Journal of Economic Growth, 2: 27–59 ( 1997)

°
c 1997 Kluwer Academic Publishers, Boston.

Empirics for Growth and Distribution:

Stratification, Polarization, and Convergence Clubs

DANNY T. QUAH
London School of Economics

This paper studies cross-country patterns of economic growth from the viewpoint of income distribution dynamics.
Such a perspective raises new empirical and theoretical issues in growth analysis: the profound empirical regularity
is an “emerging twin peaks” in the cross-sectional distribution, not simple patterns of convergence or divergence.
The theoretical problems raised concern interaction patterns among subgroups of economies, not only problems
of a single economy’s accumulating factor inputs and technology for growth.

Keywords: conditional, convergence, distribution dynamics, income distribution, inequality, trade, twin peaks

JEL classification: C13, C33, F43, O30

1. Introduction

This paper describes some recent research on patterns of growth across countries: the facts
that this paper seeks to document and to explain are given in the stylized features of Figure 1.
Figure 1 is a caricature that will inform my subsequent analysis. Below, I will describe
which aspects of this figure have been established to be accurate and which remain conjec-
ture. For the time being, however, it is useful simply to note some features in the caricature.
The horizontal axis in Figure 1 indexes time; the vertical axis, per capita incomes. Figure 1
records, for different time points, (the densities corresponding to) cross-country per capita
income distributions. As drawn, the distribution at time t shows most countries having a
medium level per capita income; there are few that are very rich, and few very poor.
Over time, cross-country income distributions fluctuate: Figure 1 makes explicit the
distribution again at t + s. In general, there is one such object for each time period.
Figure 1, therefore, is like a time-series plot, except that instead of recording the trajectory
of a scalar or vector quantity—like GNP, money, or the price level—the figure comprises
the trajectory of an entire distribution.
A first immediate question that pictures like Figure 1 address is whether poor countries are
catching up with rich ones. That would happen if, for example, the sequence of distributions
collapses over time to a degenerate point limit. But in general that need not occur, and there
are other ways whereby the poor can catch up with the rich—as illustrated, for instance, by
the criss-crossing arrows.
Figure 1, provocatively, shows the distribution at t + s to have a twin-peaks property:
there is a clustering together of the very rich, a clustering together of the very poor, and a
28 QUAH

Figure 1. Emerging twin peaks.

vanishing of the middle income class. By contrast, these features were not present at the
earlier time t: it therefore seems reasonable to call Figure 1 a picture of emerging twin
peaks.
In this work, there is nothing special about there being precisely two peaks or modes in
the time t + s distribution: What is important instead is that such features have surfaced
when previously they were absent. What also matters is that these features have a natural
interpretation in terms of polarization: those portions of the underlying population of
economies collecting in the different peaks may be said to be polarized, one group versus
another. More generally, if more than two peaks emerged, it might be natural to call
the situation stratification. With the underlying population being countries, the economic
historian’s notion of convergence clubs—of countries catching up with one another but only
within particular subgroups (e.g., Baumol 1986)—is also apposite.
This paper concerns that body of research on cross-country economic growth that attempts
to refine empirically and to understand theoretically such emerging twin-peaks properties.1
Such a focus might seem excessively narrow: it is useful, therefore, to note how this work
relates to other areas of research.
To study the dynamics of cross-section distributions of country incomes—as given in
Figure 1—is to combine simultaneously elements of macroeconomic reasoning, microeco-
nomic analysis, and econometric modeling. The researcher is concerned with macroeco-
nomic performance—measured in national income and aggregate growth—but for a rich
cross-section of individual cross-sectional units, all potentially interacting with one another.
Thus, macroeconomic theories of growth are relevant, but so are microeconomic models of
cross-sectional interaction. Figure 1 makes explicit that the success or failure of any one
country makes qualitative economic sense only in context: What does a 5% annual growth
EMPIRICS FOR GROWTH AND DISTRIBUTION 29

rate mean—is it high or is it low—if no other economy grows at less than 10% per year?
Or if no other economy has ever grown at more than 1% per year?
Arrows drawn in Figure 1 indicate a variety of intradistribution dynamics. Some countries
rich at time t + s had already been rich at time t; similarly, others poor at t + s had already
been poor at t. There is, therefore, persistence. However, there is also churning or mobility:
some of those rich at t + s had begun poor; some of those poor at t + s had begun rich.
From these and from the vanishing of the middle class between t and t + s, it is also natural
to suppose a separating: some groups of economies originally close together in the middle
class have subsequently separated, with some becoming much richer than others—even
though they had begun close together.
Figure 1 thus contains a rich spectrum of dynamic behavior. Not only is the global,
external shape of the distribution evolving—with twin peaks emerging, and stratification
and polarization settling in—but also intradistribution mobility is simultaneously occurring.
Some portions of the distribution display persistence in rich and poor states, others show
overtaking dynamics, and yet others a slowing-down in growth so that they are themselves
overtaken. Put differently, there are both shape and mobility dynamics in the distributions in
Figure 1. An appropriate econometric analysis should capture these. Moreover, researchers
might be interested not just in modeling such features in the historical record: they might
seek also to project these measured tendencies forwards from the observed sample. In
Figure 1, what if t + s is some time in the future? The econometric analysis should provide
a model that allows such calculations.
How does this generate new econometrics? Simply tracking the moments of the cross-
sectional distributions in Figure 1 will typically shed no light on many of the characteristics
I have just described. Similarly uninformative, for Figure 1, would be giving extensive
tabulations of the univariate time-series behavior of each of the underlying cross-sectional
units or indeed of documenting the multi-variate time-series characteristics of selected
subsets of those cross-sectional units. Cross-sectional and panel data regressions, if all
they do is capture the behavior of a conditional average, will be altogether unrevealing for
the dynamics of the entire cross-section distribution.2 What a researcher needs to do is to
analyze those evolving distribution dynamics directly.
Formulating the problem of economic growth in the form of Figure 1 draws an equiva-
lence between the analysis of growth and of distribution.3 It is not that higher growth can
cause or, alternatively, be driven by greater inequality, but rather the two are considered si-
multaneously. Note, however, that the distribution that is relevant here is the distribution of
income across countries, not that within a given economy. Thus, the problem considered in
this study differs from the classical set of questions prominently considered by Kuznets and
subsequently refined in Benabou (1996b), Galor and Zeira (1993), Persson and Tabellini
(1994), and many others.
From the perspective of economic growth empirics, the work described below relates to
research using convergence predictions to distinguish endogenous and neoclassical growth.
That literature is large, but helpfully summarized in Barro and Sala-i-Martin (1995) and
Sala-i-Martin (1996). However, some papers have argued that that growth and convergence
literature is uninformative for whether poor countries are catching up with rich ones, and
unrevealing in general for the dynamics of the distribution of welfare across countries (e.g.,
30 QUAH

Friedman 1992, Leung and Quah 1996, Quah 1993a, b, 1996b, c, f).4 Such ambiguity, on
the other hand, cannot taint the analysis following from Figure 1.
Finally, independent of macroeconomic analyses of aggregate growth, the study of dis-
tributions and their dynamics has long been a central part of economic analysis, not just
of personal incomes (e.g., Atkinson 1995; Cowell, Jenkins, and Litchfield 1996; Durlauf
1996; Esteban and Ray 1994; Loury 1981; Schluter 1997; and Shorrocks 1978; among
others) but also of many other economic categories including earnings, firm and industry
shares, regional economic performance, and occupational categories (e.g., Lamo 1996, Lil-
lard and Willis 1978, Lucas 1978, Konings 1994, Koopmans 1995, Quah 1996a, Singer and
Spilerman 1976, and Sutton 1995).
While the current work shares ideas with all of these, it also differs in a number of
significant ways. But those putative contributions will be easier to see at the end rather than
the beginning of the paper.
The remainder of the presentation then is organized as follows: Section 2 puts empirical
flesh on the caricature given in Figure 1: the shape and mobility dynamics sketched in
Figure 1 are broken down further into density and Tukey box plots, and stochastic kernels.
A simple illustrative model is given in Section 3—the model is highly stylized; its purpose
is only to suggest the kinds of conceptual modeling issues that will be further helpful.
Section 4 builds on those ideas and illustrates the role of conditioning in explaining the
distribution dynamics documented in Section 2. Section 5 summarizes the conclusions
from this study.

2. A First Empirical Analysis

Figure 2 plots the log of per capita incomes across 105 countries, all relative to the world
average per capita income in each year. The underlying data are drawn from the well-known
Summers-Heston (1991) dataset.
On the vertical axis in the figure, zero indicates equality with the world average. Time
proceeds sequentially along the axis marked Year. Along the axis marked Economy are the
different countries. The particular ordering on this axis gives no insight. Nor will it be used
below. For the record, however, the ordering is alphabetical within continents, beginning in
Africa with Algeria and Angola, and ending in Oceania with Vanuatu and Western Samoa.
To relate Figure 2 to Figure 1, observe that at each point on the Year axis, one can slice
across the graph, parallel to the Economy axis, and recover the point-in-time cross-country
income distribution. I have computed Figure 3.1 and Figure 3.4 doing exactly that.

2.1. Shapes

If I did no more than this, however, I would, of course, have lost important dynamic,
intradistribution information—I will return to this point below. For the time being, this pro-
cedure gives a sequence of snapshots of the resulting income distributions across countries.
Figure 3.1 and Figure 3.4 provide two views of these cross-sectional distributions. The
first, Figure 3.1, is a sequence of kernel-smoothed densities taken at roughly decade-long
EMPIRICS FOR GROWTH AND DISTRIBUTION 31

Figure 2. (Log) Relative per capita incomes across 105 countries.

intervals. The second, Figure 3.4, is a sequence of Tukey boxplots for the same underlying
data and for the same time periods.
A quick word on inference is useful here: Figure 3.1 and Figure 3.4 record properties of
the population. The data that go into these figures cannot be interpreted as a random sample.
In the language of Efron and Tibshirani (1993), these figures are direct representations of
a census: they are not pictures of a random sample from which statistical analysis can
help us infer properties of the true underlying population. These pictures already are that
population. Thus, from the perspective of statistical inference, classical random sampling
assumptions do not hold. Below, when I turn to models of endogenous cross-sectional
interaction, we will see that the departure here from a classical sampling framework goes
even deeper.
The kernel-smoothed estimates in Figure 3.1 were obtained using a Gaussian kernel.5 By
how the data are defined, 1/2 on the horizontal axis indicates one-half the world average
per capita income; 2 indicates twice the world average; and so on. Looking across three
decades, we see that in 1961 a nascent twin-peakedness—the first mode at a little less than
1, the second at slightly greater than 2.5—was beginning to be visible. By 1988 that second
peak had become pronounced. The relative income distance between the peaks doubled
from about 1.5 in 1961 to more than 3 in 1988. Finally, for the observations extant, these
tendencies appear monotone: the data show no reversals in the dynamics just described.6
For completeness, I give in Figure 3.2 and Figure 3.3 two other related snapshot density
sequences. In Figure 3.2 the distributions are in natural logs of per capita income; in
32 QUAH

Figure 3.1. Densities of relative (per capita) incomes across 105 countries.

Figure 3.3 the distributions are weighted by the relative numbers of people in each economy.
One convenient interpretation of the second is that it shows the distributions of individual
incomes across people in the world, assuming that within each economy individual personal
incomes are equally distributed, and thus equal to the level of per capita income. Properly
interpreted, the “emerging twin peaks” character remains, but is modified. In logs, the peaks
are closer together—as one would expect—but the rise of the richer peak at the expense
of the poorer remains pronounced. Weighted by populations, the distribution sequence
shows three peaks, rather than two: the rise of the higher peaks appears to be at the expense
of the middle (valley) group. Thus, although details differ, the principal message of the
“Emerging twin peaks” Figure 1 comes through in a range of perturbations on the empirical
analysis.
EMPIRICS FOR GROWTH AND DISTRIBUTION 33

Figure 3.2. Densities of log relative (per capita) incomes across 105 countries.

Figure 3.4 comprises Tukey boxplots constructed from exactly the same data used in
Figure 3.1. To understand these pictures, recall the construction of a Tukey boxplot.7 The
box in the middle of each boxplot describes central tendencies of a distribution: the thin
line inside the box locates the median; the top and bottom edges are the 75th and 25th
percentiles respectively. The middle 50% of the distribution is thus contained in the box;
the height of the box—ignoring its vertical location—is the interquartile range.
In Figure 3.4 the middle box for each of the years grows in extant: thus, the middle 50% of
the cross-section distribution can be covered only by progressively larger portions of income
space. Put differently, the middle 50% of the distribution is spreading out; or, when taken
together with the evidence in Figure 3.1, the middle-income class is vanishing—precisely
as in Figure 1.
34 QUAH

Figure 3.3. Densities of relative (per capita) incomes across 105 countries. Weighted by population.

Emanating from the middle box in each Tukey boxplot are rays reaching to upper and
lower adjacent values. If the interquartile range is r , then the upper adjacent value is the
largest income value observed no greater than the 75th percentile plus 1.5 × r . The lower
adjacent value is similarly defined, extending downwards from the 25th percentile. Indicated
by asterisks in Tukey boxplots are upper and lower outside values—observations that lie
outside the upper and lower adjacent values. From a statistical perspective, these might be
considered outliers—in the current application, however, these denote the macroeconomies
that have performed extraordinarily well or extraordinarily poorly relative to the bulk of
other macroeconomies. They represent real people and real countries, not just observations
that might be useful to delete in a statistical analysis.
Figure 3.4 shows no extraordinarily poorly-performing economies—or, more accurately,
EMPIRICS FOR GROWTH AND DISTRIBUTION 35

Figure 3.4. Boxplots, relative (per capita) incomes across 105 countries.

when economies performed especially badly, they were not alone. On the upside, by
contrast, the early part of the sample showed several outstanding performers: there is a
sprinkling of asterisks in the early boxplots. However, over time, parts of the rest of the
world have caught up with these initially very rich economies, even as other parts of the
world remained poor. Unlike the upper portion, the lower part of the boxplot has never
risen, and, indeed, relative to the median shows a continuing decline.
These two descriptions Figure 3.1 and Figure 3.4 have fleshed out and confirmed the
shape dynamics sketched in the twin-peaks picture Figure 1. We turn now to the mobility
dynamics also depicted in Figure 1 but not yet examined in the data.

2.2. Mobility

An easy way to quantify churning or intradistribution dynamics in a sequence of distributions

is to discretize the space of income values, and then simply count the observed transitions
out of and into distinct discrete cells. For instance, in Figure 4—which reproduces the
36 QUAH

Figure 4. Discretization for intradistribution dynamics.

essential features of Figure 1—one might add up the number of transitions out of cell II
into cells I and III respectively (and everywhere else), and then normalize those counts by
the total number of observations. Using discrete cells that span the space of all possible
realizations, one can then construct a transition probability matrix (as in, e.g., Quah 1993a,
b).
It is well known, however, that such a discretization can distort dynamics in important
ways when the underlying observations are continuous variables (see, e.g., Chung 1960).
The solution is not to use a discretization at all, but to retain the original set of continuous
income observations in quantifying intradistribution dynamics. Doing that is like allowing
the number of distinct cells {I, II, III, . . . } in Figure 4 to tend to infinity and then to
the continuum. The corresponding transition probability matrix tends to a matrix with a
continuum of rows and columns. In other words, it becomes a stochastic kernel, as graphed
in Figure 5.1 (in Section 4 below, I provide a more precise technical derivation of a stochastic
kernel).
The figure shows the stochastic kernel for 15-year transitions in our relative-income data,
averaging over 1961 through 1988. From any point on the axis marked Period t extending
parallel to the axis marked Period t + 15 the stochastic kernel is a probability density
function: the projection traced out is nonnegative and integrates to unity. That projection
is similar to a row of a transition probability matrix: such a row has all entries nonnegative
and summing to 1. Roughly speaking, this probability density describes transitions over 15
years from a given income value in period t.
A graph such as Figure 5.1 shows how the cross-sectional distribution at time t evolves
into that at t + 15. If most of the graph were concentrated along the 45-degree diagonal,
then elements in the distribution remain where they began. If, by contrast, most of the
EMPIRICS FOR GROWTH AND DISTRIBUTION 37

Figure 5.1. Relative income dynamics across 105 countries.

mass in the graph were rotated 90 degrees counter-clockwise from that 45-degree diagonal,
then substantial overtaking occurs—the rich become poor, and the poor rich, periodically
over 15-year horizons. If most of the graph were concentrated around the 1-value of the
Period t+15 axis—extending parallel to the Period t axis—then over a 15-year horizon, the
cross-section distribution converges towards equality. Generalizing this, if most of the mass
located parallel to the Period t axis, with projections on period t-values equal to each other,
then the kernel is one where a single (15-period) iteration takes any initial distribution to the
same long-run distribution. Dynamics over longer horizons can be studied by recursively
applying a given stochastic kernel.8
In Figure 5.1 a twin-peaks property again manifests. Over the 15-year horizon, a large
portion of the probability mass remains clustered around the main diagonal. However,
along that principal ridge, a dip appears in the middle-income portion while the kernel
itself rises towards local maxima in both poor and rich parts of the income range. Contour
plot Figure 5.2 makes this clearer. The two peaks in the stochastic kernel—because they
sit (almost) on the 45-degree diagonal—correspond to what Durlauf and Johnson (1995)
call “basins of attraction.” At the same time, however, while the middle-income class is
vanishing, portions of the cross section do transit from high to low and from low to high:
the stochastic kernel is positive almost everywhere, and communicates across the entire
range of income values.
38 QUAH

Contour plot at levels 0.2, 0.35, 0.5

Figure 5.2. Relative income dynamics across 105 countries.

Figures 5.3, 5.4, 5.5, and 5.6 provide comparable stochastic kernel representations on,
respectively, the log and population-weighted versions of the income distributions. In both
cases, the message from the unweighted per-capita income is amended but not overturned.
The twin peaks are closer together in the log case (as in just the snapshot density sequence
Figure 3.2), but the stochastic kernel again shows clearly the polarization dynamics. In the
population-weighted case, the multiplicity of peaks is again evident.
To sum up this first pass through the data, we conclude that the data confirm most of
the stylized features earlier given in Figure 1. There is a wide spectrum of intradistri-
bution dynamics—overtaking and catching-up occur simultaneously with persistence and
languishing—while overall the twin-peaks shape in the cross-sectional distribution emerges.

3. A Simple Model and Empirical Implications

What might explain these “emerging twin-peaks” regularities? In particular, what theoret-
ical models and further empirical analyses will shed light on these stylized facts?
Given this assignment, it is hard to see what insights will obtain from a conventional ap-
proach: study standard “growth and convergence” models of representative economies, and
then analyze such models using panel-data econometric methods that absorb heterogeneity
into “individual effects.” Sure, those techniques deal with data that show rich cross-section
EMPIRICS FOR GROWTH AND DISTRIBUTION 39

Figure 5.3. Log relative income dynamics across 105 countries.

and time-series variation (as in the uninformative Figure 2), but that fact alone does not
recommend them.9 Individual-effects panel data methods had been developed to take into
account the inconsistency in estimating regression coefficients when unobserved hetero-
geneity is correlated with regressors (Chamberlain, 1984, makes this particularly clear).
They were not designed to naturally provide a picture of how an entire distribution evolves.
Those regression methods average across the cross section: they can give only a picture of
the behavior of the conditional mean, not of the whole distribution. Moreover, sweeping
out individual heterogeneities, in the current application, amounts to no more than resigning
oneself to leave unexplained the (significant) differences across individual countries. But
this is precisely what we wish to understand here.
Instead, it is reasonable to guess that more insightful for studying twin-peaks behavior will
be theorizing directly in terms of the entire distribution, and permitting explicit patterns
of cross-section interaction—clustering together into distinct clumps—to endogenously
emerge. Examples of such models exist. They include those in Durlauf (1993), Ioannides
(1990), Kirman, Oddou, and Weber (1986), Townsend (1983), and Quah (1996d). Here, I
present a simplified version of the model in Quah (1996d); it is stripped down to an extent
that allows insight into key issues but at some cost in rigor and economic motivation.10
Let J be the index set of economies, taken as fixed throughout the discussion. A coalition
of economies is a subset C of J . Each economy l in J is characterized by an economy-
40 QUAH

Contour plot at levels 0.2, 0.35, 0.5

Figure 5.4. Log relative income dynamics across 105 countries.

specific stock h l , which can be interpreted as human capital. This stock is used in two nonri-
val ways: first, it represents the potential for technical progress and ongoing growth—it is the
source of useful ideas. Second, it produces nonstorable output for current consumption—it
is an input in a production technology.
Production occurs from coalitions of economies forming to jointly produce a single non-
durable consumption good. Denote the total output of coalition C by YC . Assume that YC
depends on the distribution of h l across l in C, and is increasing in each h l . Assume also
that out of total coalition output, economy l in C gets portion ψ(YC , h l ), with ψ increasing
in both arguments, and satisfying exact product exhaustion:
X
ψ(YC , h l ) = YC .
l in C

(Primitive assumptions implying these properties would be first, compensation according

to marginal product and second, the CES technology
" #1/θ
X
θ
YC = hl , 0 < θ < 1,
l in C

with θ, describing the elasticity of substitution in the CES production function, giving
isoquants between linear and Cobb-Douglas technologies. Quah (1996d) gives the natural
interpretation of these properties as economies of scale deriving from specialization.)
EMPIRICS FOR GROWTH AND DISTRIBUTION 41

Figure 5.5. Relative income dynamics across 105 countries, weighted by population.

By these assumptions, enlarging the coalition always increases total output YC . The
compensation scheme ψ then ensures that all economies unanimously agree to be in the
single grand coalition comprising the entire cross section. This, therefore, is a force for
consolidation. If this were all there were at work, the cross-sectional interaction would
be trivial: the only coalition that exists includes simultaneously all members of the cross
section J .
Turn now to the dynamics of h l . Denote for each coalition C the average value of h l ’s
across C by HC . Suppose that human capital in economy l in C evolves as

ḣ l = φ̃(h l , HC ) for l in C,

with φ̃ increasing in both arguments and homogeneous degree 1. This says simply that
human capital in economy l accumulates not just from the human capital already extant in
it, but also from the average human capital extant in the economies with which l interacts.11
Dividing by h l , this becomes the proportional growth equation:

def
ḣ l / h l = φ̃(1, HC / h l ) = φ(HC / h l ).

By construction φ is increasing in the ratio HC / h l .

42 QUAH

Contour plot at levels 0.2, 0.35, 0.5

Figure 5.6. Relative income dynamics across 105 countries, weighted by population.

It is now easy to see the force for fragmentation, and against the single grand coalition
forming. Economies in higher average h coalitions have faster proportional growth rates.
The problem with allowing a coalition to get too large is that the coalition then (typically)
lowers its average HC : this would slow growth for all the economies already in the coalition.
Economies already in good coalitions would, ceteris paribus, refuse to admit economies
that lower the coalition average HC .
The force for consolidation (the compensation ψ(YC , h l )) is a level effect—it affects cur-
rent consumption. The force for fragmentation (the growth φ(HC / h l )) is a slope effect—
it affects future consumption. Parameterizing economies’ discount rates for intertempo-
ral consumption allows calibrating the tradeoff across level and slope effects, and thus
provides a theory of endogenous coalition formation. Equilibrium is a set of coalitions
{C1 , C2 , C3 , . . .} such that no economy assigned to a coalition wishes to belong to a differ-
ent coalition agreeing to admit it. Quah (1996d) describes an equilibrium that comprises
nontrivial consecutive subsets of the cross section J of economies. Then, as shown in Fig-
ure 6.2, the distribution of incomes across economies within the same coalition converges
towards equality; those across different coalitions separate and then diverge.
The equilibrium distribution dynamics that arise depends on the functions φ and ψ and,
significantly, also on initial conditions in the distribution of h. If, the initial distribution
were not that given in Figure 6.2, but instead that of Figure 6.1, then the model implies
EMPIRICS FOR GROWTH AND DISTRIBUTION 43

Figure 6.1. Convergence, one-world.

Figure 6.2. Convergence, two-coalition polarization.

convergence of the entire cross-section, not subsets, to a single degenerate point mass.
The ideas here can be enriched in a variety of ways: if what mattered were some multidi-
mensional attribute and not just the single-dimensioned h, then equilibrium coalitions need
not be consecutive (as in Figure 6.1 and Figure 6.2), but might intermesh in some form of
a seamless web. If important stochastic disturbances perturbed each economy’s develop-
44 QUAH

ment process, then overtaking and criss-crossing across coalitions and different parts of the
income distribution might occur.
The basic insight, however, remains: by studying interactions across the cross-section,
one gets a picture of how the entire distribution evolves through time. Clustering—the
polarization or stratification that emerges in Figure 6.2—into convergence clubs then man-
ifests as a central part of the economic reasoning. For empirical analysis, it is useful to note
that these endogenous cross-section interactions result in violation of the classical random
sampling assumptions that are traditionally adopted in work with cross-sectional data.
In general, the coalitions that form in equilibrium might be only implicit: no formal
observable organization need be arranged to house them. An equilibrium such as Figure 6.2
is to a degree already consistent with the emerging twin-peaks stylized facts previously
discussed. But can empirics shed further light on the dimensions along which such coalitions
(implicitly) form? I turn next to this.

4. Conditioning

The emerging twin-peaks picture in Figure 1 is an instance of unconditional dynamics.

What the analysis leading up to Figure 6.2 has done is provided a set of predictions largely
consistent with those unconditional dynamics. Put another way, Figure 1 is a reduced form
consistent with one particular structure—that described in the previous section. The goal
of this section is to obtain independent, auxiliary evidence that might shed more light on
the link between the underlying structure and the reduced-form distribution dynamics of
Figure 1. Such an exercise might even be regarded as explaining distribution dynamics.
Note that to explain distribution dynamics, in the sense used here, is more involved than,
say discovering a particular coefficient to be significant in a regression of a dependent
variable on some right-hand side variables. What we seek is an empirical computation that
helps us understand the law of motion in an entire distribution.
The key insight is to turn on its head the reasoning behind the “unconditional dynamics”
already given. Just as stochastic kernels quantify how distributions evolve through time,
they can also describe how a set of conditioning factors alters the cross-sectional distribution
of income. Thus, to understand if a hypothesized set of factors explains emerging twin-
peaks features, we can simply ask if the stochastic kernel transforming the unconditional
distribution to a conditional one removes those same features.
To see how this works in detail, recall the technical derivation of a stochastic kernel.
Since we are concerned here with real-valued incomes, the underlying state space is the
pair (R, R), i.e., the real line R together with the collection R of its Borel sets. Let B(R, R)
denote the Banach space of bounded finitely-additive set functions on the measureable space
(R, R) endowed with total variation norm:
X
∀µ in B(R, R): |µ| = sup |µ(A j )|,
j

where the supremum in this definition is taken over all {A j : j = 1, 2, . . . , n} finite measur-
able partitions of R.
EMPIRICS FOR GROWTH AND DISTRIBUTION 45

Empirical distributions on R can be identified with probability measures on (R, R); those
are, in turn, just countably-additive elements in B(R, R) assigning value 1 to the entire
space R. Let B denote the Borel σ -algebra generated by the open subsets (relative to total
variation norm topology) of B(R, R). Then (B, B) is another measurable space.
Note that B includes more than just probability measures: an arbitrary element µ in B
could be negative; µ(R) need not be 1; and µ need not be countably-additive. On the other
hand, a collection of probability measures is never a linear space: that collection does not
include a zero element; if λ1 and λ2 are probability measures, then λ1 − λ2 and λ1 + λ2 are
not; neither is xλ1 a probability measure for x ∈ R except at x = 1. By contrast, the set of
bounded finitely-additive set functions certainly is a linear space, and as described above,
is easily given a norm and then made Banach.
Why embed probability measures in a Banach space as we have done here? A first reason
is so that distances can be defined between probability measures; it then makes sense to
talk about two measures—and their associated distributions—getting closer to one another.
A small step from there is to define open sets of probability measures, and thereby induce
(Borel) σ -algebras on probability measures. Such σ -algebras then allow modeling random
elements drawn from collections of probability measures, and thus from collections of
distributions. The data of interest when modeling the dynamics of distributions are precisely
random elements taking values that are probability measures.
This framework allows a more rigorous description of the stochastic kernels already used
above. Let Ft denote the distribution of incomes across economies at a given time t.
Associated with Ft is a measure λt in (B, B). If (Ä, F, Pr) is the underlying probability
space, then λt is the value of an F/B-measureable map 3t : (Ä, F) → (B, B). The
sequence {3t : t ≥ 0} is then a B-valued stochastic process.
How should the law of motion for such a process be modeled?
The simplest scheme for doing so is analogous to the first-order autoregression from
standard time-series analysis:

λt = T ∗ (λt−1 , u t ) = Tu∗t (λt−1 ), t ≥ 1,

where T ∗ is an operator that maps the product of measures together with generalized
disturbances u to probability measures; and Tu∗t absorbs the disturbance into the definition
of the operator. (Why the ∗ appears in T ∗ and Tu∗t will be clarified below.) This is no more
than a stochastic difference equation taking values that are entire measures; equivalently, it
is an equation describing the evolution of the distribution of incomes across economies.
To understand the structure of operators like Tu∗t , it helps to use the following:

Stochastic kernel definition: Let µ and ν be elements of B that are probability measures
on (R, R). A stochastic kernel relating µ and ν is a mapping M(µ,ν) : (R, R) → [0, 1]
satisfying:

(i) ∀y in R, the restriction M(µ,ν) (y, ·) is a probability measure;

(ii) ∀A in R, the restriction M(µ,ν) (·, A) is R-measurable;
R
(iii) ∀A in R, we have µ(A) = M(µ,ν) (y, A)dν(y).
46 QUAH

To see why this is useful, first consider (iii). In an initial period, for given y, there is
some fraction dν(y) of economies with incomes close to y. Count up all economies in that
group who turn out to have their incomes subsequently fall in a given R-measurable subset
A ⊆ R. When normalized to be a fraction of the total number of economies, this count is
precisely M(y, A) (where the (µ, ν) subscript can now be deleted without loss of clarity).
Fix A, weight R the count M(y, A) by dν(y), and sum over all possible y’s, i.e., evaluate
the integral M(y, A)dν(y). This gives the fraction of economies that end up in state A
regardless of their initial income levels. If this equals µ(A) for all measurable subsets A,
then µ must be the measure associated with the subsequent income distribution. In other
words, the stochastic kernel M is a complete description of transitions from state y to any
other portion of the underlying state space R.
Conditions (i) and (ii) simply guarantee that the interpretation of (iii) is valid. By (ii), the
right hand side of (iii) is well-defined as a Lebesgue integral. By (i), the right hand side of
(iii) is a weighted average of probability measures M(y, ·), and thus is itself a probability
measure.
How does this relate to the structure of Tu∗t ? Let b(R, R) be the Banach space under sup
norm of bounded measurable functions on (R, R). Fix a stochastic kernel M and define
the operator T mapping b(R, R) to itself by
Z
∀ f in b(R, R), ∀y in R: (T f )(y) = f (x)M(y, d x).

Since M(y, ·) is a probability measure, the image T f can be interpreted as a forwards

conditional expectation. For example, if all economies in the crossR section begin with
incomes y, and we take f to be the identify map, then (T f )(y) = x M(y, d x) is next
period’s average income in the cross section, conditional on all economies having income
y in the current period.
Clearly, T is a bounded linear operator. Denote the adjoint of T by T ∗ . By Riesz
Representation Theorem, the dual space of b(R, R) is just B(R, R) (our original collection
of bounded finitely additive set functions on R); thus T ∗ is a bounded linear operator
mapping B(R, R) to itself. It turns out that T ∗ is also exactly the mapping in (iii) of the
Stochastic Kernel Definition, i.e.,
Z
∀ν probability measures in B, ∀A in R: (T ∗ ν)(A) = M(y, A)dν(A).

(This is immediate from writing the left side as

Z Z
(T ∗ ν)(A) = 1 A d(T ∗ ν)(y) = (T 1 A )(y) dν(y) (adjoint)
Z ·Z ¸
= 1 A (x)M(y, d x) dν(y) (definition of T )
Z
= M(y, A) dν(y), (calculation)

with 1 A the indicator function for A.)

EMPIRICS FOR GROWTH AND DISTRIBUTION 47

The definition of a stochastic kernel nowhere requires that ν and its image µ under T ∗ be
sequential in time. Thus, a stochastic kernel M representing T ∗ can be used to relate any
two different distributions, in particular an unconditional observed distribution, and one
conditional on a set of explanatory factors.
To sharpen the focus, we return to the model of cross-sectional interaction given in Sec-
tion 3. In that model, while under certain conditions stratification emerges, as in Figure 6.2,
that feature is absent if we consider each economy only in relation to the other members
of its implicit coalition. In particular, conditioning each economy’s observations on the
behavior of its neighbors, the distribution dynamics imply convergence to a degenerate
point mass. This motivates the following:

Conditioning scheme definition: For a collection of economies J , a conditioning scheme

S is a collection of triples, one for each economy l in J at time instant t, with each triple
comprising:

(i) τl (t) an integer lag,

(ii) Cl (t) a subset of J , and

(iii) ω̄l (t) a set of probability weights on J never positive outside Cl (t).

The subset Cl (t) is the collection of economies associated with, or neighbors of, l at time t.
Since the weights ω̄l (t) can be positive only on Cl (t), they determine the relative importance
of the different economies in Cl (t) in the evolution of economy l at time t. Finally, τl (t) is
the lag in time: it indicates the delay with which developments in the economies in Cl (t)
affect l. Since τl (t) can be a positive constant, l’s associated neighbors Cl (t) can include l
itself.
Sometimes, it will be convenient to use an arbitrary scaled version of the probability
weights ω̄l (t), i.e., an appropriate collection of nonnegative numbers having a nonzero
sum. Such an alternative collection will be denoted ωl (t). The probability weights ω̄l (t) of
a conditioning scheme can always be constructed from ωl (t), so that referring only to the
latter is without loss.
If Y = {Yl (t): l in J and t ≥ 0} denotes the original observations on (relative) per capita
incomes, define the conditional version Y | S = Ỹ by
def
Ỹl (t) = Yl (t)/Ŷl (t)

where
def
X
Ŷl (t) = ω̄ j (t)Y j (t − τl (t)).
j in Cl (t)

In words, Y | S comprises per capita incomes relative to one’s neighbors, appropriately

weighted.12
Some special cases will help develop intuition. Fix τl (t) = τ > 0; take Cl (t) = {l},
i.e., just the economy itself; and define ωl (t) = {1 on Cl (t) and 0 elsewhere}. Then the
48 QUAH

Figure 7.1. Stochastic kernel, relative (per capita) incomes across 105 countries.

stochastic kernel mapping the unconditional distribution in Y to the conditional Y | S is

just the usual Markov stochastic kernel for τ -period transitions. It is this, for τ = 15, that
is graphed in Figure 5.
For the second example, fix τl (t) = 0; take Cl (t) = Cl (0) to be the set of economies
physically contiguous with l but excluding l itself; and let ωl (t) but the set of population
sizes at time t − τl (t) = t in each of the economies in Cl (0). Then Ŷl (t) is the per capita
income in all contiguous economies abutting on l; and Ỹl (t) is per capita income in l
relative to that in the surrounding economies. In words Y | S is per capita income relative
to one’s physical neighbors. The stochastic kernel mapping Y → Y | S would depart
from the identity map if the convergence clubs of Section 3 were spatially concentrated:
most interaction and exchange occurred within groups of countries physically close to one
another.13
Figure 7.1 displays precisely that stochastic kernel. The graph here has the same inter-
pretations as previously given for Figure 5.1 except that the axes are now Original and
Conditioned rather than Period t and Period t + 15. (For completeness, I also give Fig-
ure 7.2 and Figure 7.3 showing the densities and boxplots for Y | S.) The most prominent
change comparing stochastic kernels in Figure 7.1 and Figure 5.1 is the counterclockwise
shift in mass to parallel the Original axis. Put differently, spatial factors account for a large
part of the distribution of incomes across countries: rich economies are typically close
EMPIRICS FOR GROWTH AND DISTRIBUTION 49

Figure 7.2. Densities of relative (per capita) incomes across 105 countries.

to—interact more with—other rich ones; similarly poor economies are typically close to
other poor ones.
The snapshot densities for Y | S displayed in Figure 7.2 and Figure 7.3 no longer show
emerging twin-peaks features. In summary, it appears that the polarization earlier identified
in the unconditional distribution-dynamics of cross-country incomes is well explained by
physical geography. Not only are rich countries located close to other rich ones, such
tendencies have magnified through time.
However, while physical geography has been just shown to play an important role
in cross-country income distribution dynamics, it is likely international trade that most
economists would identify as the interactions of Section 3. This is easily incorporated in
the conditioning-schemes analysis, and constitutes the third example. Let τl (t) = 0 so that
50 QUAH

Figure 7.3. Boxplots, relative (per capita) incomes across 105 countries.

there is no delay. Fix T to be the end period of the observed sample, and take Cl (t) = Cl (T )
to be that set of l’s trading partners at time T such that their total trade (imports plus exports)
share is at least 50% of l’s total trade at T . Finally, let ωl (t) = ωl (T ) be the measured trade
shares of the different economies in Cl (t) out of l’s total trade accounts. The stochastic ker-
nel mapping Y → Y | S now describes the importance of trade in explaining cross-country
income distribution dynamics.14
Figure 8.1 displays the stochastic kernel for Y → Y | S with S trade conditioning.
Here, the counterclockwise twist in the kernel towards the vertical is even more pronounced
than in Figure 7.1: rich countries trade mostly with other rich ones; and, interestingly, the
very poorest countries, mostly with rich ones again. The way to see this is to notice that
the stochastic kernel in Figure 8.1 clusters about 1 on the high end of the Original scale,
but about a value less than 1/2 on the low end. Figure 8.2 and Figure 8.3 give the densities
and boxplots for Y | S. In Figure 8.2 the second peak that emerges in 1988 is at the
average 1. Figure 8.3 shows the conditional distributions to be relatively compact (i.e.,
EMPIRICS FOR GROWTH AND DISTRIBUTION 51

Figure 7.4. Spatial-conditioned relative income dynamics.

have their support relatively narrow); the vanishing of the middle-income class is no longer
visible.
The conditioned income distributions give information on dynamics as well. Figure 7.4
and Figure 8.4 provide stochastic kernel representations on 15-year transitions in space-
and trade-conditioned incomes. For space, Figure 7.4 shows a marked improvement in
convergence possibilities—the poor catching up with the rich—except at the very highest
income levels: the stochastic kernel is concentrated parallel to the Period t axis on the
average value of 1. For trade, however, that increase in convergence dynamics is most
obvious only for middle-income countries.
To summarize, this section has provided a set of empirical computations designed to ex-
plain the emerging twin-peaks dynamics earlier documented. Using the idea of an endoge-
nously determined set of cross-section neighbors being important, this section developed
the notion of conditioning schemes for empirical use. Applied to space and trade, we see
a first glimpse of how important and large such cross-sectional interaction effects might
be. The importance of trade here is not just expressed in blanket measures of how open an
economy is; emphasized instead are patterns of who trades with whom.
52 QUAH

Figure 8.1. Stochastic kernel, relative (per capita) incomes across 105 countries.

5. Conclusions

This paper has analyzed patterns of economic growth across countries from the perspective
of distribution dynamics. In doing so, it uncovered empirical regularities—emerging twin
peaks, incipient polarization and stratification—that are hidden to traditional methods of
empirical analysis.
Those distribution-dynamic features call for explanation. This paper has argued that such
explanation is noteworthy in two significant respects: It will (i) differ from conventional
models of growth and accumulation in the direction of theorizing in terms of the entire cross
section distribution, and (ii) depart from standard techniques of econometric analysis—
both in the empirical effects that have to be modeled and in permitting the cross-sectional
interaction that the theoretical reasoning had suggested would matter.
The paper has developed a class of conditional distribution analyses using the idea of
conditioning schemes. These showed the importance of space and trade—endogenous
cross-sectional interaction more generally—for understanding cross-country patterns of
growth. Unlike traditional studies on trade and growth, the study above emphasized not
measures of openness to trade, but instead empirical patterns of who trades with whom.
Although considerable progress has taken place, much remains to be done still in rigorous
theoretical and empirical analyses of such cross-sectional dynamics.
EMPIRICS FOR GROWTH AND DISTRIBUTION 53

Figure 8.2. Densities of relative (per capita) incomes across 105 countries.
54 QUAH

Figure 8.3. Boxplots, relative (per capita) incomes across 105 countries.
EMPIRICS FOR GROWTH AND DISTRIBUTION 55

Figure 8.4. Trade-conditioned relative income dynamics.

56 QUAH

Acknowledgments

I thank Luca Stanca for help, and conference participants at a MacArthur Foundation meet-
ing and the Santa Fe Institute for constructive criticisms on related work. Discussions with
Ken Arrow, Steven Durlauf, Yannis Ioannides, and Chuck Manski have been especially
helpful. This paper was delivered as an Invited Lecture (Economics/Econometrics) at the
1996 Econometric Society European Meetings in Istanbul, where Rob Engle and Harald
Uhlig provided many insightful comments. I have received helpful suggestions from nu-
merous others, including Willem Buiter, Stephen Redding, Tom Sargent, Peter Sinclair,
Jonathan Temple, and Adrian Wood. I thank the British Academy and the MacArthur
Foundation for financial support. All calculations and graphs below were produced using
the author’s econometrics shell tSrF.

Notes

1. Even though I have just argued that interest should not be thus confined, I will continue to use the phrase
“emerging twin-peaks” for two reasons: one, it is a convenient and evocative shorthand; and two, the cross-
country data support the twin-peaks description.
2. The statements in this paragraph should not be taken as intending anything stronger than what they actually
assert. They refer to the features of Figure 1, no more and no less.
3. Bliss (1996) and Quah (1996b) have also taken this perspective, although without the cross-sectional interaction
that will figure prominently below.
4. Other papers relevant to this debate include Ben-David (1994), Bernard and Durlauf (1996), Canova and
Marcet (1995), Desdoigts (1996), Durlauf and Johnson (1995), Galor (1996), and Jones (1997).
5. I took the data nonnegativity into account following the procedure and automatic bandwidth choice given in
Silverman (2.10 and 3.4.2, 1986).
6. Bianchi (1995)—using bootstrap tests for multimodality related to ideas in Izenman and Sommer (1988) and
Silverman (1981, 1983)—and Paap and van Dijk (1994)—using density mixture techniques—have provided
statistical descriptions on this sequence of pictures, earlier given in Quah (1993b). Jones (1997) presents, in
essence, the same picture. Cowell, Jenkins, and Litchfield (1996) have noted similar twin-peakedness in UK
personal income distributions.
7. Although this is relatively unused in economics and econometrics, it is a familiar textbook object in statistics;
see, e.g., Cleveland (1993).
8. Of course, there is no logical necessity why only 15-year horizons need be considered. Moreover, other
statistics (i.e., real-valued functionals) of a stochastic kernel might also be considered, beyond just the visual
description given in the text. For further exploration of these issues, see Durlauf and Johnson (1994), and
Quah (1993a, b, 1996a, b) who studied kernels estimated over varying horizons, and ergodic characterizations,
mobility indexes, and first passage-times calculated off estimated kernels. Desdoigts (1994), Lamo (1996),
and Schluter (1997) have used related ideas in empirical research.
9. See, however, Islam (1995) and Nerlove (1996) for an opposing view.
10. Pursuing this cross-section interactions approach reflects a bias based on instinct, not necessarily anything more
rigorous. There might well be simpler explanations for the twin peaks: e.g., distinct groups of economies
having differing preferences, giving rise to differing investment rates, or even just underlying productivity
shocks having a particular, bimodal distribution.
11. It is not essential that it be exactly the average HC that affects φ̃, just that it be some appropriate functional of
the distribution of h’s in the coalition C.
12. If the denominator Ŷl (t) were replaced by the exponential of the conditional expectation of Yl (t) conditioned
on an information set G, then Ỹl (t) is just the exponential of the expectations error log(Yl (t)) − log(Ŷl (t)) =
log(Yl (t)) − E(log(Yl (t)) | G).
EMPIRICS FOR GROWTH AND DISTRIBUTION 57

13. If we generalize beyond cross-country growth, this particular conditioning scheme provides a natural empirical
counterpart to the theoretical effects described in Benabou (1996a), Durlauf (1996), Ioannides (1996), and
others. It complements the empirical analysis of Brock and Durlauf (1995). Quah (1996e) had used a form of
conditioning scheme to study aspects of globalization in Europe.
14. I have also experimented with using only imports or only exports in this definition and with T set to the
beginning or middle rather than the end of the sample: not much changes in the conclusion. The conditioning
scheme described here is more intricate than just measuring the openness of an economy—here, information
on who trades with whom is used. Additional factor content data, as for example in Coe and Helpman (1995) or
Eaton and Kortum (1996), might additionally be exploited to more sharply focus on the learning and technology
components in cross-country interaction.

References

Atkinson, A. B.(1995). Incomes and the Welfare State: Essays on Britain and Europe. Cambridge: Cambridge
University Press.
Barro, R. J., and X. Sala-i-Martin. (1995). Economic Growth. New York: McGraw Hill.
Baumol, W. J. (1986). “Productivity Growth, Convergence, and Welfare,” American Economic Review 76, 1072–
85.
Ben-David, D. (1994). “Convergence Clubs and Diverging Economies,” Working Paper 922, CEPR.
Bénabou, R. (1996a). “Heterogeneity, Stratification, and Growth: Macroeconomic Implications of Community
Structure and School Finance,” American Economic Review 86, 584–609.
Bénabou, R. (1996b). “Inequality and Growth.” In Macroeconomics Annual, Ben Bernanke and Julio Rotemberg
(eds.), Cambridge, MA: NBER and MIT Press. Forthcoming.
Bernard, A. B., and S. N. Durlauf. (1996). “Interpreting Tests of the Convergence Hypothesis,” Journal of
Econometrics 71, 161–174.
Bianchi, M. (1995). “Testing of Convergence: A Bootstrap Test for Multimodality,” Working Paper, Bank of
England.
Bliss, C. J. (1996). “Long-run Wealth Distribution with Random Shocks,” Working Paper 23, Oxford: Nuffield
College.
Brock, W. A., and S. N. Durlauf. (1995). “Discrete Choice with Social Interactions,” Working Paper, Madison:
University of Wisconsin.
Canova, F., and A. Marcet. (1995). “The Poor Stay Poor: Nonconvergence Across Countries and Regions,”
Discussion Paper 1265, CEPR, November.
Chamberlain, G. (1984). “Panel Data.” In Zvi Griliches and Michael D. Intriligator (eds.), Handbook of Econo-
metrics vol. II. Amsterdam: Elsevier North-Holland.
Chung, K. L. (1960). Markov Chains with Stationary Transition Probabilities. Berlin: Springer-Verlag.
Cleveland, W. S. (1993). Visualizing Data. Summit, NJ: Hobart Press.
Coe, D. T., and E. Helpman. (1995). “International R&D Spillovers,” European Economic Review 39, 859–887.
Cowell, F. A., S. P. Jenkins, and J. A. Litchfield. (1996). “The Changing Shape of the UK Income Distribution:
Kernel Density Estimates.” In John Hills (ed.), New Inequalities: The Changing Distribution of Income and
Wealth in the United Kingdom. Cambridge: Cambridge University Press.
Desdoigts, A. (1994). “Changes in the World Income Distribution: A Non-parametric Approach to Challenge the
Neoclassical Convergence Argument,” PhD dissertation, Florence: European University Institute.
Desdoigts, A. (1996). “Determining Development Patterns and “Club” Formation Using Projection Pursuit
Techniques,” Working Paper, ECARE, ULB.
Durlauf, S. N. (1993). “Nonergodic Economic Growth,” Review of Economic Studies 60, 349–366.
Durlauf, S. N. (1996). “A Theory of Persistent Income Inequality,” Journal of Economic Growth 1, 75–93.
Durlauf, S. N., and P. Johnson. (1994). “Nonlinearities in Intergenerational Income Mobility,” Working Paper,
University of Wisconsin, Economics Department.
Durlauf, S. (1995). “Multiple Regimes and Cross-country Growth Behavior,” Journal of Applied Econometrics
10, 365–384.
Eaton, J., and S. Kortum. (1996). “Trade in Ideas: Patenting and Productivity in the OECD,” Journal of
International Economics 40, 251–278.
Efron, B., and R. J. Tibshirani. (1993). An Introduction to the Bootstrap. New York: Chapman and Hall.
58 QUAH

Esteban, J.-M., and D. Ray. (1994). “On the Measurement of Polarization,” Econometrica 62, 819–851.
Friedman, M. (1992). “Do Old Fallacies Ever Die?” Journal of Economic Literature 30, 2129–2132.
Galor, O. (1996). “Convergence? Inferences from Theoretical Models,” Economic Journal 106, 1056–1080.
Galor, O., and J. Zeira. (1993). “Income Distribution and Macroeconomics,” Review of Economic Studies 60,
35–52.
Ioannides, Y. M. (1990). “Trading Uncertainty and Market Form,” International Economic Review 31, 619–638.
Ioannides, Y. M. (1996). “Residential Neighborhood Effects and Inequality,” Working Paper, Economics Depart-
ment, Tufts University.
Islam, N. (1995). “Growth Empirics: A Panel Data Approach,” Quarterly Journal of Economics 110, 1127–1170.
Izenman, A. J., and C. J. Sommer. (1988). “Philatelic mixtures and multimodal densities,” Journal of the American
Statistical Association 83, 941–953.
Jones, C. I. “On the Evolution of the World Income Distribution,” Journal of Economic Perspectives. Forthcoming.
Kirman, A. P., C. Oddou, and S. Weber. (1986). “Stochastic Communication and Coalition Formation,” Econo-
metrica 54, 129–138.
Konings, J. (1994). “Gross Job Flows and Wage Determination in the UK: Evidence from Firm-level Data,” PhD
dissertation, LSE.
Koopmans, R. M. (1995). “Asymmetric Industry Structures: Multiple Technologies, Firm Dynamics, and Prof-
itability,” PhD dissertation, LSE.
Lamo, A. R. (1996). “Cross-section Distribution Dynamics,” PhD dissertation, LSE.
Leung, C., and D. T. Quah. (1996). “Convergence, Endogenous Growth, and Productivity Disturbances,” Journal
of Monetary Economics, 38, 535–547.
Lillard, L. A., and R. J. Willis. (1978). “Dynamic Aspects of Earning Mobility,” Econometrica 46, 985–1012.
Loury, G. C. (1981). “Intergenerational Transfers and the Distribution of Earnings,” Econometrica 49, 843–867.
Lucas, R. E., Jr., (1978). “On the Size Distribution of Business Firms,” Rand Journal of Economics 9, 508–523.
Nerlove, M. (1996). “Growth Rate Convergence: Fact of Artifact?” Working Paper, University of Maryland.
Paap, R., and H. K. van Dijk. (1994). “Distribution and Mobility of Wealth of Nations,” Working Paper, Tinbergen
Institute, Erasmus University.
Persson, T., and G. Tabellini. (1994). “Is Inequality Harmful for Growth?” American Economic Review 84,
600-621.
Quah, D. T. (1993a). “Empirical Cross-section Dynamics in Economic Growth,” European Economic Review 37,
426–434.
Quah, D. T. (1993b). “Galton’s Fallacy and Tests of the Convergence Hypothesis,” The Scandinavian Journal of
Economics 95, 427–443.
Quah, D. T. (1996a). “Aggregate and Regional Disaggregate Fluctuations,” Empirical Economics 21, 137–159.
Quah, D. T. (1996b). “Convergence Empirics Across Economies with (Some) Capital Mobility,” Journal of
Economic Growth 1, 95–124.
Quah, D. T. (1996c). “Empirics for Economic Growth and Convergence,” European Economic Review 40, 1353–
1375.
Quah, D. T. (1996d).“Ideas Determining Convergence Clubs,” Working Paper, Economics Department, LSE.
Quah, D. T. (1996e). “Regional Convergence Clusters Across Europe,” European Economic Review 40, 951–958.
Quah, D. T. (1996f). “Twin Peaks: Growth and Convergence in Models of Distribution Dynamics,” Economic
Journal 106, 1045–1055.
Sala-i-Martin, X. (1996). “Regional Cohesion: Evidence and Theories of Regional Growth and Convergence,”
European Economic Review 40, 1325–1352.
Schluter, C. (1997). “Topics in Distributional Analysis,” PhD dissertation, LSE.
Shorrocks, A. F. (1978). “The Measurement of Mobility,” Econometrica 46, 1013–1024.
Silverman, B. W. (1981). “Using Kernel Density Estimates to Investigate Multimodality,” Journal of the Royal
Statistical Society, Series B 43, 97–99.
Silverman, B. W. (1983). “Some Properties of a Test for Multimodality Based on Kernel Density Estimates.” In
J. F. C. Kingman and G. E. H. Reuter, (eds.), Probability, Statistics, and Analysis. vol. 79. London Mathematical
Society Lecture Note Series Cambridge: Cambridge University Press.
Silverman, B. W. (1986). Density Estimation for Statistics and Data Analysis. New York, NY: Chapman and Hall.
Singer, B., and S. Spilerman. (1976). “Some Methodological Issues in the Analysis of Longitudinal Surveys,”
Annals of Economic and Social Measurement 5, 447–474.
Summers, R., and A. Heston. (1991). “The Penn World Table (Mark 5): An Expanded Set of International
Comparisons, 1950–1988,” Quarterly Journal of Economics 106, 327–368.
EMPIRICS FOR GROWTH AND DISTRIBUTION 59

Sutton, J. (1995). “The Size Distribution of Business, Parts I and II,” Discussion Paper EI/9, 10, STICERD, LSE.
Townsend, R. M. (1983). “Theories of Intermediated Structures,” Carnegie-Rochester Conference Series on
Public Policy 18, 221–272.