capita RGDPL
POPWt = × population = × POP
real GDP chain per worker RGDPW
n then denotes the rate of growth of the 15–64 population.

The corresponding annualized average rate of growth for the variable Z within the sub–sample
[t;t+n] is computed as
log(Zt+n ) − log(Zt )
∆z =
n

The share of consumption at constant prices corresponds to the variable KC in the PWT 6.0. In
the IV procedure, the average share of consumption c/y over the sub–sample [t;t+n] is computed
as
n
1X
log(KCt+j /100)
n
j=1

The investment ratio at constant prices corresponds to the variable KI in the PWT 6.0 and is
divided by 100. In the regressions, i/y then refers to the logarithm of this variable. It is also
used to compute our accumulation variable that accounts for the overall accumulation effect,
which is given by µ ¶
KI
K ≡ log
(1 + γ)(1 + n) + δ − 1
We followed Mankiw et al. [1992] and assume a annual depreciation rate of δ = 0.03 and a rate
of growth of technical progress, γ, of 2%.37

Furthermore, the investment share is used to compute the capital stock per worker — needed
to compute the TFP — using the permanent inventory scheme

(1 + n)kt+1 = it + (1 − δ)kt
37
We performed robustness checks but did not find any significant effect of a change in either δ or γ.

40
where k = K/L. k1960 is obtained as
à 1970
!
X
1
10 KIt RGDPW1960
t=1960
k1960 =
(1 + n60−70 )(1 + γ) + δ − 1
where
log(POPW1970 ) − log(POPW1960 )
n60−70 ≡
10
We also followed Mankiw et al. [1992] and assume a annual depreciation rate of δ = 0.03 and a
rate of growth of technical progress, g, of 2%. The measure of TFP then follows

T F P = log(RGDPW) − α log(k)

where α is set to 0.35 and 0.5 to check robustness.

A.2 Education data

The education variables are borrowed from Barro and Lee [1993]. We consider essentially 3
measures of human capital. The first one is related to the overall enrollment rate in education.
Assuming that, on average, most people sped 6 years in primary schooling, 6 years in secondary
schooling and 4 years in higher schooling, we first define the index
µ ¶
6×P+6×S+4×H
log
16
where P, S and H respectively denote the total gross enrollment ratio for, respectively, primary,
secondary and higher schooling.

Our second measure of the education variable also relates to the enrollment rate but focuses on
the primary and secondary schooling. It is built as
µ ¶
DURP × P + DURS × S
log
DURP + DURS
where DURP and DURS respectively denote the duration in years of, respectively, the primary and
secondary education in 1965.

Our last education variable, Years, should be more understood as a stock since it is given by

log(HUMAN)

where HUMAN is the average schooling years in the total population over 25.

Note that these measures are only reported every 5 years in the database from 1960 to 1985. We
actually used an average of years 1960 and 1965 for the first sub–sample and 1975 and 1980 for
the second sub–sample. We restrict ourselves to 2 periods for the average for data availability
purposes.

41
A.3 Composition of the sample

Our restricted sample of 75 countries consists of:

Argentina, Australia, Austria, Belgium, Bangladesh, Bolivia, Brazil, Barbados, Botswana, Ca-
nada, Switzerland, Chile, China, Colombia, Costa Rica, Cyprus, Denmark, Dominican Re-
public, Ecuador, Egypt, Spain, Finland, Fiji, France, United Kingdom, Greece, Guatemala,
Guyana, Hong Kong, Honduras, Indonesia, India, Ireland, Iran, Iceland, Israel, Italy, Jamaica,
Jordan, Japan, Republic of Korea, Sri Lanka, Lesotho, Luxembourg, Morocco, Mexico, Mozam-
bique, Malaysia, Namibia, Nicaragua, Netherlands, Norway, Nepal, New Zealand, Pakistan,
Panama, Peru, Philippines, Papua New Guinea, Portugal, Paraguay, Romania, Singapore, El
Salvador, Sweden, Syria, Thailand, Trinidad and Tobago, Tunisia, Turkey, Taiwan, Uruguay,
USA, Venezuela, South Africa,

When we move to our enlarge sample of 106 countries we add the following 31 sub–saharan
African countries:

Burundi, Benin, Burkina Faso, Central African Republic, Ivory Coast, Cameroon, Republic of
Congo, Comoros, Cape Verde, Ethiopia, Gabon, Ghana, Guinea, The Gambia, Guinea–Bissau,
Kenya, Madagascar, Mali, Mauritania, Mauritius, Malawi, Niger, Nigeria, Rwanda, Senegal,
Seychelles, Chad, Togo, Tanzania, Uganda, Zambia.

B Kernel density estimates

The density estimates are computed using a kernel density estimator. Given a set of data
Y = {yi }N
i=1 , the Rosenblatt–Parzen kernel density estimator of Y is given by

N µ ¶
1 X y − yi
fb(y) = K
Nh h
i=1

where the kernel, K(.), is a continuous, bounded, symmetric real–valued function that integrates
√
to 1. For our purpose, we used a Gaussian kernel, K(ω) = exp(−ω 2 /2)/ 2π, with a bandwidth
parameter, h, chosen as h = 1.0592σN −1/5 where σ is the standard deviation of the data and
N is the number of observations.

42
 Figure 9: Decomposing the observed changes in the distributions (IV)

78−98
(a) Observables (y ) (b) Coefficients (βy =0)
0
0.5 0.5
1960
0.4 1998 0.4
Corr.
0.3 0.3

0.2 0.2

0.1 0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4
(c) Distribution (n,i/y) (d) Coefficients (n,i/y)
0.5 0.8

0.4
0.6
0.3
0.4
0.2
0.2
0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4

(e) Observables (n,i/y) (f) Unobservables

0.8 0.5

0.4
0.6
0.3
0.4
0.2
0.2
0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4

Note: The decomposition is based on the IV3 regression presented in table 4.

43