Rising wealth inequality: When r-g matters

Hoang Khieu(a)1

(a)
Johannes Gutenberg University Mainz

3rd February 2020

We investigate the e¤ects of the gap between the rate of return on wealth and

the growth rate, r g, on wealth inequality using a small open economy model

featuring heterogeneity in labour income and in the death rate. The heterogeneity

in the death rate implies di¤erent health types, which imply type dependence and

scale dependence. Labour income heterogeneity implies a U-shaped e¤ect of r g on

wealth inequality. Long run wealth inequality increases in r g if the distribution

of the death rate features good-health types and if r g is su¢ ciently larger than a

positive threshold, which is increasing the dispersion of labour income. The share

of good-health types and the r g’s threshold in the Survey of Consumer Finances

are 30:5% and 4:4%, respectively.

JEL Codes: C02, D31, E21

Keywords: r g, wealth inequality, health types, labour income heterogeneity,

consumption growth.

1 Introduction

[Motivation] Wealth inequality in most countries around the world is increasing and this rising

trend greatly concerns policy makers and scientists. It is probably fair to say that the recent

rise in scientists’interest in wealth inequality is due to the enormous public success of Piketty’s

(2014) book on “Capital in the Twenty-First Century”. Many outstanding economists were
1
Gutenburg School of Management and Economics, Jakob-Welder-Weg 4, 55131 Mainz, Germany, fax +
49.6131.39-25588, phone + 49.6131.39-22078, email [email protected]. I would like to thank Klaus Wälde
and seminar participants at the Johannes Gutenberg University Mainz for comments and discussions.

1
inspired by this success and reviewed the book focusing on various aspects of its …ndings

(e.g. Acemoglu and Robinson 2015; Mankiw 2015; Krusell and Smith 2015; Blume and Durlauf

2015; Jones 2015; Fischer 2017).

It seems generally accepted that wealth inequality and its increase is driven by many factors

ranging from e.g. wage inequality and heterogeneity in interest rates and not ending with tax

and transfer policies (Card and Dinardo 2002; Benhabib, Bisin and Zhu 2011; Hubmer et

al. 2019). One hypothesis put forth by Piketty (2014, 2015a,b), which received particular in-

terest, is that a higher gap between the rate of return on wealth r and the growth rate g of

an economy “will tend to greatly amplify the steady-state inequality of a wealth distribution”

(Piketty 2015a, p. 49). The widely accepted intuition behind this conjecture is that higher

r g implies a higher growth rate of wealth relative to labour income, which results in in-

creasing wealth inequality, at least between capitalists and workers. The measure of wealth

inequality embedded in this intuition is apparently scale-dependent. Theoretical support for

the hypothesis comes from Piketty and Zucman (2015, ch. 15.5.4) who study the e¤ect in a

stochastic framework yielding a Pareto distribution. They …nd that “for a given variance of

shocks, steady-state wealth concentration is always a rising function of r g”(p. 1355). Their

analysis exploits the property that multiplicative random shocks generate Pareto distributions.2

Hubmer, Krusell, and Smith (2019) however argue that Piketty’s hypothesis “mostly applies for

the very richest”and that “to understand the bulk and other side of the wealth distribution”,

earnings heterogeneity is important. Acemoglu and Robinson (2015) exploit the panel data on

r g to argue that there is “no evidence of a positive impact of r g on inequality”.

[The open questions] Most theoretical analyses so far have investigated the e¤ects of r g on

wealth inequality in a framework generating a Pareto distribution. While it is widely accepted

that the right tail of empirical wealth distributions obeys a Pareto tail, one well-known issue
2
Benhabib, Bisin and Zhu (2011) and Benhabib and Bisin (2018) also provide detailed background on Pareto
distributions generated from a model featuring multiplicative shocks.

2
(discussed e.g. by Atkinson (2017) in the context of income distribution) is where to …x the start

and the end of the right tail. For instance, only top 10% wealth share of the wealth distribution

in the Survey of Consumer Finances 2010 is well-approximated by a Pareto distribution (Cao

and Luo 2017).

Thus, it seems to be unclear at this point whether the e¤ect of the gap between r and g

is restricted to frameworks that generate a Pareto distribution of wealth. We therefore ask:

Does “r minus g”also play a role in more general setups than those usually employed to obtain

Pareto distributions? And in such general frameworks, under which conditions does “r minus

g”matter for wealth inequality?

[The objectives] It is the objective of this paper to provide a framework that allows to

study how the rate of return on wealth and the growth rate are related to the level of wealth

inequality. We want to work with a simple but ‡exible model that features distributions of

wealth (and labour income) which at the same time allows for obtaining analytically tractable

and empirically meaningful conditions for the e¤ect of r minus g (and other determinants) on

wealth inequality. This will help to formulate clearer predictions about conditions under which

wealth inequality increases.

[The setup] We consider a small open economy with free capital ‡ows and free trade. Time

is continuous. Firms produce a …nal good employing capital and labour. The price of capital

is given by a constant international interest rate. The domestic wage is given by the marginal

productivity of labour. Households supply labour inelastically and save according to the inter-

nationally given interest rate. There is ex-ante heterogeneity in wealth endowment and labour

productivity. Households draw their initial levels of wealth and labour productivity from given

distributions before becoming economically active. The heterogeneity in labour productivity

implies labour income heterogeneity. There is ex-post uncertainty in the time of death. That

is, at any instant of time households face a constant instantaneous probability of death (a death

3
rate). Households independently draw the instantaneous probability of death from some dis-

tribution before becoming economically active. This captures ex-ante heterogeneity in health

condition (heath type hereafter). We then study an optimal consumption-saving problem con-

ditional on wealth endowment, labour productivity, and the draw of the death rate. Despite its

simplicity (or maybe because of it), this framework yields very sharp and empirically testable

predictions about when and with which strength the gap between the rate of return on wealth

and the growth rate matters for an increase in wealth inequality. We describe wealth inequality

by looking at the coe¢ cient of variation of the wealth distribution.

[Findings] We analyze the e¤ects of the gap between the rate of return on wealth and the

growth rate, r g, on wealth inequality in a general framework that features any distributions

of wealth and labour income. Our …ndings on wealth inequality are therefore applicable for not

only Pareto distributions but also any empirical distributions. Our main contribution lies in

the demonstration of the important role of labour income heterogeneity for the e¤ects of the

di¤erence r g on wealth inequality.

In addition to the heterogeneity in labour income, the health types are crucial for the e¤ects

of r g on long run wealth inequality. They are de…ned by the death rate. Bad-health types

draw death rates higher than a threshold while the death rates of good-health types are below

the threshold. The death rate of the average-health type is equal to the threshold, which

implies consumption growth is equal to output growth. The higher the death rate, the lower

the growth rate of consumption. The bad-health types thus experience consumption growth

lower than output growth while consumption growth of the good-health types exceeds output

growth. Given a distribution of health types with a certain fraction of the good-health types,

long run wealth inequality is a U-shaped function of r g. Speci…cally, long run inequality …rst

decreases as r g increases up to a positive threshold. When r g is bigger than the threshold,

an increase in r g causes long run wealth inequality to rise. Labour income heterogeneity

4
is important for the threshold of r g. In particular, the r g’s threshold increases in the

dispersion of labour income. This implies when labour income distribution is more dispersed,

long run wealth inequality reaches its minimum at a higher value of r g. In the absence of

labour income heterogeneity, long run wealth inequality is always a rising function of r g

(Piketty and Zucman 2015; Jones 2015).

Thus, an increase in the gap r g leads to increasing wealth inequality in the long run

under two conditions. First, the di¤erence r g is su¢ ciently larger than a positive threshold,

which is increasing the dispersion of labour income. Second, there exist good-health types that

experience consumption growth greater than output growth. The presence of good-health types

also helps to explain a stylized fact that most OECD countries experience consumption growth

higher than output growth over the long run (see Section 2). In addition to the dispersion

of labour income, the r g’s threshold also depends on the dispersion of wealth and on the

covariance between wealth and labour income. It is increasing in the covariance between wealth

and labour income if initial wealth is more equally distributed than initial labour income. When

initial wealth is more unequally distributed than initial labour income, the threshold falls as

the covariance between wealth and labour income rises. In particular, the covariance between

wealth and labour income is irrelevant for the threshold when initial wealth and initial labour

income imply an identical coe¢ cient of variation. Using the data on wealth and labour income

from the Survey of Consumer Finances 1989-2016, we …nd that the r g’s threshold is 4:4% on

average, which is empirically plausible. The mean rate of returns to wealth with capital gains

in the Panel Study of Income Dynamics for the years 1984, 1989, and 1994 is 7:92% (Cao and

Luo 2017). When taking into account the average growth rate of 2:6%, the di¤erence r g in

the data would be around a value of 5:32%. We calibrate the distribution of health types to

match the evolution of wealth inequality over the period 1989-2016 and …nd that the share of

the good-health types is 30:5%.

5
 In addition to the di¤erence between the rate of return on wealth and the growth rate, the

share of the best-health type is crucial for long run inequality. The best-health type draws the

smallest death rate and experiences the greatest growth rate of wealth (and consumption). In

the long run, the wealth share owned by this health type dominates the wealth share held by

other health types. Hence, the smaller the share of the best-health type, the greater the level of

long run inequality. We …nd that the share of the best-health type in the Survey of Consumer

Finances is 6:2%. This …nding is somewhat surprising, but there is a fact that about 64:6%

of total wealth are owned by top 5% wealth holders and that their wealth share is increasing

(Saez and Zucman 2016).

[Related literature] There has been a heated debate on the “r minus g hypothesis” that is

fueled by three main research approaches. First, from a theoretical perspective, many authors

have discussed the e¤ect of “r minus g”, mainly in neoclassical growth frameworks or in models

generating Pareto distributions. Mankiw (2015) questions the usefulness of Piketty’s proposal

to tax wealth. Mankiw employs an optimal growth model with capitalists and workers and

shows that in a steady state r > g “arises naturally” (p. 45) where r is the (before -tax) rate

of return on capital, and g is the rate of labor-augmenting technological change and thus the

steady-state growth rate. In such a steady state, a certain level of inequality arises but this level

is stable. He continues by showing that a positive tax on wealth would generally not increase

consumption of workers.

Jones (2014, 2015) nicely contrasts this …nding by pointing to the fact that “r minus g” is

not obviously linked to inequality in a neoclassical growth model. By contrast, models yielding

Pareto distributions are much more suitable to appreciate the “r minus g”e¤ect. He develops a

model economy in which there is no labour income and heterogeneity emerges through a birth-

death process. He shows that such an economy displays a steady-state Pareto distribution of

wealth whose Pareto measure of inequality rises in “r minus g”, where r is the rate of return on

6
capital and g is the growth rate of capital per person in steady state. It is fair to say, however,

that even in such an economy with a Pareto distribution, long run inequality is stable.

Benhabib and Bisin (2018) provide a comprehensive overview of both historical background

and modern economic mechanisms that allow to understand why wealth distributions are skewed

and have an upper thick tail. They only brie‡y mention the determinants of inequality and

the relation to r: Benhabib, Bisin, and Zhu (2011) analytically study the e¤ect of taxes on the

power tail of the wealth distribution. They show that higher taxes reduce wealth inequality as

measured by their tail index.

From a quantitative perspective, there have been a number of papers that quantitatively

analyze the determinants of wealth inequality (Benhabib, Bisin and Luo 2017; Cao and Luo

2017; Khieu and Wälde 2018; Hubmer, Krusell, and Smith 2019). Khieu and Wälde (2018)

employ a framework featuring idiosyncratic risks to capital and labour income to match (almost

perfectly) the evolution of the wealth distribution of the NLSY79 cohort. They show that the

close-to-perfect …t is mainly attributed to the variations in capital returns. In line with this

…nding, Hubmer, Krusell, and Smith (2019) …nd that the observed variations in asset returns

are crucial for accounting for the observed U-shape of wealth inequality in the US over the

period 1967-2012. In connection with Piketty’s r g theory, these two papers share a view that

the growth rate g is less important than the rate of return r for explaining wealth inequality,

and that r should be perceived as an after-tax rate of return.

Third, from an empirical perspective, Acemoglu and Robinson (2015) exploit the cross-

country data for OECD countries to run a regression on inequality measured by top 1 percent

share of national income and …nd no clear correlation between r g and inequality. They

assume that all capital markets are open and all of the countries in the sample have the same

(possibly time-varying) interest rate. Under this assumption, cross-country variation in r g

will arise only because of variation in the growth rate g.

7
 The major di¤erence between our approach and those in the existing literature consists in

our simplicity and generality. We work with a small open economy model in which the interest

rate is exogenously …xed. There is ex-ante uncertainty and ex-post probability of death. This is

a simplest possible framework that can feature wealth and income distributions. Our generality

…rst lies in the fact that our framework can feature not only Pareto distributions but also any

empirical distributions of wealth and labour income. Second, we analyze the e¤ects of the rate

of return on wealth and the growth rate on the entire distribution of wealth rather than focus

on speci…c moments. Third, we allow for idiosyncratic risks to labour income though these

risks matter for the initial period only.

[Structure of the paper] Section 2 presents stylized facts about consumption growth and

output growth in OECD countries followed by Section 3 presenting a growth model for a small

open economy that can explain these facts. The model features labour income heterogeneity,

health types, r g, and wealth inequality. Section 4 discusses the distribution of wealth and how

wealth inequality is measured. Section 5 analyzes theoretical …ndings on the determinants of

long run wealth inequality. Section 6 presents empirical evidence that supports the theoretical

results and the …nal section concludes.

2 Stylized facts

In this section, using the OECD database, we present two important stylized facts about

consumption growth and output growth. We acquire these data from the online database of

the organization. The time span is 1980-2018. We then compute the di¤erence between average

consumption growth and average output growth for various time sub-periods ending in 2018

and starting from 1980. The data reveal the following stylized facts.

8
2.1 Unequal consumption and output growth

Figure 1 below shows the di¤erence between average consumption growth and average output

growth for various time periods ending in 2018 (see …g. 6 in app. I for periods starting from

1980)3 . In periods that start from the 1980s and the 1990s, the sample size reduces due to

data availability. For example, there are 37 observations in period 2005-2018 while there are

10 observations in period 1980-2018. It can be seen from the …gure that the di¤erence between

consumption growth and output growth in periods 1980-2018, 1985-2018, and 1990-2018 is less

than 1%, whereas the di¤erence in periods 1995-2018, 2000-2018, and 2005-2018 is greater than

1%.

GRC
1980-2018 GBR
1985-2018 NOR GBR
0.5 CAN
0.5 CAN
percent

NOR GRC
percent

USA USA
AUS MEX
OECD AUS ITA OECD
0 FRA
0 FRA
SWE

SWE CHE

-0.5 -0.5
KOR KOR

1990-2018 NOR 1995-2018 NOR

1
0.5 CAN
percent
percent

GBR
USA CAN
NZL NZL
AUS
GRC OECD
0.5 AUS MEX USA
ITA MEX GBR
FIN
0 FRA
FRA ITA OECD

CHE
SWE
0 GRC
JPN
SWE
CHE
AUT DEU
NLD AUT NLD
-0.5 -0.5
KOR KOR

2000-2018 NOR
CHL
2005-2018 NOR
1 CAN
1 CAN
FIN NZL
LTU NZL
FIN
percent

AUS ISR MEX USA GRC LVALTU

percent

EST FRA PRT AUS PRT USA

DNK GBR OECD DNK FRA ISRITA SWE
GRC LVA
0 HUN
ITA
JPN
SWE
0 BEL
EST
JPN MEX
LUX
GBR
OECD

AUT
BEL CZE ESP CHE AUT CZE DEU ISL POL ESP CHE
SVN
DEU LUX POL KOR TUR
TUR HUN
ISL NLD SVN NLD
KOR

-1
SVK
-1 SVK

-2
-2 IRL IRL

Figure 1 Di¤erence between average consumption growth and average output growth in

OECD countries for various time periods ending in 2018

3
See app. J for a list of OECD countries and their corresponding country codes.

9
2.2 Consumption growth exceeds output growth over the long run

When we consider the average over all OECD countries, Figure 1 shows that average consump-

tion growth is greater than average output growth in all periods except short periods 2005-2018

and 1980-1990 (see …g. 6 in app. I). There are two possible reasons for the fact that consump-

tion growth exceeds output growth on average over all OECD countries. First, there are a few

dominant countries with a big gap between consumption growth and output growth. Second,

consumption grows faster than output in the majority of the country members. In order to

better understand this, we compute the fraction of countries experiencing consumption growth

higher than output growth fo each time period. Figure 2 shows that the fraction is increasing

in the length of time periods. For example, consumption grows faster than output in less than

half of the countries in periods 1980-1990, 1980-1995, 1980-2000, 2005-2018, and 2000-2018. In

longer periods 1980-2010, 1980-2015, 1980-2018, and 1985-2018, consumption growth is bigger

than output growth in approximately 70%-80% of the OECD countries. We therefore conclude

that most OECD countries experience consumption growth exceeding output growth over the

long run.

0.7 0.8

0.6
0.6

0.5
0.4
0.4
00 8

95 8

90 8

85 8

80 8

80 0

80 5

80 0

80 5

80 0

5
20 201

19 201

19 201

19 201

19 201

01

19 199

19 199

19 200

19 200

19 201

01
-2

-2
-

-
05

80
20

19

time period time period

Figure 2 Fraction of OECD countries experiencing consumption growth higher than output

growth for various time periods ending in 2018 (left panel) and starting from 1980 (right

panel)

10
3 The model

Consider a small open economy populated by N households indexed by i = 1; :::; N . Household i

is endowed with hi e¢ ciency units of labour that are supplied inelastically. There are free capital

and goods ‡ows across borders, but labour is immobile. Firms act under perfect competition

and employ capital in an international capital market. The rate of return on capital is given

by r. Time is continuous.

3.1 The representative …rm

The representative …rm produces a …nal good Y (t) employing capital K (t) and total e¢ ciency
PN
units of labour L = i=1 hi according to a Cobb-Douglas production function

Y (t) = A (t) K (t) L1 ,0< < 1; (1)

where total factor productivity A (t) deterministically grows at a rate of gA > 0 according to4

A (t) = A (t) = A0 egA t . (2)

The initial level of total factor productivity A0 is positively given. The …nal good is freely

traded across countries and its price is normalized to one. The …rm’s pro…t maximization

problem implies that the marginal productivity of capital is equal to the internationally given

rate of return r
1
K (t)
r = A (t) : (3)
L
4
See app. B for an extension in which we allow for knowledge spillovers.

11
The equilibrium wage is equal to the marginal productivity of labour

K (t)
wL (t) = (1 ) A (t) : (4)
L

As capital returns are constant, Equation (3) implies that the optimal domestic capital-

labour ratio evolves according to the evolution of total factor productivity A(t). This in turn

implies that the evolution of the equilibrium wage given by (4) is determined by changes in

total factor productivity A(t). The equilibrium wage rate and output therefore grow at the

same rate g (see app. A for detail), i.e.

Y (t) = Y0 egt ; (5)

wL (t) = w0L egt : (6)

1 1
gA
where g = 1
; Y0 = r
1
A01 L and w0L = (1 ) A01 r
1
.

3.2 Households

Households own (save in) capital. Let us now consider one household and therefore suppress

the index i for convenience. A household supplies h e¢ ciency units of labour inelastically and

thereby receives a ‡ow of labour income w(t) = wL (t) h. The household draws the e¢ ciency

units of labour h from a distribution before she becomes economically active. Once drawn, h

remains constant. Therefore, labour income deterministically evolves according to

dw (t) = gw (t) dt: (7)

The budget constraint of a household therefore describes the change in wealth (i.e. savings)

12
as the sum of capital income and labour income, ra (t) + w (t), minus consumption, c (t),

a_ (t) = ra (t) + w (t) c (t) : (8)

Adding a no-Ponzi game condition, we can write the corresponding intertemporal budget con-

straint as (see app. C for derivation)

Z 1 Z 1
rt (r g)t
e c(t)dt = a0 + w0 e dt; (9)
0 0

where a0 and w0 = w0L h are initial wealth and initial labour income, respectively. The intertem-

poral budget constraint (9) shows that the present value of lifetime consumption must be equal

to initial wealth plus the present value of lifetime labour income. The household draws a0 from

a distribution before she becomes economically active. This captures the ex-ante heterogeneity

in …nancial background of parents. It can also be interpreted as the idiosyncratic risk to wealth

endowment. The distribution of initial wealth a0 can admit any cross-sectional distribution

of wealth. As the household draws the e¢ ciency units of labour h from a distribution, initial

labour income w0 obeys a distribution as well. This captures the ex-ante heterogeneity in la-

bour productivity/income or the idiosyncratic risk to initial labour income. We assume that

a household who draws a high level of initial wealth a0 is likely to draw a high level of labour

productivity h and therefore a high level of initial labour income w0 . This captures the fact that

children born in wealthy families are relatively more educated and therefore are more likely to

…nd high-paying jobs. Beyond ex-ante uncertainty in initial wealth and labour income, there

is also ex-post uncertainty in the time of death as in the perpetual youth model of Blanchard

(1985) building on Yaari (1965). Death is a Poisson process with an arrival rate > 0. The

household therefore faces a constant instantaneous probability of death at each instant of

time. We assume the death rate is drawn from some distribution before the household be-

13
comes economically active; this captures the ex-ante heterogeneity in health condition. Once

drawn, the death rate remains constant over the household’s life cycle.

Our model therefore di¤ers from that in Jones (2015) in three directions. First, we allow

for growing labour income. Second, our heterogeneity stems from the initial distributions while

Jones (2015) assumes heterogeneity through a birth-death process and redistribution. Third,

we allow for the heterogeneity in labour income and in the death rate. We di¤er from Piketty

and Zucman (2015) in the second and third direction. These deviations allow us to be very

‡exible and general. In particular, our framework allows for understanding the e¤ect of “r

minus g”on wealth inequality under any distributions of wealth and labour income.

The probability that the household is still alive at time t conditional on being alive at time

t = 0 is given by

t
p(t) = e : (10)

The household downweighs the utility she receives by this probability of being alive. The

intertemporal utility function of a household therefore reads

Z 1
(~+ )t
U (0) = e u (c(t)) dt; (11)
0

where ~ > 0 is the time preference rate. Denote := ~ + as the e¤ective time preference rate,

which shows that the household discounts the future more strongly in the presence of death

shocks. Instantaneous preferences of the household re‡ect constant intertemporal elasticity of

1
substitution with a parameter

8
>
>
>
< c(t)
1 1
; > 0; 6= 1;
1
u(c (t)) = (12)
>
>
>
:ln(c(t)); = 1:

14
3.2.1 Optimal behaviour

The household chooses consumption after endowment (a0 ; w0 ) and the death rate have real-

ized to maximize the intertemporal utility (11) taking into account the equilibrium wage (7)

and the budget constraint (8). This implies a deterministic maximization problem with the

corresponding straightforward solution that reads (see app. D)

(1 )r w (t)
c (t) = a (t) + : (13)
r g

Consumption is a constant share of the sum of …nancial wealth and the present value of labour

income.

This equation makes economic sense if we make the following assumptions. First, we expect

a non-negative present value of labour income. This requires the interest rate to be greater

than the growth rate

r > g: (14)

This is actually a boundedness condition for the intertemporal budget constraint to deliver a

…nite present value of labour income. Second, we also expect that consumption is increasing in

wealth and labour income. This requires

>1 : (15)
r

This condition shows that when the intertemporal elasticity of substitution is greater than one,

1
i.e. > 1, the e¤ective time preference rate must be su¢ ciently high compared to the

interest rate r. When the intertemporal elasticity of substitution is less than or equal to one,

1
i.e. 1, this condition always holds. Third, we expect a positive value of consumption,

15
and we …nd the following natural borrowing constraint (Aiyagari 1994) for any t 0;

w (t)
a (t) > anat : (16)
r g

The optimal growth rate of consumption which is resulted from this maximization problem,

and which is consistent with the solution from (13), is given by the Keynes-Ramsey rule

c_ (t) r
= gc: (17)
c (t)

We denote the growth rate of consumption for later purposes by g c .

3.2.2 The evolution of wealth

Given optimal consumption from (13), the evolution of wealth can be determined by solving

the dynamic budget constraint (8). Wealth a (t) for any t is given by (see app. E for derivation)

w0 g c ]t c
a(t) = a0 + 1 e[g eg t : (18)
r g

Initial endowment a0 and w0 in addition to the rate of return and the growth processes of the

wage and consumption …x the wealth level of the household for future point t in time.

Given initial wealth a0 , (18) shows that there are two e¤ects of the growth processes on

wealth at any point t in time: additive e¤ect and scale e¤ect. Speci…cally, initial wealth is

w0 g c ]t
…rst added up by r g
1 e[g . Second, initial wealth plus the additive e¤ect is scaled
c
by the ratio of consumption at time t to initial consumption, i.e. c(t)=c0 eg t . Thus, the

additive e¤ect depends on the di¤erence between wage growth and consumption growth while

the scale e¤ect is determined by consumption growth only. When g = g c holds, the additive

e¤ect vanishes. When g > g c holds, the “additive e¤ect” becomes negative. When g c > g

16
holds, (18) can be rewritten as

w0 1 g c ]t c
a(t) = a0 + 1 e[g (g c g) eg t : (19)
r g gc g

R1 g c ]t
As 0
1 e[g (g c g) dt = 1 holds, (19) shows that the size of the additive e¤ect is equal

g c ]t
to a certain fraction, 1 e[g (g c g), of lifetime labour income adjusted by the inverse of

w0 1
the di¤erence between consumption growth and output growth, r g gc g
.

3.2.3 Consumption and wealth dynamics

The dynamics of consumption and wealth depend on the household’s health type, which is

determined by the death rate relative to the rate of return, output growth, and preferences.

Our household …nds themself as a bad-health type when the death rate is su¢ ciently high so

that it satis…es5

>r (~ + g) ; (20)

which is equivalent to r < (~ + ) + g. As the interest rate is lower than the growth-adjusted

time preference rate, the household dissaves. Wealth and therefore consumption decrease over

time, which justi…es the negative “additive e¤ect”mentioned above. The condition (20) implies

that consumption growth of a bad-health-type household is lower than output growth, i.e.

g c < g.

The household is considered as the average-health type when the death rate is as high as

=r (~ + g) . (21)

In this case, the interest rate is equal to the growth-adjusted time preference rate, i.e. r =

(~ + ) + g. Wealth, consumption and output grow at a same rate of g = g c . The additive

5
The dynamics of wealth and consumption of a bad-health-type household in our model are of analogy to
those of an individual experiencing a low-interest-rate regime in Khieu and Wälde (2018).

17
e¤ect on wealth is equal to zero.

Our household …nds themself as a good-health type when the death rate is as low as

<r (~ + g) ; (22)

which is equivalent to r > (~ + ) + g.6 In this case, the hosuehold has an incentive to save

since the interest rate is higher than the growth-adjusted time preference rate. Wealth increases

due to both additive e¤ect and scale e¤ect. Condition (22) implies that consumption growth

of a good-health-type household is greater than output growth, i.e. g c > g.

The health types can therefore help to explain the stylized facts presented in Section 2. In

a model featuring a su¢ ciently large share of good-health types that experience consumption

growth higher than output growth, the growth rate of aggregate consumption would be greater

than the growth rate of output. Obviously, in the absence of good-health types, aggregate

consumption growth does not exceed output growth. This however does not mean that balanced

consumption and output growth implies the non-existence of good-health types. When the

share of good-health types is su¢ ciently small, aggregate consumption growth and output

growth might be identical. This demonstrates the ‡exibility of the model in explaining the

di¤erence between consumption growth and output growth in the data.

4 The distribution of wealth and the measure of wealth

inequality

We are now in a position to understand the evolution of the distribution of wealth. When a0 ,

h, and in (18) are drawn from a distribution, then a (t) obeys a distribution as well.7 There
6
This condition implies an “exploding regime” of interest rate as described in Benhabib and Bisin’s (2018).
This exploding regime is crucial for producing a fat right tail of the wealth distribution.
7
The death rate and labour productivity h appear in the evolution of wealth through consumption growth
gc = r ( ~+ ) and labour income w0 = hw0L , respectively.

18
are two interpretations one can give to this distribution of a (t). First, it is the distribution of

wealth of a household for some future point t in time when the household has not yet drawn her

endowment (a0 ; h) and the death rate : Second, when there are many households that draw

(a0 ; h) and independently from some initial distribution, then the distribution of a (t) is the

cross-sectional distribution of wealth for households in t. We adopt the second view here.8

Let us now assume that the realizations of the death rate can be either 1, 2, 3 ,..., or M,

1 < M < 1, with corresponding probabilities 1, 2, 3 ,..., or M, where j < j+1 for any
PM
j 2 Hn fM g f1; :::; M g n fM g and j=1 j = 1. We may write the expected wealth and the

variance of wealth as follows (see Wackerly et al. 2008, ch. 5.11).

X
M
E [a(t)] = E [a(t)j = j] j, (23)
j=1

X
M
var [a (t)] = var [a(t)j = j] j + var [E [a(t)j ]] , (24)
j=1

where the …rst term of the right hand side of (24) captures uncertainty in wealth endowment

and labour productivity while the second term captures randomness in the death rate. The

conditional expectation and variance are given by9

e[g gjc ]t
w0 c
E [a(t)j = j] = a0 + 1 egj t , (25)
r g
2 2
e[g gj ]t e[g gjc ]t
w0 c a0 w 0 c
var [a(t)j = j] = 2
a0 + 2 1 +2 1 e2gj t , (26)
(r g) r g
2 !2 3
2 X M X
M
w0 c c
var [E [a(t)j ]] = a0 + 4 e2gj t j egj t j
5, (27)
r g j=1 j=1

2 2
where a0 Ea0 , w0 Ew0 , a0 vara0 , w0 varw0 , a0 w 0 cov(a0 ; w0 ) > 0, and gjc
r (~+ j)
. The covariance between initial wealth and initial labour income a0 w 0 is positive as
8
Bayer, Rendall, and Wälde (2019) made the …rst attempt to provide an interpretation to the wealth distri-
bution of an individual as a cross-sectional distribution of wealth for a small open economy.
9
See app. F for the derivation of var[E [a(t)j ]].

19
it is assumed that a household that draws a high value of initial wealth a0 is likely to draw

a high value of labour productivity and therefore labour income w0 . It is straightforward to

c
verify that g1c > g2c >; :::; > gM . The …rst moment of our cross-sectional distribution of wealth

at some point t in time is therefore given by

X
M
e[g gjc ]t
w0 c
a (t) = a0 + 1 egj t j , (28a)
j=1
r g

where a (t) E [a (t)]. As most of the empirical cross-sectional distributions of wealth and

labour income have positive mean, we assume a0 > 0, w0 > 0, and a (t) > 0 for any t > 0.

Average wealth is determined by average initial wealth and average initial income plus the

growth rates g and gjc plus the interest rate plus the distribution of the death rate. Similarly,

the variance of wealth evolves according to

X
M 2 2
e[g gjc ]t
e[g gjc ]t
w0 a0 w 0 c
2
a (t) = 2
a0 + 2 1 +2 1 e2gj t j
j=1
(r g) r g
2 !2 3
w0
2 XM
c
X
M
c
+ a0 + 4 e2gj t j egj t j
5,
r g j=1 j=1

2
where a (t) var[a (t)]. The coe¢ cient of variation for our distribution of wealth is given by

v
u P
u M 2 2
e[g gjc ]t
e[g gjc ]t c
u j=1 2
a0 + w0
2 1 +2 a0 w0
1 e2gj t j
u (r g) r g
u
t 2 PM c PM c 2
p + a0 + w0
j=1 e2gj t j j=1 egj t j
var [a (t)] r g
CVa (t) = PM h i :
E [a (t)]
j=1 a0 + w0
r g
1 e[g gjc ]t
egjc t
j
(28b)

Given the distributions of initial wealth, initial labour income, and the death rate, a change

in the coe¢ cient of variation is due to changes in both the standard deviation and the mean.

In particular, when the standard deviation rises relative to the mean, the coe¢ cient of vari-

20
ation rises. Hence, we measure wealth inequality by the coe¢ cient of variation.10 One typical

property of the coe¢ cient of variation is scale-independence. Speci…cally, it is unchanged when

wealth of all individuals grows at a same rate. As the mean of the distribution of wealth at

any point in time is assumed to be positive, so is the coe¢ cient of variation.

5 Wealth inequality in the long run

We now study the e¤ect of the interest rate r; the growth rate g, the di¤erence r g, and

initial conditions on long run wealth inequality. The long run in our setup is when time goes

to in…nity. The coe¢ cient of variation (28b) shows that wealth inequality arises due to three

sources of heterogeneity: heterogeneity in wealth endowment, in labour income, and in the

death rate (or health condition). In order to gain a better intuition for the …nal results, we

proceed by adding up every source of heterogeneity. We start with the heterogeneity in the

death rate only.

5.1 Identical wealth endowment and labour income

When households draw wealth endowment and labour productivity from (univariate) degenerate

2 2
distributions, i.e. a0 = 0, w0 = 0, and a0 w 0 = 0, the coe¢ cient of variation from (28b)

collapses to r
w0
PM 2gj t c PM c
gj t
2
a0 + r g j=1 e j j=1 e j
CVa (t) = PM h i c : (29)
e[g gj ]t egj t j
w0 c
j=1 a0 + r g
1

The following proposition provides the level of wealth inequality in the long run and shows how

it depends on the distribution of the death rate.

Proposition 1 Consider a small open economy populated by M 2 health types. Assume

there exists a k, 1<k M , such that gkc = g. (i) There exists a stationary coe¢ cient of
10
See Allison (1978) for discussion of other measures of inequality.

21
variation in the long run, i.e.
r
1 1
lim CVa (t) = : (30)
t!1 1

(ii) Inequality measured by the coe¢ cient of variation is decreasing in 1 but is independent of

r g.

Proof. See app. G.1

Good-health types experience positive additive e¤ect and therefore accumulate wealth over

time. Bad-health types dissave as they experience negative additive e¤ect. Among good-health-

type households, the households that draw the smallest death rate experience the greatest scale

e¤ect on wealth and therefore accumulate wealth fastest. This health type is considered as the

best-health type. In the long run, the wealth share owned by this best-health type dominates

the wealth share of the other health types. Thus, long run wealth inequality is determined by

the share of the best-health type, 1, relative to the share of the other health types, 1 1,

as shown in (30). The lower the share of the best-health type, the higher the level of long run

wealth inequality.

5.2 Identical wealth endowment

We now allow for labour income heterogeneity and assume initial wealth is drawn from a

2
univariate degenerate distribution, i.e. a0 = 0. The covariance between wealth and labour

income is therefore equal to zero, i.e. a0 w 0 = 0. The coe¢ cient of variation is given by

s
PM 2 2 2 PM PM 2
e[g gjc ]t c c c
w0
j=1 (r g)2 1 e2gj t j + a0 + r g
w0
j=1 e2gj t j j=1 egj t j

CVa (t) = PM h i ;
e[g gjc ]t c
j=1 a0 + r g
w0
1 egj t j
(31)

which shows that wealth inequality arises due to the heterogeneity in labour income and in

the death rate. Long run wealth inequality and its determinants are revealed in the following

22
proposition.

Proposition 2 Consider a small open economy populated by M 2 health types. Assume

there exists a k, 1<k M , such that gkc = g. (i) There exists a stationary coe¢ cient of

variation in the long run, i.e.

v
u 2
u w0
1 1
u (r g)2 1
lim CVa (t) = t 2 + : (32)
t!1 w0 1 1
a0 + r g

(ii) Inequality measured by the coe¢ cient of variation is decreasing in r g.

Proof. See app. G.2.

Compared to the level of long run wealth inequality from (30), wealth inequality given

by (32) is further ampli…ed due to the heterogeneity in labour income. As discussed above,

long run wealth inequality is characterized by the best-health type, which experiences positive

additive e¤ect and the greatest scale e¤ect on wealth accumulation. Since the best-health-type

households all draw the smallest death rate, they experience identical scale e¤ect. Therefore,

it is the additive e¤ect on wealth accumulation that determines long run wealth inequality

measured by the coe¢ cient of variation.

The size of the additive e¤ect in the long run is equal to the present value of lifetime labour

income w0 =(r g). An increase in r g leads to decreasing additive e¤ect on wealth across

households. This simultaneously causes both the average wealth and the variance of wealth in

the long run to fall. When initial wealth is positive, the decreasing e¤ect of r g is weaker on

long run average wealth than on the standard deviation of wealth in the long run. This results

in decreasing wealth inequality in the long run. The following corollary provides the level of

long run inequality when wealth endowment is zero.

Corollary 1 Assume initial wealth is equal to zero. (i) The coe¢ cient of variation in the long

23
run reads
s
2
w0 1 1 1
lim CVa (t) = + : (33)
t!1 w0 1 1

(ii) Inequality measured by the coe¢ cient of variation is independent of r g:

In this case, the decreasing e¤ects of r g on the average wealth and on the standard

deviation of wealth are of equal size. The e¤ects of r g on long run inequality are therefore

neutralized. Hence, heterogeneity in labour income alone is not su¢ cient to translate the e¤ects

of r g on long run wealth inequality. Positive initial wealth is necessary. The following corollary

provides the mechanism through which increasing the di¤erence r g leads to decreasing wealth

inequality in the long run.

Corollary 2 An increase in the di¤erence between the rate of return on wealth and the growth

rate leads to decreasing wealth inequality in the long run due to the heterogeneity in labour

income.

Increasing the gap r g tends to shift the wealth distribution to the left due to decreasing

additive e¤ect on wealth accumulation. This tends to lower the average wealth as a result.

As the additive e¤ect di¤ers across households due to labour income heterogeneity, households

earning high labour income experience stronger (decreasing) e¤ect than those earning low labour

income. This produces a decreasing e¤ect on the standard deviation of wealth. When wealth

endowment is zero, these two e¤ects are cancelled out as shown in Corollary 1, leaving long

run wealth inequality unchanged. When households are endowed with positive wealth, this

increases the average wealth but has no e¤ect on the variance of wealth since wealth endowment

is deterministic. In e¤ect, the decreasing e¤ect of the di¤erence r g is stronger on the standard

deviation of wealth than on the average wealth. The e¤ect of r g on wealth inequality through

the scale e¤ect is neutralized since wealth inequality is measured by the coe¢ cient of variation.

Overall, wealth inequality in the long run falls due to decreasing additive e¤ect.

24
 2
Lemma 1 The greater the dispersion index of labour income, w0 = w0 , the stronger the de-

creasing e¤ect of r g on long run wealth inequality.

Proof. See app. G.3.

Proposition 1 shows that the di¤erence r g has no e¤ect on long run inequality in the

absence of heterogeneity in labour income and in wealth endowment. In the presence of la-

bour income heterogeneity, Proposition 2 shows that r g has a decreasing e¤ect on long run

inequality. A direct implication therefore is the more households di¤er in labour income, the

stronger the decreasing e¤ect of r g on long run inequality. The degree of heterogeneity can

be measured by the dispersion index.

5.3 Identical labour income

When we allow for heterogeneity in wealth endowment and abstract from labour income het-

erogeneity, the coe¢ cient of variation reads

s
PM c
2 2gj t
2 PM c PM c 2
j=1 a0 e j + a0 + w0
r g j=1 e2gj t j j=1 egj t j

CVa (t) = PM h i ; (34)

e[g gjc ]t c
j=1 a0 + w0
r g
1 egj t j

which shows that wealth inequality arises due to the heterogeneity in wealth endowment and

in the death rate. The following proposition pins down long run wealth inequality and its

determinants.

Proposition 3 Consider a small open economy populated by M 2 health types. Assume

there exists a k, 1<k M , such that gkc = g. (i) There exists a stationary coe¢ cient of

variation in the long run, i.e.

v
u 2
1 1
lim CVa (t) = u a0 1
t 2 + : (35)
t!1 w0 1 1
a0 + r g

25
(ii) Inequality measured by the coe¢ cient of variation is increasing in r g.

Proof. See app. G.4.

As discussed in Section 5.2, it is the best-health type that characterizes wealth inequality

in the long run. Since the best-health type experiences identical scale e¤ect, the additive e¤ect

on wealth accumulation matters for long run wealth inequality. Increasing the di¤erence r g

causes the average wealth to decrease as the additive e¤ect is weaker. Since the additive e¤ect is

identical across households (due to identical labour income), this leaves the variance of wealth

unchanged. The coe¢ cient of variation declines as a result. The following corollary shows how

this result di¤ers when households earn zero labour income.

Corollary 3 Assume labour income is zero. (i) The coe¢ cient of variation in the long run

reads
s
2
a0 1 1 1
lim CVa (t) = + : (36)
t!1 a0 1 1

(ii) Inequality measured by the coe¢ cient of variation is independent of r g:

When labour income is zero, the additive e¤ect is zero. This switches o¤ the e¤ect of r g

on the wealth distribution in the long run. Thus, heterogeneity in wealth endowment alone is

not su¢ cient to translate the e¤ects of r g on long run wealth inequality. Positive labour

income is necessary.

Corollary 4 An increase in the di¤erence between the rate of return on wealth and the growth

rate leads to increasing wealth inequality in the long run due to the heterogeneity in wealth

endowment.

Corollary 3 shows that the di¤erence r g has no e¤ect on long run wealth inequality in

the absence of labour income as the additive e¤ect on wealth accumulation disappears. When

labour income is in place, so is the additive e¤ect. An increase in r g tends to shift the wealth

26
distribution to the left since the additive e¤ect decreases. As the additive e¤ect is identical

across households, this tends to lower the average wealth but leaves the variance of wealth

unchanged. Since wealth inequality is measured by the coe¢ cient of variation, the e¤ect of

r g on the wealth distribution through the scale e¤ect is neutralized. In e¤ect, long run

wealth inequality rises in r g through the additive e¤ect on wealth accumulation.

2
Lemma 2 The higher the dispersion index of wealth endowment, a0 = a0 , the stronger the

increasing e¤ect of r g on long run wealth inequality.

Proof. See app. G.5.

In the absence of heterogeneity in wealth endowment and in labour income, long run inequal-

ity is independent of the di¤erence r g as shown in Proposition 1. When the heterogeneity

in wealth endowment is in place, Proposition 3 shows that long run inequality is increasing in

r g. Lemma 2 is therefore a direct implication of Proposition 1 and Proposition 3.

5.4 Independence between wealth and labour income

We now allow for all sources of heterogeneity but assume independence between wealth and

labour income, i.e. a0 w 0 = 0. The coe¢ cient of variation reads

v
u PM
u 2 2
e[g gjc ]t c
u j=1
2
a0 + w0
2 1 e2gj t j
u (r g)
u
t 2 PM c PM c 2
+ a0 + w0
r g j=1 e2gj t j j=1 egj t j
CVa (t) = PM h i : (37)
j=1 a0 + w0
r g
1 e[g gjc ]t
e gjc t
j

At any point t in time, wealth inequality is determined by the heterogeneity in wealth endow-

ment, labour income, and the death rate, and by all the primitive parameters. The following

proposition provides the level of wealth inequality and its determinants in the long run.

27
Proposition 4 Consider a small open economy populated by M 2 health types. Assume

there exists a k, 1<k M , such that gkc = g. (i) There exists a stationary coe¢ cient of

variation in the long run, i.e.

v
u 2
u 2 + w0
1 1
u a0 (r g)2 1
lim CVa (t) = t 2 + : (38)
t!1 w0 1 1
a0 + r g

(ii) Long run wealth inequality measured by the coe¢ cient of variation rises in r g if the gap

is su¢ ciently large, i.e.

2
d w0 = w0
lim CVa (t) > 0 , r g> 2
: (39)
d (r g) t!1 a0 = a0

(iii) Long run wealth inequality is decreasing in r g if

r g< : (40)

Proof. See app. G.6.

The distributions of initial wealth and labour income, the share of the best-health type, and

the di¤erence between the rate of return on wealth and the growth rate …x the level of wealth

inequality in the long run. One can easily verify that long run wealth inequality from (38) is

greater than that from (32) or (35).

Section 5.1 shows that the di¤erence between the rate of return on wealth and the growth

rate has no e¤ect on long run wealth inequality in the absence of heterogeneity in wealth

endowment and in labour income. In Section 5.2, when labour income heterogeneity is allowed

for, the di¤erence r g produces decreasing e¤ect on long run wealth inequality. The more

dispersed the distribution of labour income, the stronger the (decreasing) e¤ect of r g on

long run inequality. In Section 5.3, when labour income heterogeneity is abstracted and the

28
heterogeneity in wealth endowment is in place, an increase in the di¤erence r g leads to rising

long run wealth inequality. The more dispersed the distribution of initial wealth, the stronger

the (increasing) e¤ect of r g on long run inequality.

An increase in the di¤erence r g therefore has two opposite e¤ects on long run wealth

inequality when all sources of heterogeneity are in place. First, it causes long run inequality to

decrease through the heterogeneity in labour income. Second, it leads to rising inequality in

the long run through the heterogeneity in wealth endowment. These two opposite e¤ects imply

a threshold of r g above which increasing the di¤erence r g leads to rising wealth inequality,

and below which long run wealth inequality falls as the gap r g gets larger. A higher (lower)

threshold means the …rst e¤ect is relatively stronger (weaker) than the second one because

the range within which long run inequality is decreasing in r g is larger (smaller). Such a

threshold of r g therefore increases in the dispersion of labour income since the more dispersed

the distribution of labour income, the stronger the …rst e¤ect. The threshold is decreasing in

the dispersion of initial wealth since the more dispersed the distribution of initial wealth, the

stronger the second e¤ect.

5.5 Full model

We now turn to the full model to have a full picture of the conjectured relationship between

r g and long run wealth inequality. The evolution of wealth and the coe¢ cient of variation

in the full model are respectively given by (18) and (28b). To obtain meaningful results, we

make the following assumptions.

Assumption 1 The dispersion index of the initial wealth distribution is greater than that of

the distribution of initial labour income, i.e.

2 2
a0 w0
> :
a0 w0

29
Assumption 2 The dispersion index of the initial wealth distribution is greater than the ratio

of the covariance between initial wealth and initial labour income to the average initial labour

income, i.e.
2
a0 a0 w 0
> :
a0 w0

These assumptions are strongly supported by the empirical evidence that wealth distribu-

tions are more dispersed than labour income distributions (see section 6.1). The following

proposition provides the level of long run wealth inequality and its determinants.

Proposition 5 Consider a small open economy populated by M 2 health types. Assume

there exists a k, 1<k M , such that gkc = g. (i) There exists a stationary coe¢ cient of

variation in the long run, i.e.

v
u 2
u 2 + w0
+2 a0 w0
1 1
u a0 (r g)2 r g 1
lim CVa (t) = t 2 + : (41)
t!1 w0 1 1
a0 + r g

(ii) Long run wealth inequality measured by the coe¢ cient of variation rises in r g if the gap

is su¢ ciently large, i.e.

2
d w0 = w0 a0 w 0 = a0
lim CVa (t) > 0 , r g> 2
: (42)
d (r g) t!1 a0 = a0 a0 w 0 = w 0

(iii) Long run wealth inequality is decreasing in r g if

r g< : (43)

Proof. See app. G.7.

The distributions of initial wealth and labour income, their covariance, the share of the

best-health-type households, and the gap r g …x long run wealth inequality. Similar to the

case of independence between wealth and labour income, there are two opposite e¤ects of r g

30
on long run wealth inequality given by the coe¢ cient of variation (41): the …rst e¤ect causes

long run inequality to decrease through the heterogeneity in labour income while the second

e¤ect leads to rising long run inequality through the heterogeneity in wealth endowment. When

the di¤erence r g is greater than the threshold , the …rst e¤ect is dominated, leading to

increasing wealth inequality. By contrast, long run wealth inequality is decreasing in r g if

the di¤erence r g is less than the threshold . A higher dispersion index of labour income

strengthens the …rst e¤ect while the second e¤ect is stronger as wealth endowment is more

dispersed. This justi…es why the r g’s threshold, , is increasing in the dispersion of labour

2 2
income, w0 = w0 , but decreasing in the dispersion of wealth endowment, a0 = a0 .

What are the e¤ects of the covariance between wealth and labour income on long run

inequality and on the r g’s threshold? Comparing (38) and (41), it can be seen that the

covariance between wealth and labour income further ampli…es long run wealth inequality.

This is because the e¤ect of labour income on wealth accumulation is additive. The correlation

between initial wealth and initial labour income therefore ampli…es the variance of wealth in

the long run but has no e¤ect on the average wealth.

To understand the e¤ect of the covariance between wealth and labour income on the r g’s

threshold, one could compare the threshold in the absence of the covariance with that in

presence of the covariance, i.e. compare and . If < holds, then the covariance

between wealth and labour income strengthens the increasing e¤ect of r g. By contrast, if

> holds, then the covariance weakens the increasing e¤ect of r g. If is equal to ,

the covariance between wealth and labour income is irrelevant for the r g’s threshold. The

following proposition reveals the relation between and .

Proposition 6 (Covariance between wealth and labour income and the r g’s threshold)

(i) If CVa0 > CVw0 holds, the r g’s threshold is strictly positive and smaller than the ratio

of the dispersion index of the initial labour income distribution to that of the initial wealth

31
 2 2
distribution, i.e. 2 0; w0 = w0 = a0 = a0 .

(ii) If CVa0 < CVw0 holds, the r g’s threshold is greater than the ratio of the dispersion

index of the initial labour income distribution to that of the initial wealth distribution, i.e.
2
w0 = w0
> 2 .
a0 = a0

(iii) If CVa0 = CVw0 holds, the r g’s threshold is equal to the ratio of the dispersion index of
2
w0 = w0
the initial labour income distribution to that of the initial wealth distribution, i.e. = 2 .
a0 = a0

Proof. See app. G.8.

Thus, the e¤ects of the covariance between initial wealth and initial labour income on the

r g’s threshold depend on the distributions of initial wealth and initial labour income. When

initial wealth is more unequally distributed than initial labour income, i.e. CVa0 > CVw0 , is

less than , i.e. the covariance between initial wealth and initial labour income strengthens

the increasing e¤ect of r g on long run wealth inequality. By contrast, when initial labour

income is more unequally distributed, i.e. CVa0 < CVw0 , the covariance between wealth and

labour income weakens the increasing e¤ect of r g since it increases the r g’s threshold,

i.e. > . When the coe¢ cients of variation of the distributions of initial wealth and initial

labour income are identical, the covariance between wealth and labour income has no in‡uence

on the r g’s threshold.

That the mechanisms through which the di¤erence r g produces (decreasing and increasing)

e¤ects on long run wealth inequality have been understood, the following proposition provides

conditions under which increasing the di¤erence between the rate of return on wealth and the

growth rate leads to rising wealth inequality in the long run.

Proposition 7 (Long run wealth inequality and r g)

Given a distribution of labour income, a distribution of wealth endowment, and a distribution

of health types, long run wealth inequality rises in r g if

(i) r g is su¢ ciently larger than a positive threshold, and

32
(ii) there exist good-health types that experience consumption growth higher than output growth.

Proof. See app. G.9.

The rate of return on wealth r in our setup is perceived as an after-tax rate of return, i.e.

r = r~ , where r~ and are the gross rate of return and the tax rate, respectively. Thus, a

decrease in r can be translated to an increase in a tax on wealth . Proposition 7 therefore

suggests a necessary condition under which a wealth tax produces desirable e¤ects, i.e. reduce

long run inequality. That is the rate of return on wealth must be su¢ ciently higher than the

growth rate.

6 Empirical evidence

Proposition 7 provides two conditions under which increasing the di¤erence between the rate

of return on wealth and the growth rate leads to rising wealth inequality in the long run. In

this section, we test the empirical plausibility of these conditions.

6.1 How high is the r g’s threshold?

We compute the r g’s threshold using the data on wealth and labour income extracted from

the Survey of Consumer Finances 1989-2016. Let us …rst check the validity of Assumption 1

and Assumption 2. Figure 3 shows that the dispersion indices of wealth are much higher than

those of labour income and are also much higher than the ratios of the covariance between

wealth and labour income to average labour income. Since the assumptions are validated, the

r g’s thresholds are reported in Table 1.

33
 400

300
2
/
200 a a
2
/
w w
100 /
aw w

19 9
19 2
19 5
20 8
20 1
20 4
20 7
20 0
20 3
16
8
9
9
9
0
0
0
1
1
19
Figure 3 The dispersion indices of wealth and labour income in the Survey of Consumer

Finances 1989-2016 (wealth and labour income expressed in million US$ in 2016 prices)

The second row of Table 1 shows that the threshold above which an increase in r g leads to

increasing wealth inequality ranges from 0:8% to 15:4%. The (unweighted) average threshold

is 4:4%.11 Cao and Luo (2017) compute returns to wealth from the Panel Study of Income

Dynamics and report that the mean rates of (annualized) returns on wealth with and without

capital gains are 7:92% and 5:94%, respectively. When we take the rate of return on wealth

with capital gains and the average growth rate of 2:6% (see Table 2) into account, the di¤erence

r g in the data would be around a value of 5:32%. This demonstrates the empirical plausibility

of the r g’s threshold.

Table 1 The empirical thresholds of r g and the coe¢ cients of variation of wealth and

labour income

1989 1992 1995 1998 2001 2004 2007 2010 2013 2016

1.3% 0.8% 4.1% 2.8% 1.8% 9.3% 2.4% 15.4% 1.0% 4.7%

1.7% 1.0% 3.5% 2.6% 1.8% 6.2% 2.1% 8.1% 1.1% 3.4%

CVa 4.2 6.2 5.3 5.0 4.4 4.3 4.6 5.9 6.0 5.7

CVw 3.2 4.6 6.4 5.5 4.4 8.0 5.7 12.0 5.3 9.3

11
Piketty (2014) states that “the central value...” of r g “...observed over the long run is 4 5 percent a
year” (p.143).

34
 Though the dispersion indices of wealth are signigicantly higher than those of labour income,

the coe¢ cients of variation of wealth (the fourth row) are not always greater than those of

labour income (the last row). This is because the dispersion index is scale-dependent while the

coe¢ cient of variation is scale-independent. This justi…es why the r g’s threshold is mostly

larger in the presence of the correlation between wealth and labour income (second row) than

in the absence of it (third row). The covariance between wealth and labour income in the data

therefore tends to lessen the increasing e¤ect of r g on wealth inequality.

6.2 How large is the share of good-health types?

Section 2 presented a stylized fact that consumption growth exceeds output growth over the

long run in most OECD countries. Though it su¢ ces to show the existence of good-health

types that experience consumption growth higher than output growth, it is of great interest

to calibrate the share of the good-health types in the Survey of Consumer Finances. Also, as

long run wealth inequality depends on the share of the best-health type, it is natural to ask

how large it is. The health types in the model imply di¤erent growth rates of consumption and

wealth across types and individuals, which then implies type dependence and scale dependence

(Gabaix et al. 2016). In this section, we calibrate the distribution of health types to match

the evolution of wealth inequality in the US measured by the coe¢ cient of variation. In our

calibration exercise, health types are de…ned by consumption growth.12

1
The time preference rate ~ and the intertemporal elasticity of substitution are set equal

to 1% and 1, respectively. The growth rate g is set equal to the average annual growth rate based

on the data presented in Section 2. Consumption growth rates of the worst-health type and

the best-health type are set respectively equal to the lowest and highest rates of consumption

growth in the data. We set the rate of return on wealth r to 7:92% as reported by Cao and Luo
12
Di¤erent health types are de…ned by di¤erent death rates, which imply di¤erent rates of consumption
growth.

35
(2017)13 . We choose the shares of health types (the distribution of health types) to match the

coe¢ cient of variation for wealth in 2016 using the coe¢ cients of variation for wealth and labour

income in 1989 as initial conditions. In e¤ect, we solve the following minimization problem

min CVamodel (2016) CVadata (2016) ;

M
f j gj=1

where CVamodel (2016) is given from (28b) and CVadata (2016) is given from Table 1. We iterate

the calibration exercise for M = 2; 4; :::; 100. In the calibration, we always allow for the average-

health type, bad-health types, and good-health types. When M = 2, there are a good-health

type and the average-health type only. The optimal number of health types is the one that

produces the smallest absolute value of the di¤erence betwen the coe¢ cient of variation in the

model and that in the data.

Table 2 Parameterization

Parameter Description Value

1
Intertemporal elasticity of subsitution 1

~ Discount rate 1%

r Rate of return on wealth 7:92%

g Output growth rate 2:6%

M Number of health types 12

g1c Consumption growth rate of the best-health type 5:7%

c
gM Consumption growth rate of the worst-health type 1:3%

1 Share of the best-health type 6:2%

With 12 health types (6 bad-health types, 5 good-health types, and the average-health

type), the model …ts the empirical coe¢ cient of variation in 2016 perfectly (see Figure 5 in
13
This is the annualized rate of return on wealth with capital gains (Cao and Luo 2017 , …g. 1).

36
app. H for detail)14 . Table 2 shows that the share of the best-health type that dominates the

other health types in the long run is 6:2% (the asterisk in Figure 4). This …nding is somewhat

surprising, but there is empirical evidence that top 5% wealth holders own approximately 64:6%

of total wealth, and that their wealth share is rising (Saez and Zucman 2016). The share of

good-health types is 30:5%.

Figure 4 displays the health types characterized by consumption growth rates and their cor-

responding share. The circle in the graph represents the average-health type that experiences

consumption growth rate of 2:6% (as high as the growth rate). The growth rates of consump-

tion of bad-health types range between 1:3% and 2:6% while good-health types experience

consumption growth rates between 2:6% and 5:7%.

0.15
average-health type
share of health types

best-health type

0.1

0.05
-0.02 0 0.02 0.04 0.06
health types by consumption growth

Figure 4 Share of health types by consumption growth

7 Conclusion and outlook

This paper uncovers the e¤ects of the di¤erence between the rate of return on wealth and the

growth rate on long run wealth inequality for a small open economy. Households draw initial

wealth and initial labour income from given distributions and face a constant instantaneous

probability of death. The death rate is also drawn from a distribution, which catures ex-ante
14
This perfect …t is not surprising as the health types imply type dependence and scale dependence that
generate fast transitions (Gabaix et al. 2016).

37
heterogeneity in health condition. The realized death rate of a bad-health type is lower than a

threshold while a good-health type draws a death rate higher than the threshold. The death

rate of the average-health type is equal to the threshold, which implies consumption growth

equal to output growth. Thus, the bad-health types experience consumption growth lower

than output growth while consumption growth of the good-health types is greater than output

growth.

Households accumulate wealth according to two e¤ects: additive e¤ect and scale e¤ect.

The additive e¤ect depends on labour income and therefore di¤ers across households. The

scale e¤ect is related to consumption growth and is identical across households that draw an

identical death rate. The best-health-type households draw the smallest death rate and hence

experience the biggest scale e¤ect on wealth. These households therefore experience fastest

wealth accumulation, and their wealth share dominates the wealth share of other households in

the long run. Thus, long run wealth inequality is characterized by the best-health type. These

households draw the lowest death rate and therefore experience identical scale e¤ect on wealth

accumulation (due to identical consumption growth). As wealth inequality is measured by the

coe¢ cient of variation which is scale-independent, only the additive e¤ect matters for long run

wealth inequality.

The additive e¤ect in the long run is equal to the present value of lifetime labour income.

An increase in r g reduces the additive e¤ect, which causes two simultaneous e¤ects on long

run wealth inequality. First, it leads to decreasing long run inequality due to the heterogeneity

in labour income. Second, it leads to rising long run inequality due to the heterogeneity in

wealth endowment. The more dispersed the distribution of labour income, the stronger the

…rst e¤ect. A higher dispersion index of the initial wealth distribution strengthens the second

e¤ect. We …nd a positive threshold of r g, above which long run wealth inequality is a rising

function r g. The threshold is increasing the dispersion of labour income but decreasing in

38
the dispersion of wealth endowment.

In addition to the dispersion indices of labour income and wealth endowment, the r g’s

threshold is a function of the covariance between wealth and labour income. Using the data on

wealth and labour income from the Survey of Consumer Finances 1989-2016 we …nd that the

r g’s threshold is 4:4% on average, which is empirically plausible since the di¤erence r g

observed in the data is close to a value of 5:32%.

We calibrate the distribution of health types to match the evolution of wealth inequality in

Survey of Consumer Finances 1989-2016. Starting from the distributions of wealth and labour

income in 1989, the shares of health types are chosen such that the coe¢ cient of variation

in 2016 in the model …ts that in the data. With 12 health types, the model perfectly …ts

the empirical coe¢ cient of variation in 2016. Of 12 health types, there are 5 good-health

types, which account for 30:5% of population. The share of the best-health type that becomes

a “wealth dictator” in the long run is 6:2%. This …nding is greatly supported by empirical

evidence that top 5% wealth holders account for approximately 64:6% of total wealth and that

their wealth share is rapidly rising (Saez and Zucman 2016).

Our model predicts that a good-health-type individual experiences higher consumption

growth as she has an incentive to save more than the bad-health-type individual. Though

a health type in our model is de…ned through a death rate, a healthy individual and an un-

healthy individual in a survey can be distinguished e.g. by considering their days of sick leave or

how they report their health condition. Thus, depending on how a healthy individual is de…ned

in a survey, this gives rise to a question: Do healthy individuals save more than unhealthy

individuals? Another prediction by the model is that the smaller the share of the healthiest in-

dividuals, the higher the level of wealth inequality. A question related to this prediction is: Are

higher medical costs associated with rising wealth inequality? A further related question would

be: Does health insurance reduce wealth inequality? These would be interesting questions for

39
future research.

Labour income inequality is one of the drivers of wealth inequality. Our …nding of the

U-shaped e¤ect further shows that labour income inequality is crucial for understanding the

e¤ects of r g on wealth inequality. Future empirical research that investigates e¤ects of

interest rates and output growth on wealth distributions should not ignore the role of labour

income heterogeneity.

8 Appendix

A The growth rates of wage, capital and output

We start from the optimal rules of the …rm

1
K (t)
r = A (t) ; (A.1)
L
K (t)
wL (t) = (1 ) A (t) : (A.2)
L

The capital-labour ratio K (t) =L given by (A.1) is a function of total factor productivity A(t),

the international interest rate r, and the elasticity of output with respect to capital

1
K (t) 1 1
= (A (t)) 1 ,t 0: (A.3)
L r

The evolution of capital stock therefore reads

K_ (t) 1 _
A(t)
=
K (t) 1 A(t)
gA
= g
1

40
The initial capital stock demanded to produce the …nal good at time t = 0 is given by

1 1
1
K0 = A01 L:
r

Rewriting the equilibrium wage using (A.3) yields

1 1
wL (t) = (1 ) (A (t)) 1 : (A.4)
r

Taking log and deriving with respect to time yields

w_ L (t) 1 _
A(t)
L
=
w (t) 1 A(t)
gA
= g:
1

This ODE has a solution of a form wL (t) = w0L egt . Given the production function Y (t) =

A(t)K (t) L1 , taking log and deriving with respect to time yields

Y_ (t) A_ (t) K_ (t)

= +
Y (t) A (t) K (t)
gA
= gA +
1
gA
= g:
1

Solving this simple ODE gives Y (t) = Y0 egt .

B The representative …rm in the presence of knowledge spillovers

In the presence of knowledge spillovers, total factor productivity A (t) evolves according to

A (t) = AegA t K(t) L , A > 0, (gA ; ) 2 (0; 1)2 ; (A.5)

41
which captures the fact that an increase in capital stock leads to a higher level of technology

through knowledge spillovers (Arrow 1962; Romer 1986; Romer 1994). The evolution of total

factor productivity (2) also implies that an increase in the total supply of labour lowers the level

of technology. This is because …rms have less incentive to “discover and implement labour-saving

innovations” (Romer 1994). Equation (A.5) therefore captures positive externality of capital

investment and negative externality of growing labour supply. As the e¤ect of an increase in

capital K or labour L on total factor productivity A through Equation (A.5) is an external

e¤ect, the exponent measures the private e¤ect of an increase in capital on output. The

…rm chooses capital K and labour L to maximize its pro…t taking into account their private

e¤ects. The …rm’s optimal rule implies the marginal productivity of capital is equal to the

internationally given rate of return r

1
K (t)
r = A (t) : (A.6)
L

Rewriting (A.6) taking into account the evolution of technology (2) yields

1
K (t)
r = AegA t ; (A.7)
L

where = + measures the aggregate e¤ect of an increase in capital. The equilibrium wage

is equal to the marginal productivity of labour

K (t)
wL (t) = (1 ) AegA t : (A.8)
L

The capital-labour ratio K (t) =L given by (A.7) is a function of, inter alia, the rate of return r

1 1
K (t) 1 1
= AegA t ,t 0: (A.9)
L r

42
The initial capital stock demanded to produce the …nal good at time t = 0 is therefore given

by
1
1 1
K0 = A 1 L:
r

Rewriting the equilibrium wage using (A.9) yields

1
1 1
L gA t
w (t) = (1 ) Ae : (A.10)
r

We now compute the growth rates of wage and output. We …rst rewrite the evolution of

total factor productivity using the capital-labour ratio

A (t) = AegA t K(t) L

K (t)
= AegA t
L
1
= AegA t AegA t 1

r
1 +
1 1
= AegA t ; (A.11)
r

where
1
K (t) 1 1
= AegA t 1
(A.12)
L r

Taking log and then deriving with respect to time gives

_
A(t) 1 +
= gA
A(t) 1
1
= gA
1

Wage is given by
1
1
wL (t) = (1 ) AegAt 1
:
r

43
 Taking log and computing time derivative gives

w_ L (t) gA
L
= g:
w (t) 1

1
Solving this simple ODE of wL (t) results in wL (t) = w0L egt , where w0L = (1 ) A1 r
1
.

Similarly, taking log and computing time derivative of the capital-labour ratio (A.12) yields

K_ (t) gA
= :
K (t) 1

Given the production function Y (t) = A(t)K (t) L1 , taking log and deriving with respect

to time yields

Y_ (t) A_ (t) K_ (t)

= +
Y (t) A (t) K (t)
(1 ) gA gA
= +
1 1
gA
= g:
1

1
Solving this simple ODE yields Y (t) = Y0 egt , where Y0 = r
1
A 1 L.

C The intertemporal budget constraint

The dynamic budget constraint can be written as

a_ (t) ra (t) = w (t) c (t) :

rt
Multiplying both sides by e gives

rt rt
e [a_ (t) ra (t)] = e [w (t) c (t)] ;

44
 d
of which the left hand side is dt
e rt a (t). Hence, this equation reads

rt rt
de a (t) = e [w (t) c (t)] dt:

Integrating from 0 to T yields

Z T Z T
rt rt
de a (t) = e [w (t) c (t)] dt
0 0
Z T
rt T rt
e a (t) 0
= e [w (t) c (t)] dt
0

Z T
rT rt
e a (T ) a (0) = e [w (t) c (t)] dt
0
Z T
rT rt
e a (T ) = e [w (t) c (t)] dt + a (0) :
0

The no-Ponzi game condition implies

Z T
rT rt
lim e a (T ) = lim e [w (t) c (t)] dt + a (0)
T !1 T !1 0

= 0;

which implies

Z 1 Z 1
rt rt
e c (t) dt = a (0) + e w (t) dt
0 0
Z 1
[r g]t
= a (0) + w(0) e dt:
0

D The optimal behaviour

The household’s problem reads

45
 Z 1
t
max
1
U (0) = e u (c(t)) dt (A.13)
fc(t)gt=0 0

subject to

dw(t) = gw(t)dt (A.14)

a_ (t) = ra(t) + w (t) c (t) : (A.15)

The household chooses a consumption path to maximize her intertemporal utility (A.13) subject

to the law of motion of wage (A.14) and the budget constraint (A.15). Let us denote the value

function by V (a; w) that maps the two dimensional state space (a; w) into real numbers through

a mapping V . The Bellman equation for the household’s problem therefore reads

dV (a(t); w(t))
V (a(t); w(t)) = max u(c(t)) + :
c(t) dt

Computing the di¤erential dV (a(t); w(t)), taking the constraints (A.14) and (A.15) into

account yields, suppressing the time argument t for simplicity,

V (a; w) = max fu(c) + V1 (a; w)(ra + w c) + gwV2 (a; w)g (A.16)

c

The …rst-order condition for this problem reads

u0 (c(a; w)) = V1 (a; w): (A.17)

1
1 (a+ 2 w) 3
We guess V (a; w) = 1
, 1 6= 0. Using the …rst order condition (A.17) yields

c(a; w) = (a + 2 w); (A.18)

46
 1
where 1 . Plugging (A.18) into the Bellman equation (A.16) yields

h 1 i
1
(a + 2 w) 1 + 1 [(1 )r ]

+ w(a + 2 w) 1 (1 )(g 2 +1 r 2) + 3 1 = 0: (A.19)

As Equation (A.19) must hold for any a and w, it implies

(1 )r
= ; (A.20)
1
2 = ; (A.21)
r g

1
3 = : (A.22)

Thus, the optimal consumption reads

(1 )r w (t)
c (t) = a (t) + : (A.23)
r g

E The evolution of wealth over time

Inserting the closed form solution (A.23) into the budget constraint (A.15) yields

a_ (t) g c a ( ) = w0 egt ; (A.24)

where

r
gc ;
r ( + g)
:
(r g)

gc t
Multiplying (A.24) by e and integrating gives

47
 gc t w0 (g g c )t
a(t)e = e + k;
g gc

where k is the constant of integration. Further simpli…cation gives

w0 c
a(t) = egt + keg t : (A.25)
r g

w0
Evaluating (A.25) at t = 0 gives k = a0 + r g
. Thus, the evolution of wealth is described

by

w0 g c ]t c
a(t) = a0 + 1 e[g eg t : (A.26)
r g

F Expectation and variance of an exponential of a random variable

Consider a random variable X. We would like to compute E ef (X) and var ef (X) . They are

given by

X
E ef (X) = ef (xj ) j ;

2
var ef (X) = E e2f (X) E ef (X)
X X 2
= e2f (xj ) j ef (xj ) j :

When X = and f (X) = g c t, we obtain

X
M
c
gc t
E e = egj t j ;
j=1
!2
X
M X
M
gc t 2gjc t gjc t
var e = e j e j :
j=1 j=1

48
We now consider

w0 g c ]t ct
E [a(t)j ] = a0 + 1 e[g eg
r g
w0 ct w0
= a0 + eg egt :
r g r g

When is random, var[E [a(t)j ]] is given by

w0 ct
var [E [a(t)j ]] = var a0 + eg
r g
2
w0 c
= a0 + var eg t
r g
2 !2 3
2 X M X
M
w0 c c
= a0 + 4 e2gj t j egj t j
5:
r g j=1 j=1

G Proofs

G.1 Proof of Proposition 1

(i) The coe¢ cient of variation in the long run reads

r
w0
PM 2gj t c PM gj t c 2
a0 + r g j=1 e j j=1 e j
lim CVa (t) = lim PM h i c
t!1 t!1
+ r w0g 1 e [g gjc ]t
egj t j
j=1 a0
s
PM 2gc t PM c 2
2g1c t
a0 + r g
w0
j=1 e
j
j j=1 egj t j e
= lim PM h i
e[g gjc ]t
e[gj g1c ]t
t!1 w0 c
j=1 a0 + r g
1 j

r
w0
PM 2[gj c g1c ]t PM [gjc g1c ]t
2
a0 + r g j=2 e j + 1 j=2 e j + 1
= lim P h i h i
e[g gjc ]t
e[gj g1c ]t
e[g g1c ]t
t!1 M w0 c w0
j=2 a0 + r g
1 j + a0 + r g
1 1
r
PM PM 2
e2[gj g1c ]t
e[gj g1c ]t
w0 c c
a0 + r g j=2 j + 1 j=2 j + 1
= lim P h i h i
t!1 M [gjc g1c ]t w0
e[gj
c g1c ]t
e[g g1c ]t w0
e[g g1c ]t
j=2 a0 e + r g j + a0 + r g
1 1

49
 As g1c > gjc holds for any j > 1, limt!1 e[gj g1c ]t
c
= 0. Also, since there exists a k,

k 2 f2; 3; :::; M 1g such that gkc = g holds, g1c > g and therefore limt!1 e[g g1c ]t
= 0

hold. Therefore we obtain

r
1 1
lim CVa (t) = :
t!1 1

(ii) This directly follows (i).

G.2 Proof of Proposition 2

(i) The coe¢ cient of variation in the long run reads

v
u
u PM 2 2
e[g gjc ]t c
u
u j=1 (r g)2
w0
1 e2gj t j
u
t w0
2 PM 2gj t c PM gj t c 2
+ a0 + r g j=1 e j j=1 e j
lim CVa (t) = lim PM h i
e[g gjc ]t
t!1 t!1 c
j=1 a0 + r g
w0
1 egj t j
v
u
u PM 2 2
e[g gjc ]t c

c
u
u j=1 (r g)
w0
2 1 e2gj t j
e g1 t u
t 2 PM c PM c 2
+ a0 + w0
r g j=1 e2gj t j j=1 egj t j
= lim PM h i
e[g gjc ]t
t!1 c
e g1c t + w0
1 egj t
j=1 a0 r g j
v
u
u PM 2 2
e[g gjc ]t
e2[gj g1c ]t
c
u w0
1 j
u j=1 (r g)2
u
t 2 PM PM 2
e2[gj g1c ]t
e[gj g1c ]t
w0 c c
+ a0 + r g j=1 j j=1 j
= lim PM h i
e[g gjc ]t
e[gj g1c ]t
t!1 w0 c
j=1 a0 + r g
1 j
v
u PM
u 2 2 2 2
e[gj g1c ]t
e[g g1c ]t
e[g g1c ]t
w0 c w0
u j=2 (r g)2 j + 2 1 1
u (r g)
u
t 2 PM PM 2
e2[gj g1 ]t e[gj g1c ]t
w0 c c c
+ a0 + r g j=2 j + 1 j=2 j + 1
= lim P h i c h i
t!1 M
+ w0
1 e[ g gjc ]t
e[gj g1c ]t
+ + w0
1 e[g g1c ]t
j=2 a0 r g j a0 r g 1

50
 v
u PM
u 2 2 2 2
e[gj g1c ]t
e[g g1c ]t
e[g g1c ]t
w0 c w0
u j=2 (r g)2 j + (r g)2
1 1
u
u
t w0
2 PM 2[gj c g1c ]t PM [gjc g1c ]t
2
+ a0 + r g j=2 e j + 1 j=2 e j + 1
= lim P h i h i
t!1 M [gjc g1c ]t w0
e[gj
c g1c ]t
e[g g1c ]t w0
e[g g1c ]t
j=2 a0 e + r g j + a0 + r g
1 1

As g1c > gjc holds for any j > 1, limt!1 e[gj g1c ]t
c
= 0. Also, since there exists a k,

k 2 f2; 3; :::; M 1g such that gkc = g holds, g1c > g and therefore limt!1 e[g g1c ]t
= 0

hold. Therefore we obtain

v
u 2
u w0
1 1
u (r g)2 1
lim CVa (t) = t 2 + :
t!1 w0 1 1
a0 + r g

(ii) Deriving the long run coe¢ cient of variation with respect to r g gives

d 1 1 d
lim CVa (t) = [F (r g)] 2 F (r g) ,
d(r g) t!1 2 d (r g)

where
2
w0
(r g)2 1 1 1
F (r g) = 2 + :
w0 1 1
a0 + r g

d
d(r g)
limt!1 CVa (t) < 0 holds if and only if

d
F (r g) < 0
d (r g)
2 2 2 2
w0 w0
1 (r g) 3 a0 + r g
2 (r wg)0 2 a0 + r g
w0
(r g)2
w0

4 <0
1 w0
a0 + r g

h 2 2 i
w0 1 w0 w0 w0 w0
1 2 a0 + r g (r g)2 (r g) 2 r g a0 + r g
4 <0
1 w0
a0 + r g
2
w0
a0 < 0;
r g

51
 which always holds.

G.3 Proof of Lemma 1

Long run wealth inequality reads

v
u 2
u w0
1 1
u (r g)2 1
lim CVa (t) = t 2 +
t!1 w0 1 1
a0 + r g
s
2
w0 1 1 1
= 2 + (r g) :
((r g) a0 + w0 ) 1 1

Consider the following limits

s
2
w0 1 1 1
lim (r g) = + ;
(r g)!0 w0 1 1
r
1 1 ~:
lim (r g) =
(r g)!1 1

According to Proposition 2, long run wealth inequality falls from to ~ as the di¤erence r g

rises. Notice that, as w0 = w0 is positive, is a one-to-one function of w0 = w0 . Thus, a higher

value of ( ~ unchanged) implies a higher slope of (r g) at each value of r g, i.e. (r g)

decreases faster. In other words, when is bigger, the decreasing e¤ect of r g on (long run

2
inequality) is stronger. Clearly, a higher dispersion index w0 = w0 implies a higher coe¢ cient

of variation w0 = w0 and therefore a higher value of .

G.4 Proof of Proposition 3

(i) The coe¢ cient of variation in the long run reads

s
PM c
2 2gj t
2 PM c PM c 2
j=1 a0 e j + a0 + r g
w0
j=1 e2gj t j j=1 egj t j

lim CVa (t) = lim PM h i

e[g gjc ]t
t!1 t!1 c
j=1 a0 + r g
w0
1 egj t j

52
 s
g1c t
PM c
2 2gj t
2 PM c PM c 2
e j=1 a0 e j + a0 + r g
w0
j=1 e2gj t j j=1 egj t j

= lim PM h i
e[g gjc ]t
t!1 c
e g1c t + w0
1 egj t
j=1 a0 r g j
s
PM 2 2[gj g1 ]t
2 PM PM 2
e2[gj g1c ]t
e[gj g1c ]t
c c w0 c c
j=1 a0 e j + a0 + r g j=1 j j=1 j

= lim PM h i
e[g gjc ]t
e[gj g1c ]t
t!1 w0 c
j=1 a0 + r g
1 j
v
u
u PM 2 2[gjc g1c ]t
u a0 e j + 2
a0 1
u j=2
u
t 2 PM PM 2
e2[gj g1 ]t e[gj g1c ]t
w0 c c c
+ a0 + r g j=2 j + 1 j=2 j + 1
= lim P h i c h i
t!1 M
+ w0
1 e[ g gjc ]t
e[gj g1c ]t
+ + w0
1 e[g g1c ]t
j=2 a0 r g j a0 r g 1

v
u
u PM 2 2[gjc g1c ]t
u a0 e j + 2
a0 1
u j=2
u
t w0
2 PM 2[gj c g1c ]t PM [gjc g1c ]t
2
+ a0 + r g j=2 e j + 1 j=2 e j + 1
= lim P h i h i
t!1 M [gjc g1c ]t w0
e[gj
c g1c ]t
e[g g1c ]t w0
e[g g1c ]t
j=2 a0 e + r g j + a0 + r g
1 1

As g1c > gjc holds for any j > 1, limt!1 e[gj g1c ]t
c
= 0. Also, since there exists a k,

k 2 f2; 3; :::; M 1g such that gkc = g holds, g1c > g and therefore limt!1 e[g g1c ]t
= 0

hold. Therefore we obtain

v
u 2
1 1
lim CVa (t) = u a0 1
t 2 + :
t!1 w0 1 1
a0 + r g

(ii) Deriving the long run coe¢ cient of variation with respect to r g gives

d 1 1 d
lim CVa (t) = [P (r g)] 2 P (r g) ,
d(r g) t!1 2 d (r g)

where
2
a0 1 1 1
P (r g) = 2 + :
w0 1 1
a0 + r g

53
 d
d(r g)
limt!1 CVa (t) > 0 holds if and only if

d
P (r g) > 0
d (r g)
2 w0 w0
1 2 a0 a0 + r g (r g)2
4 > 0;
1 w0
a0 + r g

which always holds.

G.5 Proof of Lemma 2

Long run wealth inequality reads

v
u 2
1 1
lim CVa (t) = u a0 1
t 2 + (r g) :
t!1 w0 1 1
a0 + r g

Consider the following limits

r
1 1 ~;
lim (r g) =
(r g)!0 1
s
2
a0 1 1 1
lim (r g) = + :
(r g)!1 a0 1 1

According to Proposition 3, long run wealth inequality rises from ~ to as the di¤erence

r g rises. Notice that, as a0 = a0 is positive, is a one-to-one function of a0 = a0 . Thus, a

higher value of ( ~ unchanged) implies a higher slope of (r g) at each value of r g, i.e.

(r g) increases faster. In other words, when is bigger, the increasing e¤ect of r g on

2
(long run inequality) is stronger. Clearly, a higher dispersion index a0 = a0 implies a higher

coe¢ cient of variation a0 = a0 and therefore a higher value of .

54
G.6 Proof of Proposition 4

(i) The coe¢ cient of variation in the long run reads

v
u PM
u 2 2
e[g gjc ]t c
u j=1
2
a0 + w0
2 1 e2gj t j
u (r g)
u
t w0
2 PM 2gj t c PM gj t c 2
+ a0 + r g j=1 e j j=1 e j
lim CVa (t) = lim PM h i
e[g gjc ]t
t!1 t!1 c
j=1 a0 + r g
w0
1 egj t j
v
u PM
u 2 2
e[g gjc ]t c
u j=1
2
a0 + w0
2 1 e2gj t j
c u (r g)
e g1 t u
t 2 PM c PM c 2
+ a0 + w0
r g j=1 e2gj t j j=1 egj t j
= lim PM h i
e[g gjc ]t
t!1 c
e g1c t + w0
1 egj t
j=1 a0 r g j
v
u
u PM 2 2[gjc g1c ]t
PM 2 2
e[g gjc ]t
e2[gj g1c ]t
c
u j=1 a0 e j + w0
1 j
u j=1 (r g)2
u
t 2 PM PM 2
e2[gj g1c ]t
e[gj g1c ]t
w0 c c
+ a0 + r g j=1 j j=1 j
= lim PM h i
e[g gjc ]t
e[gj g1c ]t
t!1 w0 c
j=1 a0 + r g
1 j
v
u PM 2 2[gc gc ]t
u 2
u j=2 a0 e
j 1
j + a0 1
u
u P
u 2
[gjc g1c ]t e[g g1c ]t
2 2
[g g1c ]t
2
u + M w0
j=2 (r g) 2 e j + w0
2 1 e 1
u (r g)
u
t 2 P
M 2[gjc g1c ]t PM [gc gc ]t 2
+ a0 + r w0g j=2 e j + 1 j=2 e
j 1
j + 1
= lim P h i c c h i
t!1 M w0 [ g gjc ]t [ gj g1 ]t w0 [ g g1c ]t
j=2 a0 + r g 1 e e j + a0 + r g 1 e 1

v
u PM 2 2[gc gc ]t
u 2
u j=2 a0 e
j 1
j + a0 1
u
u P
u 2
[gjc g1c ]t e[g g1c ]t
2 2
[g g1c ]t
2
u + M j=2
w0
2 e j + w0
2 1 e 1
u (r g) (r g)
u
t 2 P
M 2[gjc g1c ]t PM [gc gc ]t 2
+ a0 + r w0g j=2 e j + 1 j=2 e
j 1
j + 1
= lim P h i h i
t!1 M [gjc g1c ]t w0 [gjc g1c ]t [ g g1c ]t w0 [ g g1c ]t
j=2 a0 e +r g e e j + a0 + r g 1 e 1

As g1c > gjc holds for any j > 1, limt!1 e[gj g1c ]t
c
= 0. Also, since there exists a k,

k 2 f2; 3; :::; M 1g such that gkc = g holds, g1c > g and therefore limt!1 e[g g1c ]t
= 0

55
 hold. Therefore we obtain

v
u 2
u 2 + w0
1 1
u a0 (r g)2 1
lim CVa (t) = t 2 + :
t!1 w0 1 1
a0 + r g

(ii) Deriving the long run coe¢ cient of variation with respect to r g gives

d 1 1 d
lim CVa (t) = [G (r g)] 2 G (r g) ,
d(r g) t!1 2 d (r g)

where
2
2 w0
a0 + (r g)2 1 1 1
G (r g) = 2 + :
w0 1 1
a0 + r g

d
d(r g)
limt!1 CVa (t) ? 0 holds if and only if

d
G (r g) ? 0
d (r g)
2 2 2 2
w0 w0 2 w0 w0 w0
1 (r g) 3 a0 + r g
2 a0 + (r g)2 a0 + r g (r g)2
4 ?0
1 w0
a0 + r g
h 2 2 i
w0 1 2 w0 w0 w0
1 2 a0 + r g (r g)2 a0 + (r g)2 w0 r g a0 + r g
4 ?0
1 w0
a0 + r g

2 2
?0
2 w0 w0 w0
a0 + 2 w0 a0 +
(r g) r g r g
2
a0
2
a0 w0
w0
?0
r g
2
a0
2
a0 w0 ? w0
r g
2
w0 = w0
r g? 2
a0 = a0

56
G.7 Proof of Proposition 5

(i) The coe¢ cient of variation in the long run reads

v
u P
u M 2 2
e[g gjc ]t
e[g gjc ]t c
u j=1 2
a0 + w0
2 1 +2 a0 w0
1 e2gj t j
u (r g) r g
u
t 2 PM c PM c 2
+ a0 + r g
w0
j=1 e2gj t j j=1 egj t j
lim CVa (t) = lim PM h i
e[g gjc ]t
t!1 t!1 c
j=1 a0 + w0
r g
1 egj t j

v
u P
u M 2 2
e[g gjc ]t
e[g gjc ]t c
u j=1 2
a0 + w0
2 1 +2 a0 w0
1 e2gj t j
g1c t u
(r g) r g
e u
t w0
2 PM 2gj t c gj tPM c 2
+ a0 + r g j=1 e j j=1 e j
= lim PM h i c
e[g gj ]t egj t j
t!1 c
e g1c t + w0
1
j=1 a0 r g

v
u
u PM 2 2[gjc g1c ]t
PM 2 2
e[g gjc ]t
e2[gj g1c ]t
c
u j=1 a0 e j + w0
1 j
u j=1 (r g)2
u
u PM
e[g gjc ]t
e2[gj g1c ]t
c
u +2 a0 w0
1 j
u j=1 r g
u
t w0
2 PM 2[gj c g1c ]t PM [gjc g1c ]t
2
+ a0 + r g j=1 e j j=1 e j
= lim PM h i
e[g gjc ]t
e[gj g1c ]t
t!1 w0 c
j=1 a0 + r g
1 j

v
u
u PM 2 2[gc gc ]t
u 2
j=2 a0 e j + a0 1
j 1
u
u
u PM 2 2 2 2
u + w0
e[ gjc g1c ]t
e[ g g1c ]t w0
j + (r g)2 1 e[g g1c ]t
u j=2 (r g)2 1
u
u hP i
u
e2[gj g1 ]t e[(g g1 )+(gj g1 )]t j + ra0 wg0 1 e[g g1c ]t
M a0 w0 c c c c c
u +2 j=2 r g 1
u
u 2 P PM [gc gc ]t 2
t M 2[gjc g1c ]t
+ a0 + r w0g j=2 e j + 1 j=2 e
j 1
j + 1

= lim PM h i h i
e[g gjc ]t
e[gj g1c ]t
e[g g1c ]t
t!1 w0 c w0
j=2 a0 + r g
1 j + a0 + r g
1 1

57
 v
u
u PM 2 2[gc gc ]t
u 2
j=2 a0 e j + a0 1
j 1
u
u
u P 2 2 2 2
u + M w0 [gjc g1c ]t e[g g1c ]t w0
e[g g1c ]t
u j=2 (r g)2 e j + (r g)2 1 1
u
u hP i
u
e2[gj g1 ]t e[(g g1 )+(gj g1 )]t j + ra0 wg0 1 e[g g1c ]t
M a0 w0 c c c c c
u +2 j=2 r g 1
u
u 2 P PM [gc gc ]t 2
t M 2[gjc g1c ]t
+ a0 + r w0g j=2 e j + 1 j=2 e
j 1
j + 1

= lim P h i h i :
t!1 M [gjc g1c ]t w0
e[gj
c g1c ]t
e[g g1c ]t w0
e[g g1c ]t
j=2 a0 e + r g j + a0 + r g
1 1
(A.27)

As g1c > gjc holds for any j > 1, limt!1 e[gj g1c ]t
c
= 0. Also, since there exists a k,

k 2 f2; 3; :::; M 1g such that gkc = g holds, g1c > g and therefore limt!1 e[g g1c ]t
= 0

hold. Therefore we obtain

v
u 2
u 2 + w0
+2 a0 w0
1 1
u a0 (r g)2 r g 1
lim CVa (t) = t 2 + : (A.28)
t!1 w0 1 1
a0 + r g

(ii) Deriving the long run coe¢ cient of variation with respect to r g gives

d 1 1 d
lim CVa (t) = [Q (r g)] 2 Q (r g) ,
d(r g) t!1 2 d (r g)

where
2
2 w0 a0 w0
a0 + (r g)2
+2 r g 1 1 1
Q (r g) = 2 + :
w0 1 1
a0 + r g

d
d(r g)
limt!1 CVa (t) ? 0 holds if and only if

d
Q (r g) ? 0
d (r g)
2 2 2
2
1 2 (r wg)0 3 2 (ra0g)
w0
2 a0 + w0
r g
2 a0 + (r g)2
w0
+2 a0 w0
r g a0 + r g
w0 w0
(r g)2
4 ?0
1 w0
a0 + r g

58
 h 2 2 i
w0 1 2 w0 a0 w0 w0 w0
1 2 a0 + r g (r g)2 a0 + (r g)2
+2 r g w0 r g
+ a0 w0 a0 + r g
4 ?0
1 w0
a0 + r g
2 2
?0
2 w0 a0 w 0 w0 w0
a0 + 2 +2 w0 + a0 w0 a0 +
(r g) r g r g r g
2
a0
?0
2 w0 a0 w 0 w0
a0 w0 a0 w 0 a0
r g r g

This is equivalent to

2
a0 a0 w 0 w0
2
a0 w0 a0 w 0 a0 ? w0
: (A.29)
r g

2
Under Assumption 2, a0 w0 a0 w 0 a0 > 0 holds. (A.29) is therefore equivalent to

2
w 0 a0 a0 w 0 w0
r g? 2
a0 w 0 a0 w 0 a0
2
w0 = w0 a0 w 0 = a0
r g? 2
:
a0 = a0 a0 w 0 = w 0

G.8 Proof of Proposition 6

(i) Deriving the threshold with respect to the covariance a0 w 0 yields

2 1 2 1
d a0 = a0 a0 w 0 = w 0 a0 w0 = w0 a0 w 0 = a0 w0
= 2 :
d a0 w 0 2
a0 = a0 a0 w0 = w0

Rearranging this gives

2 2 2 2
d w0 = w0 a0 = a0
= 2:
d a0 w 0 2
a0 = a0 a0 w0 = w0

d
d a0 w0
< 0 holds if and only if

2 2 2 2
w0 = w0 a0 = a0 <0

2 2 2 2
w0 = w0 < a0 = a0

59
 a0 w0
> ;
a0 w0

which is true as CVa0 > CVw0 is assumed. So is decreasing in a0 w 0 . Let us compute

the limit

2
w0 = w0 a0 w 0 = a0
lim = lim 2
a0 w0 !0 a0 w0 !0 a0 = a0 a0 w 0 = w 0
2
w0 = w0
= 2
;
a0 = a0

and the limit

2
w0 = w0
2
w0 = w0 a0 w 0 = a0 a0 w0
1= a0
lim = lim 2
= lim 2
a0 = a0
a0 w0 !1 a0 w0 !0 a0 = a0 a0 w 0 = w0 a0 w0 !0
1= w0
a0 w0

1= a0 w0
= = > 0:
1= w0 a0

This implies that the threshold is strictly positive and less than the ratio of the disper-
2
w0 = w0
sion indices, i.e. 2 0; 2 .
a0 = a0

d
(ii) d a0 w0
> 0 holds if and only if

2 2 2 2
w0 = w0 a0 = a0 >0

2 2 2 2
w0 = w0 > a0 = a0

a0 w0
<
a0 w0

CVa0 < CVw0 ;

which is true by assumption. So is increasing in a0 w 0 . Consider the limit

2
w0 = w0
lim = 2
;
a0 w0 !0 a0 = a0

60
 2
w0 = w0
which implies > 2 .
a0 = a0

(iii) Consider the following di¤erence

2 2
w0 = w0 a0 w 0 = a0 w0 = w0
= 2 2
a0 = a0 a0 w 0 = w 0 a0 = a0
2 2 2 2 2 2 2 2
w 0 a0 =( w0 a0 ) a0 w 0 a0 = a0 a0 w 0 = ( a0 w0 ) + a0 w 0 w0 = w0
= 2 2
a0 = a0 a0 w 0 = w 0 a0 = a0
2 2 2 2
a0 w 0 w0 = w0 a0 w 0 a0 = a0
= 2 2
a0 = a0 a0 w 0 = w0 a0 = a0
2 2 2 2
a0 w 0 w0 = w0 a0 = a0
= 2 2
a0 = a0 a0 w 0 = w0 a0 = a0
h i2 h i2
w0 a0
a0 w 0 w0 a0
= 2 2
:
a0 = a0 a0 w 0 = w 0 a0 = a0

2
w0 = w0
As CVa0 = CVw0 is assumed, this implies = 0 or = = 2 .
a0 = a0

G.9 Proof of Proposition 7

We would like to show

A ^ B =) C, (A.30)

where

A = “The di¤erence r g is greater than a positive threshold”,

B = “There exist good-health types”,

c = “Increasing r g leads to rising wealth inequality in the long run”.

It can be seen that Proposition 5 holds for a small open economy populated by bad-health

types, good-health types, and the average-health type (k < M ) or for a small open economy

populated by good-health types and the average-health type (k = M ). Thus, to show (A.30),

it is su¢ cient to show that Proposition 5 also holds for a small open economy populated by

61
good-health types only. That is, the proposition holds when gjc > g for all j = 1; 2; :::; M .

Recall the coe¢ cient of variation in the long run (A.27)

v
u
u PM 2 2[gc gc ]t
u 2
j=2 a0 e j + a0 1
j 1
u
u
u P 2 2 2 2
u + M w0 [gjc g1c ]t e[g g1c ]t w0
e[g g1c ]t
u j=2 (r g)2 e j + (r g)2 1 1
u
u hP i
u
e2[gj g1 ]t e[(g g1 )+(gj g1 )]t j + ra0 wg0 1 e[g g1c ]t
M a0 w0 c c c c c
u +2 j=2 r g 1
u
u 2 P PM [gc gc ]t 2
t M 2[gjc g1c ]t
+ a0 + r w0g j=2 e j + 1 j=2 e
j 1
j + 1

lim CV (t) = lim P h i h i :

t!1 t!1 M [gjc g1c ]t w0
e[gj
c g1c ]t
e[g g1c ]t w0
e[g g1c ]t
j=2 a0 e + r g j + a0 + r g
1 1

As gjc > g holds for all j 2 H, limt!1 e[g g1c ]t

= 0 holds. In addition, as g1c > gjc holds for any

j > 1, limt!1 e[gj g1c ]t

c
= 0 Therefore we obtain

v
u 2
u 2 + w0
+2 a0 w0
1 1
u a0 (r g)2 r g 1
lim CVa (t) = t 2 + ;
t!1 w0 1 1
a0 + r g

which is identical to A.28). By Proposition 5, long run wealth inequality measured by the

coe¢ cient of variation is a rising function of r g if the di¤erence r g is greater than the

positive threshold . This completes the proof.

H The empirical …t

So we solve the following minimization problem

min CVamodel (2016) CVadata (2016) ;

M
f j gj=1

where CVamodel (2016) is given from (28b) and CVadata (2016) is given from Table 1. This problem

is numerically solved for M = 2; 4; :::; 100. When M = 12, the di¤erence CVamodel (2016) CVadata (2016)

is smallest. Figure 5 shows that the …t for 1989 is perfect by construction. The …t then becomes

62
worse for 1992 2013 as these years are not targeted. The …t for 2016 is perfect again.

6.5

coefficient of variation
data
6 model

5.5

4.5

4
89
92
95
98
01
04
07
10
13
16
19
19
19
19
20
20
20
20
20
20
year

Figure 5 Coe¢ cients of variation in the data and model

I Consumption growth and output growth in OECD countries for

periods starting from 1980

GRC
1980-1990 GRC
1980-1995
GBR
1 GBR

1
percent

percent

USA
0.5 USA
OECD
0 AUS CAN FRA
0 AUS CAN FRA
OECD

SWE
NOR -0.5
-1 KOR KOR
NOR
SWE

1980-2000 GBR 1980-2005 GBR

GRC
GRC

0.5 0.5 USA

percent

percent

USA
AUS OECD
OECD
CAN
0 AUS
CAN FRA 0 FRA
NOR

-0.5 NOR SWE -0.5 SWE

-1 KOR
-1 KOR

GRC
1980-2010 GBR GRC
1980-2015
GBR
0.5
0.5 CAN USA
CAN NOR USA
AUS
AUS NOR OECD
percent
percent

FRA OECD FRA

0 0
SWE
SWE

-0.5 -0.5
KOR
-1 KOR
-1

Figure 6 Di¤erence between average consumption growth and average output growth in

OECD countries for various time periods starting from 1980

63
J Country codes

Table 3 Country codes of OECD countries

Country Code Country Code

Australia AUS Korea KOR

Austria AUT Latvia LVA

Belgium BEL Lithuania LTU

Canada CAN Luxembourg LUX

Chile CHL Mexico MEX

Czech Republic CZE Netherlands NLD

Denmark DNK New Zealand NZL

Estonia EST Norway NOR

Finland FIN Poland POL

France FRA Portugal PRT

Germany DEU Slovak Republic SVK

Greece GRC Slovenia SVN

Hungary HUN Spain ESP

Iceland ISL Sweden SWE

Ireland IRL Switzerland CHE

Isreal ISR Turkey TUR

Italy ITA United Kingdom GBR

Japan JPN Unitedt States USA

64
References

Acemoglu, D., and J. A. Robinson (2015): “The Rise and Decline of General Laws of Capital-

ism,”Journal of Economic Perspectives, 29(1), 3–28.

Aiyagari, S. R. (1994): “Uninsured Idiosyncratic Risk and Aggregate Saving,” Quarterly

Journal of Economics, 109, 659–84.

Allison, P. D. (1978): “Measures of Inequality,”American Sociological Review, 43(6), 865–880.

Arrow, K. J. (1962): “The Economic Implications of Learning by Doing,”Review of Economic

Studies, 29, 155–173.

Atkinson, A. B. (2017): “Pareto and the Upper Tail of the Income Distribution in the UK:

1799 to the Present,”Economica, 84(334), 129–156.

Bayer, C., A. D. Rendall, and K. Wälde (2019): “The invariant distribution of wealth and

employment status in a small open economy with precautionary savings,”Journal of Math-

ematical Economics, 85, 17–37.

Benhabib, J., and A. Bisin (2018): “Skewed Wealth Distributions: Theory and Empirics,”

Journal of Economic Literature, 56(4), 1261–91.

Benhabib, J., A. Bisin, and M. Luo (2017): “Earnings Inequality and Other Determinants of

Wealth Inequality,”American Economic Review: Papers & Proceedings, 107(5), 593–597.

Benhabib, J., A. Bisin, and S. Zhu (2011): “The Distribution of Wealth and Fiscal Policy in

Economies with Finitely Lived Agents,”Econometrica, 79(1), 123–157.

Blanchard, O. J. (1985): “Debt, De…cits, and Finite Horizons,” Journal of Political Economy,

93, 223–47.

65
Blume, L., and S. Durlauf (2015): “Capital in the Twenty-First Century: A Review Essay,”

Journal of Political Economy, 123(4), 749–777.

Cao, D., and W. Luo (2017): “Persistent Heterogeneous Returns and Top End Wealth Inequal-

ity,”Review of Economic Dynamics, 26, 301–326.

Card, D., and J. E. Dinardo (2002): “Skill-Biased Technological Change and Rising Wage

Inequality: Some Problems and Puzzles,”Journal of Labor Economics, 20(4), 733–783.

Fischer, T. (2017): “Thomas Piketty and the rate of time preference,” Journal of Economic

Dynamics and Control, 77, 111–133.

Gabaix, X., J.-M. Lasry, P.-L. Lions, and B. Moll (2016): “The Dynamics of Inequality,”

Econometrica, 84(6), 2071–2111.

Hubmer, J., P. Krusell, and A. Smith (2019): “Sources of U.S. Wealth Inequality: Past, Present,

and Future,”mimeo extending NBER Working Paper 23011.

Jones, C. (2014): “Simple Models of Pareto Income and Wealth Inequality,” mimeo Stanford,

pp. 1–13.

(2015): “Pareto and Piketty: The Macroeconomics of Top Income and Wealth Inequal-

ity,”Journal of Economic Perspectives, 29(1), 29–46.

Khieu, H., and K. Wälde (2018): “Capital Income Risk and the Dynamics of the Wealth

Distribution,”IZA Discussion Paper No. 11840.

Krusell, P., and A. A. Smith, Jr. (2015): “Is Piketty‘s Second Law of Capitalism Fundamental?,”

Journal of Political Economy, 123(4), 725–748.

Mankiw, N. (2015): “Yes, r > g. So What?,”American Economic Review: Papers & Proceed-

ings, 105(5), 43–47.

66
Piketty, T. (2014): Capital in the 21st Century. Harvard University Press.

Piketty, T. (2015a): “About Capital in the Twenty-First Century,”American Economic Review:

Papers & Proceedings, 105(5), 48–53.

Piketty, T. (2015b): “Putting Distribution Back at the Center of Economics: Re‡ections on

Capital in the Twenty-First Century,”Journal of Economic Perspectives, 29(1), 67–88.

Piketty, T., and G. Zucman (2015): Wealth and Inheritance in the Long Run,pp. 1304–1366.

Handbook of Income Distribution, Atkinson, A. and Bourgignon, F. (eds.), Volume 2B, 2015

Elsevier B.V.

Romer, P. M. (1986): “Increasing Returns and Long-Run Growth,” Journal of Political Eco-

nomy, 94, 1002–1037.

Romer, P. M. (1994): “The Origins of Endogenous Growth,”Journal of Economic Perspectives,

8(1), 3–22.

Saez, E., and G. Zucman (2016): “Wealth Inequality in the United States Since 1913: Evidence

from Capitalized Income Tax Data,”Quaterly Journal of Economics, 131(2), 519–578.

Wackerly, D., W. Mendenhall, and R. Schea¤er (2008): Mathematical Statistics with Applica-

tions, 7th ed. Thomson Brooks/Cole.

Yaari, M. (1965): “Uncertain Lifetime, Life Insurance, and the Theory of the Consumer,”

Review of Economic Studies, 32(2), 137–150.

67