Distribution Dynamics and Nonlinear Growth: Davide Fiaschi

Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

Journal of Economic Growth, 8, 379±401, 2003

# 2003 Kluwer Academic Publishers. Manufactured in The Netherlands.

Distribution Dynamics and Nonlinear Growth


DAVIDE FIASCHI
University of Pisa, Dipartimento di Scienze Economiche, Via Ridol® 10, 56124 Pisa, Italy

ANDREA MARIO LAVEZZI


University of Pisa, Dip.to di Scienze Economiche, Via Ridol® 10, 56124 Pisa, Italy

This paper extends the distribution dynamics approach to the study of the shape of the growth process of a cross-
section of countries. We ®rst identify some empirical implications of a nonlinear Solovian growth model with
multiple equilibria. These implications are then tested by a novel de®nition of the state space, which jointly
considers income and growth rate. The main ®ndings are that nonlinearity is a salient feature of the overall
picture, and that the cross-section dynamics is compatible with the existence of multiple equilibria. We also
discuss how the hypothesis of conditional convergence may be challenged in the light of our results.

Keywords: nonlinear growth, distribution dynamics, conditional convergence, nonparametric methods

JEL classi®cation: O11, O40, C14, C21

1. Introduction

This paper is a contribution to the debate on the empirical analysis of economic growth,
recently surveyed by Durlauf and Quah (1999) and Temple (1999). The huge empirical
work on growth produced over the last years still leaves many questions open, both in the
identi®cation of growth patterns in the available data, and in the search for causal
explanations of growth.
The main concern of this paper is with the former issue: we focus on nonlinearities in
the growth process from the point of view of the evolution of cross-section distributions
(see Quah, 1993). In particular, as argued by Barro and Sala-i-Martin (1999, pp. 51±52):
``One empirical implication of the model [with poverty traps] . . . is that there would exist a
middle range of values of k [capital] . . . for which the growth rate [of capital] gk is
increasing in k and, hence, in y [income]. . . . Our reading of the evidence across
countries . . . does not support this hypothesis. . . . Fundamentally, these policy conclusions
[e.g. donation by an international organization of discrete quantity of capital to a country
to allow it to escape from a poverty trap] depend on the idea that production involves a
380 DAVIDE FIASCHI AND ANDREA LAVEZZI

range of decreasing returns, followed by a range of increasing returns. . . . We do not know,


however, of empirical evidence that supports the underlying pattern of decreasing/
increasing returns. Therefore, these policy inferences are only speculative.'' As is
apparent from the quotation, the presence of a range of income levels where growth rates
are increasing is a necessary condition for the existence of poverty traps. Hence, the
detection of nonlinearities is relevant for the design of economic policies. From another
point of view, the presence of nonlinearities would cast doubts on those contributions
aiming to estimate the speed of convergence.
In this paper we show that growth appears as a nonlinear process. We identify three
income ranges characterized by a different relation between income and growth rate. At
low income levels the relation is negative, at intermediate levels it is positive and, ®nally,
at high income levels it is again negative. In particular, countries in the intermediate
income range appear to experience rapid growth (take-off ) with increasing growth rates.
Quah (1993) pioneered the distribution dynamics approach. This approach is
informative on some aspects of the mobility of countries across income levels, but fails
to reveal the shape of the growth process. For this reason we introduce a novel de®nition of
the state space, which jointly takes into account income levels and growth rates, in contrast
to previous works which considered only income levels. This allows us to capture the
presence of nonlinearities.
Speci®cally, we ®rst present a graphical analysis, providing evidence of a range of
(relative) income levels where growth rates are increasing. Then we consider a nonlinear
Solovian growth model with multiple equilibria, and derive its empirical predictions in
terms of state space dynamics. Finally, we test for these predictions by estimating Markov
transition matrices. We also argue that our ®ndings question the hypothesis of conditional
convergence and the results of contributions on the speed of convergence, due to their
common assumption of a concave production function, implying growth paths
monotonically converging to the steady state.
Our ®ndings relate to the literature on the transition from stagnation to growth, in which
the nonlinearity of the growth process is a by-product of the process of development. In
early stages of development economies are in a state of stagnation and they take-off to a
state of sustained economic growth in later stages of development due to structural change
(Rostow, 1960), or an endogenous technological progress and demographic transition
(Galor and Weil, 2000). In addition, the ®ndings are related to the theory of club
convergence (Galor, 1996) and its quantitative representation (Quah, 1997).
Although nonlinearities are crucial features of the new growth theory, empirical practice
mostly concentrates on linear models (see Durlauf, 2001). Notable exceptions are Durlauf
and Johnson (1995), Liu and Stengos (1999) and Kalaitzidakis et al. (2001), but these
papers focus on the nonlinear effects of some explanatory variables. In particular, Liu and
Stengos (1999) and Kalaitzidakis et al. (2001) provide some evidence consistent with our
results of the existence of an income range where the growth rate is increasing, but their
analysis regards initial income levels (see also Durlauf et al., 2001).
The rest of the paper is organized as follows: Section 2 contains the preliminary
graphical analysis; Section 3 discusses the empirically testable implications of a Solovian
model with nonlinearities; Section 4 reports the quantitative results; Section 5 concludes.
The appendices gather some extensions of the empirical analysis.
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 381

2. Graphical Analysis

The basic unit of observation in this paper is the annual per capita GDP level and the
corresponding annual growth rate. Data are from the Penn World Table 5.6 for the 1960±
1989 period and cover 120 countries.1 In particular, by GDP we mean the relative income
with respect to the (world) average of the period.2
In Figure 1 the growth rate is plotted against the (relative) per capita GDP for all the
observations in the sample. A nonparametric estimation of the relationship between these
two variables identi®es a nonlinear relationship between the level of GDP and the growth
rate.3
From the nonparametric regression in Figure 1 we observe that growth is ®rst high and
then decreasing for the lowest levels of GDP.4 The subsequent range of GDP is
characterized by increasing growth rates that, after reaching a peak, show a tendency to
decrease. In the last range of GDP the growth rate settles at a steady positive value. Liu and
Stengos (1999) and Kalaitzidakis et al. (2001) ®nd a similar relationship between log of
initial per capita GDP and growth rate controlling for cross-country heterogeneity in
investment rates, population growth and human capital. We control our results for cross-
country heterogeneity in Appendix B.
By looking only at Figure 1 one may conclude that there exists a tendency towards
convergence. In fact, the estimate always lies above the x-axis, while the richest countries
show decreasing growth rates with respect to the GDP level. However, this conclusion is
questionable if we take into account the growth performances of many African countries,
which appear stuck at relatively low GDP levels (see also Quah, 1993).
The point is that, since the observations for individual countries span only a part of the

Figure 1. Relative GDP levels vs growth rates.


382 DAVIDE FIASCHI AND ANDREA LAVEZZI

GDP range reported in Figure 1, the pooling of cross-country data can be misleading in the
identi®cation of actual growth patterns. In this regard we will see that a complementary
analysis based on transition matrices allows us to avoid such mistakes, since it permits us
to keep track of the growth path of each individual country.

3. A Simple Growth Model

In this section we present a simple growth model that can account for the facts emerging
from the previous graphical analysis. Then we derive its empirical implications, which are
tested in Section 4.
Figure 2 reports an example of a nonlinear Solovian growth model, where y and yE
respectively represent the level of per capita income and of per capita income in ef®ciency
_ and y_ E =yE are their respective growth rates. Denote the path of y_E =yE by
units, while y=y
yE E
GP : along GPy the model has three equilibria, two of which are stable ( yE1 and yE3 ), and a
region of increasing growth rate with respect to income: ‰ y2 ; y Š. The income level yE2
E ^E

represents a threshold: countries with initial income below (above) yE2 converge to yE1 … yE3 †
yE
following the growth path GP . In the literature different growth models display multiple
equilibria. In a Solow-type model, multiple equilibria can be the result of non-monotonic
decline of the average product of capital, as shown in Barro and Sala-i-Martin (1999,
p. 50), or of a non-constant saving rate, as shown in Galor (1996, p. 1059). We refer to
these references for more details.
Since the database allows for the calculation of growth rates of per capita GDP, it is

Figure 2. Growth model and state space de®nition.


DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 383

E
convenient to move from the space …y_ E =yE ; yE † to …y=y;_ yE †. GPy , de®ned in the space
E E E E
…y_ =y ; y †, can be mapped into the space …y=y;_ y † by a simple upward shift of the curve
equal to the growth rate of the exogenous technological progress (in Figure 2 we denote it
with g). The resulting growth path of per capita income is indicated by GPy.
In Figure 2 we superimpose a partition of the space …y=y; _ yE † in 12 regions based on
three levels of per capita income in ef®ciency units, yI , yEII and yEIII (which de®ne four
E

income classes), and two levels of the growth rate, g ‡ and g ‡‡ (which de®ne three
growth rate classes).
The partition in Figure 2 identi®es regions in the space …y=y; _ yE † characterized by
speci®c relationships between income level and growth rate. In particular, we expect
persistence in income class I at medium growth rates. Moreover, this class should contain
an attractor, hence we expect that in equilibrium a relevant fraction of countries
concentrates here. By contrast, countries in income class II should show either an
increasing growth rate or a decreasing/low growth rate if they locate, respectively, on the
right or on the left of the unstable equilibrium yE2 . These countries should show a tendency
to move either into class I or into class III. Countries in income class III should show
increasing (at least up to y^E ) and/or persistently high growth rates with respect to all the
other income classes. Also, we expect countries in this income class to show a higher
tendency to move into income class IV than into class II. Finally, income class IV, like
income class I, should behave like an attractor, with persistence at medium growth rates.
These predictions are tested in the following section, where the partition of the space
_ yE † represents the state space of a Markov process.
…y=y;

4. Empirical Analysis

In this section ®rst we discuss the methodology used for the empirical investigation and
then present the results.

4.1. Methodology for Empirical Investigation

Following Quah (1993), the growth dynamics of the sample is represented by Markov
transition matrices, i.e. we consider a discrete state space and one-year transitions. The
analysis of the transition matrix and the associated ergodic distribution allows a researcher
to capture dynamics normally missed by standard cross-country regressions (see Durlauf
and Quah, 1999).
However, in the literature on distribution dynamics, the state space has been de®ned so
far only in terms of income levels. The novelty of the paper, as remarked, is that the state
space is de®ned in terms of both income levels and growth rates. By this de®nition, it is
possible to detect at the same time the presence of two types of heterogeneity: across time
and across countries. The presence of time-heterogeneity in the growth path (i.e.
nonlinearities) implies that different income classes are characterized by different growth
rate dynamics, whereas cross-sectional heterogeneity can be re¯ected in the tendency
384 DAVIDE FIASCHI AND ANDREA LAVEZZI

towards polarization of the cross-country distribution. Clearly, these two tendencies can be
present at the same time.
A drawback in de®ning the states in terms of both income level and growth rate is the
complex choice of income and growth rate classes, since the number of states tends to be
high even for a low number of classes for each one of two variables. For example,
choosing ®ve income level classes as in Quah (1993) and three growth rate classes (the
minimum if we want to detect possible nonlinear dynamics) generates 15 states.
Before proceeding, we describe the steps to make inference on estimates of transition
probabilities. Assume that the process is stationary and has k states. The elements of the
unobservable transition matrix, pij …i; j [ {1; . . . ; k}†, i.e. transition probabilities, can be
estimated by p^ij ˆ nij =ni, where ni is the number of observations in state i, and nij is the
number of observed transitions from state i to state j. These estimates, as shown for
instance in Norris (1997, pp. 56±57), are the maximum likelihood estimators of the true
transition probabilities pij . Stuart and Ord (1994, p. 260), show that the distribution of
~
ni ˆ …ni1 ; . . . ; nik † converges to a n-variate normal distribution, with means ni pij ,
p
variances ni pij …1 pij † and covariances cov…nij ; niq † ˆ ni pij piq . Thus ni …^ pij pij †
tends towards the normal distribution N…0; pij …1 pij ††. Therefore, for transition
probabilities of different rows
s
^2ij s
s ^2mq
pij p^mq †=
s ˆ …^ ‡ ;
ni nm
converges to a Gaussian distribution under the null hypothesis p^ij ˆ p^mq , where i=m and
^2ij ˆ p^ij …1 p^ij †, while for transition probabilities of the same row
s
s
^2ij s
s ^2iq 2 cov…^ pij ; p^iq †
s ˆ …^pij p^iq †= ‡ ;
ni ni ni
converges to a Gaussian distribution under the null hypothesis p^ij ˆ p^iq , where
pij ; p^iq † ˆ p^ij p^iq . Finally, to make inference on the elements of ergodic distribution
cov…^
we utilize a procedure based on delta method (see Trede, 1999).5

4.2. De®nition of the State Space

In Section 3 we showed that a de®nition of three growth rate classes and four income
classes, i.e. a total of 12 states, is suf®cient to generate empirically testable implications of
the presence of nonlinearities. Moreover, the discussion of the previous section suggests
keeping the number of states as low as possible.6 Thus we adopt the de®nition of the state
space in Table 1,7 where mI is the average annual per capita GDP.8
The procedure for de®ning the class boundaries consists of two steps. First we set the
growth rate classes on the basis of an estimate of the growth rate of the exogenous
technological progress, and then we set the GDP classes by a comparison of Figure 1 with
Figure 2.9
As regards the growth rate classes, the theoretical model shows that the class limits
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 385

Table 1. State space de®nition.

Income\Growth rate 5 0.8% 0.8%±2.8% 4 2.8%

0--0:3mI I I‡ I ‡‡
0:3mI --0:9mI II II ‡ II ‡‡
0:9mI --2:5mI III II ‡ III ‡‡
> 2:5mI IV IV ‡ IV ‡ ‡

should be set in order to obtain a central class which includes the growth rate of
technology. The identi®cation of this rate reasonably concerns the subset of the wealthiest
countries. On inspection of Figure 1 it seems that the richest countries are converging
towards a growth rate equal to 1.8 percent. Hence a range of + 1 percent should plausibly
include the shift of the technological frontier. Therefore, the three resulting growth rate
classes are de®ned as: ‰… ?, 0.8%†; …0.8%; 2.8%†; …2.8%; ‡ ?%†Š.10
In order to avoid the problems related to discretization of an otherwise continuous state
space, in Appendix C we report the kernel density estimation for the growth rate
transitions within every GDP class.
The de®nition of relative per capita GDP classes directly follows from imposing these
growth rate classes on Figure 1, taking into account the representation of the state space in
Figure 2.

4.3. Results

Table 2 reports the transition matrix that obtains by applying the de®nition of states in
Table 1.11
The ®rst column contains the number of observations for every state. The ergodic
distribution is reported in Table 3.12
Table 4 shows the ergodic distribution with respect to a normalization of the
distribution's mass in every GDP class. This representation of the ergodic distribution
highlights the growth rate dynamics within each GDP class in steady state.

Table 2. Transition matrix.

Obs States I I‡ I ‡‡ II II ‡ II ‡‡ III III ‡ III ‡‡ IV IV ‡ IV ‡‡

423 I 0.54 0.14 0.32 0 0 0 0 0 0 0 0 0


118 I‡ 0.42 0.20 0.37 0 0 0 0 0 0 0 0 0
337 I ‡‡ 0.39 0.12 0.45 0.02 0 0.01 0 0 0 0 0 0
470 II 0.03 0.01 0.01 0.47 0.14 0.34 0 0 0 0 0 0
221 II ‡ 0 0 0 0.35 0.22 0.42 0 0 0 0 0 0
593 II ‡‡ 0 0 0 0.26 0.16 0.53 0.01 0 0.04 0 0 0
202 III 0 0 0 0.06 0.01 0.04 0.46 0.16 0.26 0 0 0
132 III ‡ 0 0 0 0 0.01 0 0.23 0.17 0.55 0.01 0.02 0.02
445 III ‡‡ 0 0 0 0 0 0 0.16 0.16 0.65 0 0 0.02
93 IV 0 0 0 0 0 0 0.05 0.02 0.02 0.29 0.30 0.31
125 IV ‡ 0 0 0 0 0 0 0 0.02 0.01 0.23 0.34 0.39
201 IV ‡‡ 0 0 0 0 0 0 0 0 0 0.16 0.27 0.56
386 DAVIDE FIASCHI AND ANDREA LAVEZZI

Table 3. Ergodic distribution.

I I‡ I ‡‡ II II ‡ II ‡‡ III III ‡ III ‡‡ IV IV ‡ IV ‡‡

0.19 0.06 0.15 0.10 0.05 0.13 0.05 0.03 0.10 0.03 0.04 0.07

Table 4. Ergodic distribution normalized for every GDP class.

‡ ‡‡

I 0.48 0.14 0.38


II 0.38 0.17 0.45
III 0.27 0.17 0.56
IV 0.22 0.31 0.47

Table 5. First and last year distribution vs ergodic distribution.

I II III IV

1960 0.20 0.47 0.22 0.11


1989 0.31 0.34 0.19 0.16
Ergodic 0.41 0.28 0.18 0.14

Finally, Table 5 contains the cross-country income distribution of the ®rst and last year,
along with the ergodic distribution in terms of GDP classes only.
Table 5 provides information on the speed of convergence of the actual distribution to
its ergodic limit, and on the tendency to move from one GDP class into another. Figure 3 is
a graphical representation of the results, which are discussed in details below.
Figure 3 is a simpli®ed version of Figure 2, in which we report for each element of the
state space the transition probabilities from Table 2, and the income distribution dynamics
from Table 5 (respectively, the element of the distribution in 1960, p1960 , and of the
ergodic distribution, pe ). The reported probabilities refer to transitions across growth rate
classes (see note 13): we denote a transition to the adjacent higher (lower) growth rate
class by : …;†, while a jump to the furthest higher (lower) growth class by :: …;;†; ®nally, a
transition to the same growth rate class is denoted by ;:.
In Appendix D we report the estimates of transition matrices with 5-year transitions and
with 3-year average growth rates and the related tables. Our aim is to check whether our
results depend on the possible autocorrelation of shocks. This is particularly relevant for
low income countries, where measurement error can induce serial correlation between
growth rates. Overall, our results do not seem to be affected by this phenomenon.

4.3.1. Nonlinearities in the Growth Process

In the following we assess for each GDP class whether the predictions of our theoretical
model ®nd support in the data.
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 387

Figure 3. State space de®nition and estimated transition probabilities.

As expected, GDP class I behaves like an attractor, as the fraction of countries in this
class is increasing and takes on a relatively high value in steady state (see Table 5). From
the transition matrix we observe a remarkable tendency of deceleration of growth. A
country in this class is much more likely to show a low growth rate after a high growth rate,
with respect to all the other GDP classes. The relevant probabilities are: Pp^I ‡ ‡ ; ˆ 0:41,
p^II ‡ ‡ ; ˆ 0:27, p^III ‡ ‡ ; ˆ 0:16 vs p^IV ‡ ‡ ; ˆ 0:16, where p^ir; w ˆ q [ fI; ...; IVg p^ir; qw
with i [ {I; . . . ; IV} and r, w [ f ; ‡ ; ‡ ‡ g.13 In particular, the probability of
maintaining a high growth rate for two consecutive periods is almost equal to the
probability of having low growth …^ pI ‡ ‡ ; ‡ ‡ ˆ 0:46 vs p^I ‡ ‡ ; ˆ 0:41†.14 In addition, a
country in GDP class I has a relatively high probability of persisting at low growth rates.
Speci®cally, p^I ; I ˆ 0:54 is higher than p^i ;i …i ˆ II; III; IV†.15
The prediction formulated in Section 3 on persistence at medium growth rates does not
®nd a direct support. Namely, the transition probabilities towards the medium growth rate
classes are not high, and the fraction of countries in the ergodic distribution in the medium
growth rate class (0.14) is lower than the other two (see the ®rst row of Table 4). However,
we attribute the lack of persistence at medium growth rates to the higher growth volatility
that exists at low income levels.16 A comparison of the transition matrix in Table 2 with the
transition matrix with 3-year average growth rates in Table 15 provides partial support to
this hypothesis. In Table 15 we observe a clear increase in p^I ; ‡ , p^I ‡ ; ‡ and p^I ‡ ‡ ; ‡ .
Also, the fraction of countries that are in the medium growth rate class in GDP class I in
steady state increases from 0.14 to 0.23 (see the ®rst row of Table 17).17 Further support is
388 DAVIDE FIASCHI AND ANDREA LAVEZZI

given by the kernel density estimation in Appendix C, where peaks of marginal densities in
GDP class I correspond to medium growth rates. This suggests that the estimated high
values of the probabilities of transitions to extreme (low/high) growth rates can be due to
the discretization of the state space.
In GDP class II the fraction of countries tends to reduce (see Table 5, second column), as
expected if this class contains the unstable equilibrium. A country in GDP class II shows a
higher probability of persisting at low growth rates than moving to the highest growth rate
class …^ pII ; ˆ 0:50 vs p^II ; ‡ ‡ ˆ 0:35†, as in GDP class I. However, unlike in GDP
class I, the probability of persisting at high growth rates is much higher than the
probability of moving to a lower growth rate class …^ pII ‡ ‡ ; ‡ ‡ ˆ 0:57 vs
p^II ‡ ‡ ; ˆ 0:27†.18 Also, the fact that in GDP class II there is a higher probability of
increasing a medium growth rate than in GDP class I re¯ects the possibility of entering the
acceleration regime (i.e. moving from GDP class II to GDP class III). Inspection of the
kernel density estimation in Appendix C con®rms that, with respect to GDP class I, there
exists such tendency of a change in the growth regime. A tendency also re¯ected in the
values in Table 4, where the probability of high (low) growth rate is increased (decreased)
with respect to GDP class I.
In GDP class III the fraction of countries tends to reduce (see Table 5, third column), as
expected. More importantly, GDP class III should be characterized by (i) acceleration and
by (ii) persistence at high growth rate. In our framework these features are captured by:
(i) the probability of moving from a state with a ``medium'' growth rate to a state with
a ``high'' growth rate and, (ii) the probability of exhibiting persistently high growth rates.
In the transition matrix the set of relevant probabilities for point (i) is given by:
p^I ‡ ; ‡ ‡ , p^II ‡ ; ‡ ‡ , p^III ‡ ; ‡ ‡ and p^IV ‡ ; ‡ ‡ . The estimated values are, respectively:
p^I ‡ ; ‡ ‡ ˆ 0:37, p^II ‡ ; ‡ ‡ ˆ 0:42, p^III ‡ ; ‡ ‡ ˆ 0:57 and pIV ‡ ; ‡ ‡ ˆ 0:40.19 Thus, it
appears that a country in income class III is relatively more likely to show accelerating
growth. More precisely, we ®nd that the probability of a country increasing an already
sustained growth rate rises with income in the ®rst three GDP classes, and then decreases
in the fourth.
As regards point (ii), note that for a country with a high growth rate, the probability
of maintaining such rate of growth is highest in income class III. The relevant
value is p^III ‡ ‡ ; ‡ ‡ ˆ 0:67 against p^I ‡ ‡ ; ‡ ‡ ˆ 0:46, p^II ‡ ‡ ; ‡ ‡ ˆ 0:57 and
p^IV ‡ ‡ ; ‡ ‡ ˆ 0:56.20 Estimates of transition matrices with 5-year transitions and with 3-
year average growth rates in Appendix D show that our ®ndings do not seem to depend on
the possible autocorrelation of shocks. Again, the kernel density estimation in Appendix C
con®rms that GDP class III is the ``acceleration'' income class. In addition, countries in
GDP class III are relatively more likely to have a high growth rate in steady state (see the
last column in Table 4).21
In GDP class IV we observe ®rst that the fraction of countries tends to increase (see
Table 5, fourth column), and this is compatible with the presence of an attractor in this
GDP class. Also, we expected a tendency to persist at medium growth rates. This fact ®nds
support in the transition matrix in the high estimated values of p^IV ; ‡ , p^IV ‡ ; ‡ and
p^IV ‡ ‡ ; ‡ . Estimates of the transition matrices with 5-year transitions or with 3-year
average growth rates in Appendix D also suggest that countries in GDP class IV are likely
to settle their growth rate at a medium level (e.g. p^IV ‡ ; ‡ is equal to 0.43 and 0.55,
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 389

respectively, for estimates with 5-year transitions and with 3-year average growth rates).
Moreover, in the stochastic kernel for GDP class IV in Appendix C the peaks of the
conditional densities are found at medium growth rates for any initial growth rate. Finally,
on inspection of the ergodic distribution in Table 4, we see that countries in GDP class IV
have the highest probability of a medium growth rate (0.31) with respect to other GDP
classes.22
Overall, these results are broadly in accordance with the predictions on the shape of the
growth process and on the cross-country distribution dynamics from Figure 2.

4.3.2. Existence of Poverty Traps

A recurrent question in growth empirics is the existence of poverty traps.23 Comparison


between the initial, ®nal and ergodic distribution in terms of only GDP classes in Table 5
shows that, as noted, a process of polarization is taking place, in the sense that the weight
of middle income classes tends to fall.24
In accordance with Figure 2 it appears that, even in steady state, a subset of countries is
in a poverty trap, as the lowest relative income class persistently contains a considerable
fraction of countries. Hence, in the present framework a poverty trap must be understood
with respect to the entire distribution, and does not correspond to absolutely low income
levels from which a country cannot escape. Again, we take these facts as providing support
for the existence of multiple stable equilibria as in Figure 2.
Finally, another issue related to the ergodic distribution is the speed of convergence to
the ergodic limit. Table 5 shows that full convergence is far from being achieved for low
income classes, but not for the high GDP classes. Following Shorrocks (1978, pp. 1021±
1022), the asymptotic half life of this process equals about 49 periods (years).25 This value
should be taken as evidence that the speed of convergence is slow and not as a precise
forecast of the time needed for convergence. In fact, this would require the stability of the
process for a century or more, which we cannot take for granted, as we observed only a
limited period of time.

4.3.3. On Conditional Convergence

Our ®ndings challenge the hypothesis of conditional convergence. In particular, we


question the presence of conditional convergence according to the predictions of the
neoclassical concave growth model. In such a model, conditional convergence occurs
through a monotonic decrease or increase in the growth rate, until all countries grow at the
common exogenous growth rate (the steady state levels of income may be different across
countries because of cross-country heterogeneity in the determinants of the steady state).26
Our results are in contrast with this prediction.
To see this, consider the standard Solovian model with a concave production function
and cross-country heterogeneity in our state space in Figure 4. Conditional convergence
and the concavity of the production function imply that a country's growth rate should
decrease (increase) if, respectively, it is converging from the left (right) to its long-run
390 DAVIDE FIASCHI AND ANDREA LAVEZZI

Figure 4. Conditional convergence and state space de®nition.

equilibrium level of per capita income (in e®ciency units). In Figure 4 we report four
equilibria: yE1 , yE2 , yE3 and yE4 .
In terms of our state space de®nition, if a country is converging from the left to its steady
state, it will be either in the ``high'' growth rate class or in the ``medium'' growth rate
class, and its growth rate would decrease or remain stable. Otherwise, if a country is
converging from the right, it will be either in the medium growth rate class or in the low
growth rate class, and would show increasing or steady growth rates. Finally, if a country is
near its steady state, it will be in the medium growth rate class and it would show steady
growth rates. All this implies that the probability of observing an increase in the growth
rate of a country with a medium growth rate would be low (as well as the probability of a
low growth rate). In terms of our state space de®nition, the medium growth rate class
would be an attractor, independent of which per capita GDP classes we consider.
However, we have just highlighted that, especially for a country in GDP class III, the
probability of moving from a medium to a high growth rate …^ pIII ‡ ; ‡ ‡ ˆ 0:57† is very
high compared to the probabilities of maintaining the same growth rate …^ pIII ‡ ; ‡ ˆ 0:19†,
or exhibiting a decrease in it …^ pIII ‡ ; ˆ 0:24†.27
The small range of the medium growth rate class and the high growth volatility for some
GDP classes imply that the mass of probability in this growth rate class is always lower
compared with the other two classes. However, given a medium growth rate, we observe a
substantial symmetry in the probabilities of a high or a low growth rate for GDP classes I
and II (the values of such probabilities are p^I ‡ ; ˆ 0:42 vs p^I ‡ ; ‡ ‡ ˆ 0:37 and
p^II ‡ ; ˆ 0:35 vs p^II ‡ ; ‡ ‡ ˆ 0:42).28 By contrast, GDP class III shows a strong and
highly signi®cant asymmetry (0.24 vs 0.57), as does GDP class IV (0.23 vs 0.40).
This evidence appears robust to the possible autocorrelation of shocks. From Appendix D
we see that the estimates with 5-year transitions and 3-year average growth rates con®rm
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 391

that countries in GDP class III have a strong tendency to increase their growth rate, given a
medium growth rate (respectively, p^III ‡ ; ‡ ‡ ˆ 0:49 vs p^III ‡ ; ˆ 0:31 in Table 11 and
p^III ‡ ; ‡ ‡ ˆ 0:51 vs p^III ‡ ; ˆ 0:28 in Table 15).29 Interestingly, in GDP class IV, where
we know convergence in income takes place, a country with a medium growth rate has the
highest probability of maintaining the same growth rate level (for 5-year transitions:
p^IV ‡ ; ‡ ˆ 0:43 vs p^IV ‡ ; ˆ 0:21 and p^IV ‡ ; ‡ ‡ ˆ 0:3530 and, with 3-year averages,
p^IV ‡ ; ‡ ˆ 0:55 vs p^IV ‡ ; ˆ 0:11 and p^IV ‡ ; ‡ ‡ ˆ 0:3531). This is in agreement with our
claims, since the hypothesis of (conditional) convergence in GDP class IV (like in GDP
class I) would hold in our framework.
To explicitly control for cross-country heterogeneity we conducted an analysis in the
spirit of Liu and Stengos (1999). In Appendix B we report the details of a semiparametric
regression, where the linear terms are the usual determinants of the steady state in the
standard Solow model (investment ratios, growth rates of population, rates of depreciation
and technical progress), while relative initial per capita GDP enters nonlinearly. The
existence of a range of income where growth rates are increasing is con®rmed, and the
turning points of the relation income-growth rate broadly correspond to those identi®ed in
the previous sections (see also Kalaitzidakis et al., 2001).
We do not discuss further the presence of conditional convergence in the sample, which
can occur in the presence of nonlinearities in the production function. However, we argue
that our results question the hypothesis of conditional convergence derived by the standard
growth concave model. This is particularly relevant for all contributions aiming to
estimate the speed of convergence, which generally assume monotonically convergent
growth paths.

5. Conclusions

The main result of the paper is the detection of nonlinearities in the growth process. In
particular, our ®ndings suggest a range of incomes characterized by an accelerating growth
rate, which eventually decelerates once a country catches up with the wealthier countries.
However, this only holds for a subset of countries, because we ®nd evidence of poverty
traps, in the sense that the cross-country distribution persistently displays a considerable
fraction of countries lagging at low income levels.
Moreover, our ®ndings suggest that the hypothesis of conditional convergence based on
a concave production function should be rejected. In this regard, estimates of speed of
convergence should take into account the presence of nonlinearities.
Many models can account for the presence of nonlinearities. In particular, traditional
development theories provide interesting insights on the transition from stagnation to
growth, such as the focus on structural change, which modern growth literature has
recently started to explore intensively (see, e.g. Galor and Weil, 2000). This is the direction
of our current research.
392 DAVIDE FIASCHI AND ANDREA LAVEZZI

Appendix

A. Country List of Sample PWT

The following is the list of countries included in the sample PWT (source: Penn-World
Table 5.6). The series are real per capita GDP in constant (1985) dollars using the Chain
Index.

Algeria Angola Reunion Brazil


Benin Rwanda Chile Belgium
Botswana Senegal Colombia Cyprus
Burkina Faso Seychelles Ecuador Czechoslovakia
Burundi Somalia Guyana Denmark
Cameroon South Africa Paraguay Finland
Cape Verde Swaziland Peru France
Central Afr. Togo Suriname Germany West
Chad Tunisia Uruguay Greece
Comoros Uganda Venezuela Iceland
Congo Zaire Bangladesh Ireland
Egypt Zambia China Italy
Gabon Zimbabwe Hong Kong Luxembourg
Gambia Barbados India Malta
Ghana Canada Indonesia Netherlands
Guinea Costa Rica Iran Norway
Guinea-Biss. Dominican Republic Israel Portugal
Ivory Coast El Salvador Japan Romania
Kenya Guatemala Jordan Spain
Lesotho Haiti Korea Rep. Sweden
Madagascar Honduras Malaysia Switzerland
Malawi Jamaica Myanmar Turkey
Mali Mexico Pakistan U.K.
Mauritania Nicaragua Philippines U.S.S.R.
Mauritius Panama Saudi Arabia Yugoslavia
Morocco Puerto Rico Singapore Australia
Mozambique Trinidad & Tobago Sri Lanka Fiji
Namibia U.S.A. Syria New Zealand
Niger Argentina Taiwan Papua N. Guinea
Nigeria Bolivia Thailand Austria

B. Controlling for Cross-country Heterogeneity

To control for cross-country heterogeneity we perform a semiparametric analysis in the


spirit of Liu and Stengos (1999). We build up a new data set by adding to our data on per
capita GDP data on investment ratios and population growth rates extracted from Penn
World Table 6.1. This implies a reduction of the sample to 106 countries (the countries
excluded are: Reunion, Somalia, Swaziland, Tunisia, Haiti, Puerto Rico, Suriname,
Myanmar, Saudi Arabia, Czechoslovakia, Germany West, Malta, USSR and Yugoslavia).
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 393

For the semiparametric regression we use the ``gam'' package available under the R
environment (codes and dataset are available in our websites). We take 5-year averages on
investment and population, and normalize the level of per capita GDP at the beginning of
each 5-year period to the sample average of the period (the total number of observations is
636: 106 countries for 6 periods 1960, 1965, 1970, 1975, 1980, 1985). Then we estimate
the following equation:

gi ˆ a ‡ b1 log…INVi † ‡ b2 log…POPi † ‡ s…log…iniGDPi †† ‡ ui ;

where g is the average growth rate over the ®ve-year period, INV is the investment ratio,
POP is the growth rate of population (we add 0.05 as in Durlauf and Johnson (1995) and in
Liu and Stengos (1999) to capture the combined effect of the depreciation rate and
technological change) and iniGDP is the relative initial-period per capita GDP. Notice that,
in accordance with our paper and differently from Liu±Stengos, we use relative per capita
GDP. We do not take into account the time dummies D70 and D80 for the seventies and
eighties, as in Liu and Stengos, since we consider that normalizing per capita GDP to
world average should capture the time effect (at any rate, including the dummies does not
alter our results). The estimation of the nonlinear component is reported in Figure 5.
Figure 5 shows that, in accordance with our results, there exists a range of per capita
GDP where growth rates are increasing. Bowman and Azzalini (1997, p. 154); suggest a
simple test to compare linear and nonlinear speci®cations: in our case the nonlinear
speci®cation provides a better description of the data (we refer to Bowman and Azzalini,
1997, pp.150±163, for details). We checked this result with annual data and with 3- and 10-
year averages obtaining a similar picture. These ranges are compatible with those used in
the transition matrix derived from Figure 1. In particular, the turning points broadly
correspond to those identi®ed in the paper. Finally, the residuals from the regression
plotted against relative per capita GDP levels do not show any discernible pattern.

Figure 5. Relative initial GDP vs growth rate.


394 DAVIDE FIASCHI AND ANDREA LAVEZZI

C. Kernel Density Estimation

Figure 6 contains the contour plots from kernel density estimations for 5-year transitions
of the growth rate within every GDP class (see Durlauf and Quah, 1999 for details on
stochastic kernels).32 The vertical and horizontal axes respectively refer to growth rate in
year t …gt † and growth rate in year t ‡ 5 …gt ‡ 5 †. We superimpose a grid representing our
growth rate classes and a 45 line, which helps to identify the probabilities of acceleration
or deceleration of growth rates.
We see that in GDP class I for high/medium initial growth rates, the highest probability
is to have a decelerating growth rate, since the peaks of the conditional densities are above
the 45 line starting from a current growth rate approximately equal to 1.8 percent. In GDP
class II, for an initial growth rate below 2.0 percent the peaks of conditional densities
highlight an acceleration of growth rate, as they move to the right of 45 line with respect
to GDP class I. Also, in this GDP class, for initially high growth rates, the peaks of
conditional distributions are still above the 45 line, but, with respect to GDP class I, they

Figure 6. Kernel density estimation.


DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 395

correspond to higher growth rates. In GDP class III the threshold separating accelerating
from decelerating growth rates becomes approximately 3.5 percent. Finally, in GDP class
IV 2.5 percent approximately represents the limit separating the range of accelerating
growth rates from the decelerating growth rates.
Overall, in GDP class I transitions towards medium/low levels of growth rates appear
relatively more probable; in GDP class II the probability of transitions towards medium/
high growth rates increases with respect to GDP class I. In GDP class III there is a further
change in the shape of the conditional probabilities in favor of transitions towards high
growth rates. Finally, in GDP class IV we observe a concentration of the overall
conditional probability mass at medium growth rates.

D. Other Estimates

In this appendix we report alternative estimates. We do not provide any discussion of


results, except that they support our previous ®ndings.

D.1. Estimates with GDP Classes Taken from Quah (1993)

We combine the de®nition of income classes of Quah (1993) with our de®nition of growth
rate classes, obtaining a state space with 15 states in Table 6.33
The resulting transition matrix is in Table 7, while the ergodic distribution for all the
states is in Table 8. The ergodic distribution for each GDP class is in Table 9.
Finally, in Table 10 we report the distribution of the ®rst and last year distribution vs the
ergodic distribution for only GDP classes.

D.2. Estimates with 5-year Transitions

In this section we calculate the transition matrix refers to 5-year transitions, that is we
consider transitions from …yt ; gt † to …yt ‡ 5 ; gt ‡ 5 †. In Tables 11±14 we report the results.

Table 6. Income states de®ned in Quah (1993).

Income/Growth rate 5 0.8% 0.8%±2.8% 4 2.8%

0±0:25mI I I‡ I ‡‡
0:25mI ±0:5mI II II ‡ II ‡‡
0:5mI ±mI III III ‡ III ‡‡
mI ±2mI IV IV ‡ IV ‡‡
> 2mI V V‡ V ‡‡
396 DAVIDE FIASCHI AND ANDREA LAVEZZI

Table 7. Transition matrix.

States I I‡ I ‡ ‡ II II ‡ II ‡ ‡ III III ‡ III ‡ ‡ IV IV ‡ IV ‡ ‡ V V‡ V‡‡

I 0.55 0.13 0.32 0 0 0 0 0 0 0 0 0 0 0 0


I‡ 0.43 0.21 0.35 0 0 0 0 0 0 0 0 0 0 0 0
I‡‡ 0.36 0.11 0.43 0.05 0.01 0.03 0 0 0 0 0 0 0 0 0
II 0.06 0.03 0.05 0.47 0.14 0.26 0 0 0 0 0 0 0 0 0
II ‡ 0 0 0.02 0.38 0.23 0.37 0 0 0 0 0 0 0 0 0
II ‡ ‡ 0 0 0 0.30 0.16 0.44 0.04 0.02 0.05 0 0 0 0 0 0
III 0 0 0 0.04 0.01 0.07 0.45 0.12 0.31 0 0 0 0 0 0
III ‡ 0 0 0 0 0.01 0.01 0.32 0.18 0.47 0 0 0.02 0 0 0
III ‡ ‡ 0 0 0 0 0 0 0.23 0.15 0.58 0.01 0 0.03 0 0 0
IV 0 0 0 0 0 0 0.06 0.01 0.05 0.45 0.17 0.27 0 0 0
IV ‡ 0 0 0 0 0 0 0 0 0 0.31 0.19 0.50 0 0 0
IV ‡ ‡ 0 0 0 0 0 0 0 0 0 0.15 0.14 0.67 0.01 0 0.02
V 0 0 0 0 0 0 0 0 0 0.04 0.01 0.02 0.33 0.30 0.31
V‡ 0 0 0 0 0 0 0 0 0 0 0 0 0.20 0.34 0.45
V‡‡ 0 0 0 0 0 0 0 0 0 0 0 0 0.16 0.24 0.60
N Obs 344 99 280 295 124 284 286 132 431 132 78 278 131 163 303

Table 8. Ergodic distribution.

I I‡ I‡‡ II II ‡ II ‡ ‡ III III ‡ III ‡‡ IV IV ‡ IV ‡ ‡ V V‡ V‡‡

0.14 0.04 0.12 0.08 0.03 0.07 0.06 0.03 0.09 0.04 0.02 0.07 0.04 0.06 0.10

Table 9. Ergodic distribution normalized for every GDP class.

‡ ‡‡

I 0.48 0.14 0.38


II 0.43 0.17 0.39
III 0.34 0.15 0.50
IV 0.28 0.16 0.56
V 0.21 0.28 0.50

Table 10. Distribution of the ®rst and last year vs ergodic distribution for only GDP classes.

I II III IV V

1960 0.14 0.25 0.32 0.12 0.16


1989 0.26 0.18 0.22 0.16 0.18
Erg. distr. 0.30 0.19 0.18 0.13 0.20
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 397

Table 11. Transition matrix.

N. Obs States I I‡ I‡‡ II II ‡ II ‡ ‡ III III ‡ III ‡ ‡ IV IV ‡ IV ‡ ‡

347 I 0.50 0.13 0.36 0.01 0 0 0 0 0 0 0 0


92 I‡ 0.51 0.11 0.34 0.02 0.01 0.01 0 0 0 0 0 0
296 I‡‡ 0.40 0.13 0.39 0.03 0.02 0.04 0 0 0 0 0 0
387 II 0.06 0.02 0.04 0.34 0.15 0.36 0.01 0 0.02 0 0 0
182 II ‡ 0.04 0.01 0.03 0.32 0.21 0.35 0 0.01 0.03 0 0 0
552 II ‡ ‡ 0.02 0.01 0.01 0.31 0.12 0.42 0.02 0.02 0.07 0 0 0
170 III 0 0 0 0.07 0 0.09 0.25 0.17 0.38 0.01 0.01 0.02
112 III ‡ 0 0 0 0.04 0 0.04 0.26 0.17 0.38 0.01 0.04 0.07
394 III ‡ ‡ 0 0 0 0.03 0.01 0.02 0.23 0.12 0.51 0.02 0.02 0.05
84 IV 0 0 0 0 0 0 0.11 0.02 0.02 0.25 0.27 0.32
94 IV ‡ 0 0 0 0 0 0 0.04 0.02 0.02 0.17 0.41 0.33
170 IV ‡ ‡ 0 0 0 0 0 0 0.02 0.01 0.05 0.22 0.26 0.44

Table 12. Ergodic distribution.

I I‡ I‡‡ II II ‡ II ‡ ‡ III III ‡ III ‡ ‡ IV IV ‡ IV ‡ ‡

0.22 0.06 0.17 0.09 0.04 0.11 0.05 0.03 0.09 0.03 0.05 0.06

Table 13. Ergodic distribution normalized for every GDP class.

‡ ‡‡

I 0.48 0.14 0.38


II 0.38 0.17 0.45
III 0.30 0.17 0.53
IV 0.23 0.34 0.43

Table 14. Distribution of the ®rst and last year vs ergodic distribution for only GDP classes.

I II III IV

1960 0.20 0.47 0.22 0.11


1989 0.31 0.34 0.19 0.16
Ergodic distr. 0.45 0.24 0.17 0.14
398 DAVIDE FIASCHI AND ANDREA LAVEZZI

D.3. Estimates with 3-year Average Growth Rates

In this section we consider 3-year average growth rates, that is the annual observations on
income are considered against the average growth rate for the subsequent three years. That
is, we consider transitions from, e.g. …yt ; gt; t ‡ 2 † to …yt ‡ 3 ; gt ‡ 3; t ‡ 5 †, where gt; t ‡ 2 is the
average annual growth rate from period t to period t ‡ 2 and g^t ‡ 3; t ‡ 5 is the average annual
growth rate from period t ‡ 3 to period t ‡ 5. In Tables 15±18 we report the results.

Table 15. Transition matrix.

Obs States I I‡ I‡‡ II II ‡ II ‡ ‡ III III ‡ III ‡ ‡ IV IV ‡ IV ‡ ‡

126 I 0.50 0.22 0.28 0 0 0 0 0 0 0 0 0


65 I‡ 0.40 0.34 0.25 0 0.02 0 0 0 0 0 0 0
87 I‡ ‡ 0.38 0.16 0.39 0.01 0.01 0.05 0 0 0 0 0 0
134 II 0.08 0.02 0.03 0.45 0.15 0.27 0 0 0 0 0 0
84 II ‡ 0.01 0 0 0.36 0.24 0.39 0 0 0 0 0 0
199 II ‡ ‡ 0 0 0 0.22 0.17 0.50 0.03 0.02 0.07 0 0 0
57 III 0 0 0 0.12 0.02 0.14 0.39 0.07 0.26 0 0 0
51 III ‡ 0 0 0 0.02 0 0.04 0.24 0.20 0.45 0.02 0.02 0.02
146 III ‡ ‡ 0 0 0 0 0 0 0.14 0.19 0.59 0.02 0.01 0.05
21 IV 0 0 0 0 0 0 0.05 0.10 0.10 0.14 0.29 0.33
55 IV ‡ 0 0 0 0 0 0 0.02 0.02 0.02 0.09 0.53 0.33
55 IV ‡ ‡ 0 0 0 0 0 0 0 0 0.02 0.22 0.29 0.47

Table 16. Ergodic distribution.

I I‡ I‡‡ II II ‡ II ‡ ‡ III III ‡ III ‡ ‡ IV IV ‡ IV ‡ ‡

0.22 0.12 0.15 0.09 0.05 0.11 0.04 0.03 0.08 0.02 0.04 0.05

Table 17. Ergodic distribution normalized for every GDP class.

‡ ‡‡

I 0.45 0.23 0.31


II 0.36 0.19 0.45
III 0.25 0.19 0.56
IV 0.18 0.39 0.43

Table 18. Distribution of the ®rst and last year vs ergodic distribution for only GDP classes.

I II III IV

1960 0.20 0.47 0.22 0.11


1989 0.30 0.35 0.20 0.15
Ergodic distr. 0.49 0.25 0.15 0.11
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 399

Acknowledgments

We are very grateful to Carlo Bianchi, Eugene Cleur, Steven Durlauf, seminar participants at Bologna, Cagliari,
Pisa, Roma, Siena and Teramo, two anonymous referees and the editor for their comments and suggestions. The
usual disclaimers apply.

Notes

1. Appendix A contains the country list. Our sample is very similar to Quah's (1993) sample.
2. This allows us to take into account the presence of an underlying long-run trend affecting all countries
(typically referred to as exogenous technological progress). This approach is utilized in many works on
distribution dynamics (see Durlauf and Quah, 1999). Actually, following Jones (1997), US labor productivity
could be used as a proxy for the level of technology. In a companion paper we utilize the absolute per capita
GDP level (see Fiaschi and Lavezzi, 2002), and ®nd again evidence of nonlinearities.
3. The nonparametric estimate is obtained with the statistical package included in Bowman and Azzalini
(1997). We used the standard settings suggested by the authors (e.g. optimal normal bandwidth and weights
on observations according to their density). To test the robustness of this estimate, we ran an alter-
native nonparametric regression by using the plug-in method to calculate the kernel bandwidth, and
obtained a similar picture. We refer to Bowman and Azzalini (1997) for more details. Data sets and codes
used in the empirical analysis are available on the authors' websites (http://www-dse.ec.unipi.it/®aschi and
http://www-dse.ec.unipi.it/lavezzi/index.html).
4. Even if not immediately evident from Figure 1, an inverse relationship is likely to exist between growth
volatility and the level of GDP. We discuss this point in a companion paper (see Fiaschi and Lavezzi, 2003).
5. A statistical appendix on this procedure is available upon request.
6. In Appendix D we present an analysis with 15 states following the de®nition of Quah (1993). The results,
again, support our claim on the presence of nonlinearities.
7. Given this de®nition of GDP classes, the distribution of observations in GDP class I, II, III and IV is the
following: (0.26, 0.38, 0.23, 0.12).
8. We tested the robustness of our results by performing estimations with slightly different de®nitions of the
class boundaries, without obtaining any signi®cant change.
9. The choice of the number of states and of their boundaries could be based on a more rigorous statistical
procedure (see, e.g. Hansen, 2000). We leave this non-trivial extension for future work.
10. Notice the similarity with the classi®cation of countries in Jones (1997). Jones de®nes the growth rate of a
country in the post WWII period as ``Fast'', ``Intermediate'' or ``Slow'' according to the following three
classes: ‰… ?; 0.4%†; …0.4%; 2.4%†; …2.4%; ?%†Š. Such boundaries are simply obtained by adding and
subtracting one percentage point to the average growth of the US productivity on the period considered
(1960±1988) equal to 1.4 percent, assuming that the United States represent the technology frontier for that
period. As suggested by one referee, we tried asymmetric growth rate classes around 1.8 percent: 1.3 percent
below and 1 percent above and obtained results highly in accordance with those presented below.
11. In this matrix and in those in the appendices, rows may not sum to one due to rounding.
12. The ergodic distribution represents the long-run, or invariant, distribution. Its existence is generally
guaranteed if the process is irreducible, aperiodic and positive persistent. For an introduction to Markov
Chains theory see, e.g. Isaacson and Madsen (1976).
13. These ``modi®ed'' transition probabilities simply refer to each (stochastic) submatrix corresponding to a
GDP class. For example, the value p^I ‡ ‡ ; ˆ 0:41 is obtained by summing p^I ‡ ‡ ;I ˆ 0:39 and
p^I ‡ ‡ ; II ˆ 0:02. Tests of equality between p^I ‡ ‡ ; and, respectively, p^II ‡ ‡ ; , p^III ‡ ‡ ; and p^IV ‡ ‡ ;
return, respectively, the following p-values: 0, 0 and 0.
14. The p-value of the test of equality between p^I ‡ ‡ ; ‡ ‡ and p^I ‡ ‡ ; is equal to 0.13. The likely autocorrelation
of shocks can bias upwards the latter value. Indeed, the transition matrix with 5-year transitions in Appendix
D contains a similar value of the probability of maintaining a high growth rate or moving to a low growth rate
for this GDP class …^ pI ‡ ‡ ; ‡ ‡ ˆ p^I ‡ ‡ ; ˆ 0:43†.
400 DAVIDE FIASCHI AND ANDREA LAVEZZI

15. In particular, the difference with the complementary probability …1 p^I ; I † ˆ 0:46, is highly signi®cative.
16. See Fiaschi and Lavezzi (2003).
17. Notice that 0.23 is higher than the fraction of countries found in the medium growth rate class in GDP classes
II and III (see Table 17).
18. All these values are statistically different.
19. The test of equality between p^III ‡ ; ‡ ‡ and, respectively, p^I ‡ ; ‡ ‡ , p^II ‡ ; ‡ ‡ and p^IV ‡ ; ‡ ‡ returns, respectively,
the following p-values: 0.001, 0.005 and 0.004.
20. Also in this case the hypothesis of equality between p^III ‡ ‡ ; ‡ ‡ and the other probabilities can be rejected.
Respectively, the test gives the following p-values: 0, 0.001 and 0.003.
21. Tests of equality of 0.56 with, respectively, 0.38, 0.45 and 0.47 return the following p-values: 0, 0 and 0.01.
22. Tests of equality of this value with, respectively, 0.14, 0.17 and 0.17 return the following p-values: 0, 0 and 0.
23. This issue is clearly related to the controversy on (conditional) convergence among countries. We refer to
Durlauf and Quah (1999) for an exhaustive discussion of the argument.
24. A similar result is found by Quah (1993) even if his evidence is clearer than ours due to a different de®nition
of GDP classes. In particular, he de®nes the following GDP classes: ‰…0; 0.25†; …0.25; 0.5†; …0.5; 1†; …1; 2†;
…2; ?†Š; and obtains the following ergodic distribution: ‰0.28; 0.15; 0.12; 0.15; 0.30Š; where the mass is
concentrated at the extreme classes (``twin peaks''). If we de®ne the income classes in the same way,
considering that our sample is slightly different, the following ergodic distribution is obtained: ‰0.30; 0.19;
0.18; 0.13; 0.20Š (see Appendix D). In the ergodic distribution in Table 5 we do not observe the highest value
at the extreme GDP classes. However, our GDP class III overlaps with Quah's GDP class V (the upper limit of
our GDP class III is 2.5 and the lower limit of GDP class V of Quah is 2): hence it is likely that some countries
which are in our ergodic distribution in class III would be in GDP class V adopting Quah's de®nition. Notice
that the union of our GDP classes III and IV forms a class with almost the same limits of the union of Quah's
GDP classes IV and V, i.e. ‰0.9; + ?† vs ‰1; + ?†: with our data they present almost equal masses of
probability in the ergodic distribution (0.32 vs 0.33). This con®rms that the difference in the ergodic
distributions is essentially caused by the different de®nition of the GDP classes.
25. The asymptotic half life is de®ned as h ˆ log 2=log jl2 j, where l2 is the second largest eigenvalue of the
transition matrix. In our case l2 ^ 0:986. This measure of the speed of convergence is based on the time the
process takes from period t to reach half of the distance from its equilibrium level (the ergodic distribution).
Therefore there is a discrepancy between this result and results based on b-convergence, estimating an half
life of the process of about 23 years (see Barro and Sala-i-Martin, 1999, pp. 431 and 446).
26. In our framework GDP classes I and IV should show a similar behavior. However, in the former class such
behavior may be blurred by the high growth volatility.
27. The p-value of test of equality between p^III ‡ ; ‡ ‡ and the complementary probability is equal to 0.052.
28. Test of equality between p^I ‡ ; ˆ 0:42 and p^I ‡ ; ‡‡ ˆ 0:37 has a p-value equal to 0.27, while p^II ‡ ; ˆ 0:35
and p^II ‡ ; ‡‡ ˆ 0:42 has a p-value equal to 0.12. It is worth remarking that, according to our model, there
could be countries in GDP class II with medium growth rate, whose growth rate is increasing. On the
contrary, in GDP class I growth rates should be always converging towards the medium growth rate class.
Therefore, the low p-value of GDP class II could be explained by this fact.
29. The p-values are, respectively, 0.015 and 0.027.
30. The p-values are, respectively, 0.002 and 0.19.
31. The p-values are, respectively, 0 and 0.055.
32. We do not report kernel density estimate for 1-year transitions and 3-year average growth rates, as they
con®rm the results presented here.
33. Given this de®nition of GDP classes, the distribution of observations in GDP class I, II, III and IV is the
following: …0.22; 0.21; 0.25; 0.15; 0.18†.

References

Barro, R. J., and X. Sala-i-Martin. (1999). Economic Growth. London: MIT Press.
Bowman, A. W., and A. Azzalini. (1997). Applied Smoothing Techniques for Data Analysis. Oxford: Clarendon
Press.
DISTRIBUTION DYNAMICS AND NONLINEAR GROWTH 401

Durlauf, S. N. (2001). ``Manifesto for a Growth Econometrics,'' Journal of Econometrics 100, 65±69.
Durlauf, S. N., and P. A. Johnson. (1995). ``Multiple Regime and Cross-Country Growth Behaviour.'' Journal of
Applied Econometrics 10, 365±384.
Durlauf, S. N., and D. T. Quah. (1999). ``The New Empirics of Economic Growth,'' Ch. 4, In John B. Taylor and
Michael Woodford (eds.), Handbook of Macroeconomics. Vol. IA, Amsterdam: North-Holland.
Durlauf, S. N., A. Kourtellos, and A. Minkin. (2001). ``The Local Solow Growth Model,'' European Economic
Review 45, 928±940.
Fiaschi, D., and A. M. Lavezzi. (2002). ``Nonlinear Economic Growth: Evidence from a Large Sample of
Countries,'' Mimeo, University of Pisa.
Fiaschi, D., and A. M. Lavezzi. (2003). ``Explaining Growth Volatility,'' Mimeo, University of Pisa.
Galor, O. (1996). ``Convergence? Inferences from Theoretical Models,'' Economic Journal 106, 1056±1069.
Galor O., and D. N. Weil. (2000). ``Population, Technology, and Growth: from the Malthusian Regime to the
Demographic Transtition and Beyond,'' American Economic Review 90, 806±828.
Hansen, B. E. (2000). ``Sample Splitting and Threshold Estimation,'' Econometrica 68, 575±603.
Isaacson, D. L., and R. W. Madsen. (1976). Markov Chains: Theory and Applications. New York: John Wiley.
Jones, C. I. (1997). ``On the Evolution of World Income Distribution,'' Journal of Economic Perspectives 11, 19±
36.
Kalaitzidakis, P., T. P. Mamuneas, A. Savvides, and T. Stengos. (2001). ``Measures of Human Capital and
Nonlinearities in Economic Growth,'' Journal of Economic Growth 6, 229±254.
Liu, Z., and T. Stengos. (1999). ``Non-linearities in Cross-Country Growth Regressions: A Semiparametric
Approach,'' Journal of Applied Econometrics 14, 527±538.
Norris, J. R. (1997). Markov Chains. Cambridge: Cambridge University Press.
Quah, D. T. (1993). ``Empirical Cross-section Dynamics for Economic Growth,'' European Economic Review 37,
426±434.
Quah, D. T. (1997). ``Empirics for Growth and Distribution: Strati®cation, Polarization, and Convergence Clubs,''
Journal of Economic Growth 2, 27±59.
Rostow, W. W. (1960). The Stages of Economic Growth. Oxford: Oxford University Press.
Shorrocks, A. F. (1978). ``The Measurement of Mobility,'' Econometrica 46, 1013±1024.
Stuart, A., and K. Ord. (1994). Kendall's Advanced Theory of Statistics, 6th ed., vol. I. London: Edward Arnold.
Trede, M. (1999). ``Statistical Inference for Measures of Income Mobility,'' JahrbuÈcher fur NationaloÈkonomie
und Statistik 218, 473±490.
Temple, J. (1999). ``The New Growth Evidence,'' Journal of Economic Literature 37, 112±156.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy