Inequality and Growth

Esther Duflo
March 2003

CROSS SECTIONAL EVIDENCE

Can aggregate data say something informative
about the causal relationship between inequality and growth?
Wouldnt that be nice?
Earlier literature: Alesina-Rodrik (94) , AlesinaPerotti (96), Perotti (96), Benabou (96)
- Cross section.

(yit+a yit)/a = yit + Xit + git + i + it

over long periods of data.
- Negative correlation between beginning of
period inequality and long-run growth.
Consensus: negative (causal) relationship between inequality and growth.
2

PANEL DATA EVIDENCE

Deininger and Squire panel data set on inequality (available on the world Bank Web Site).
- Forbes (98), Li-Zou (98). Observe that level
of inequality may well be correlated with countries specific propensity to grow (the fixed effect). Estimate the same relationship

(yit+a yit)/a = yit + Xit + git + i + it

But with shorter periods, and a=5.
Want to use fixed effects. However, in the
presence of a lagged endogenous regressor, fixed
effect is biased.
3

Use GMM estimator developed by Arellano and

Bond (91).
First difference the data:

(a + 1)
(yit+a yit)/a =
(yit yita)
a
+(Xit Xita) + (git gita) + it ita
use yita, Xita and gita (and previous lags)
as instruments for the differences.
Find that when using fixed effects or Arellano
and Bond, the relationship is positive and significant.

POTENTIAL PROBLEMS

This literature (the ancient and the recent)

has commonly assumed a linear relationship between growth and observed levels of
inequality. Barro (99) is a partial exception
(interact inequality with wealthw/o fixed
effect. Finds that the negative relationship
holds only in levels).

There are reasons to think that the true

relationship is not linear in levels
Theoretical reasons
Measurement issues

OUTLINE
This lecture :

1. Examine what the theory predicts about

(a) Causal relationship/Correlation between
growth and inequality
(b) The functional form of this relationship

2. Show empirically that the linearity assumption is violated

3. Examine the consequences of this violation

for the interpretation of the data

A INEQUALITY AND GROWTH :

WEALTH-BASED MODELS
(REMINDER AND SUMMARY).
Basic model:
Production function that is concave in physical
capital + imperfection in the capital markets
(e.g. people can only borrow their inherited
wealth) a mean preserving spread in income
today reduces income tomorrow.
Empirical implications:
In cross section a more equal economy grows
faster than an unequal one (causal effect)
As the economy becomes richer, inequality
decline, and growth rates declines. Reduction
in inequality are correlated with reduction in
growth.
7

 Need to control for convergence, but convergence is non-linear in past growth (it slows
down).
All decreases in inequality are endogenous
in this model, and have the causal effect of
reducing the growth rate: also associated with
decline in growth rate.

WEALTH-BASED MODELS: IMPLICATION

FOR THE ESTIMATION
Level specification:
(yit+a yit)/a = yit + Xit + git + i + it
An interaction term between inequality and
wealth should be here
Effect of inequality is non linear, or even non
monotonic
Control for distance from steady state is
critical, but difficult...
Correlation between i and (yit+a yit)/a
First difference specification:
(a + 1)
(yit yita)
a
+(Xit Xita) + (git gita) + it ita

(yit+a yit)/a =

If the data is generated by a convergence process (and not by shocks on inequality, is mechanically positive.

B INEQUALITY AND GROWTH :

POLITICAL ECONOMY MODELS
Starting with Alesina, Rodrik (1994): inequality increase redistribution and redistribution reduces growth. Other models in this tradition: Benhabib-Rustichini(98), Tornell-Velasco(92),
Perroti (96)
Reduced form specification present in the
literature:

(yit+a yit)/a = yit + Xit + h(gita) + i + it,

However, why should one estimate only the
reduced form? Benabou (96): models imply
a link between inequality and redistribution,
and between redistribution and growth. These
channels are not seen empirically...
8

B INEQUALITY AND GROWTH :

POLITICAL ECONOMY
MODELS-EXTENSION
Now, suppose that changes in inequality can
be costly in any direction in the short run.
Modifying inequality implies making a transfer to either the poor or the rich, and making
transfers is costly:
New institution Acemoglu and Robinson
(2000),
Costly turmoilAlesina Perotti (1996)
Change in regime (Rodrik (99)). See table
1 )

B INEQUALITY AND GROWTH :

POLITICAL ECONOMY MODELS-SIMPLE
MODEL
A simplified model with two groups (A and B).
A gets g% of the output.
In each period, there is a growth opportunity: potential growth is y.
In each period, one group chosen at random can hold up the rest of the economy: It
can demand a increase in its share before the
growth takes place.
If the other group agrees the growth is y,
otherwise there is no growth
10

 In the data, it would imply a non-linear relationship between changes in inequality and
subsequent growth:
(yit+a yit)
= yit +Xit+k(git gita)+i +it
a
Is this non-linear relationship present? If it
is, what does, what constraint does it impose
on the estimation of the relationship between
inequality and growth?

11

4-IMPLICATION FOR THE LITERATURE:

B- NON PARAMETRIC ESTIMATION OF
FIRST-DIFFERENCE EQUATION
This differenced form miss-specified in the presence of non linearities.
Partially linear model:
(a + 1)
(yit yita)
a
+(Xit Xita) + (git gita) + it ita

(yit+a yit)/a =

Kernel and series estimation of (.) Inverted

U-curve. Robust to using other specifications.
Non-linearity in existing models:
(yit+a yit) = yit+a yit
(a
+ 1)(yit yita) a(Xit Xita)
Estimate the relationship:
(yit+a yit)/a = (git gita) + it
12

4-IMPLICATION FOR THE LITERATURE:

C- CONSEQUENCES OF NON-LINEARITY
Random effects.
(yit+a yit)/a = yit + Xit + git + i + it
Equation miss-specified (non-linear term in past
changes in omitted).
Fixed effects and First differences.
(a + 1)
(yit+a yit)/a =
(yit yita)
a
+(Xit Xita) + (git gita) + it ita
Impose a linear structure on a non-linear equation: average the negative and positive parts
of the curve.

13

4-IMPLICATION FOR THE LITERATURE:

D- CONSEQUENCE OF NON LINEARITY
Arellano and Bond: Instrument difference
with lagged levels.
-Overall change in inequality are negatively related to past inequality.
-This is driven by the fact that when inequality
is high, it tends to decrease. When it is low it
does not tend to increase
-Result of the IV estimation: average effect on
part of the distribution that is affected by the
instrument...: negative effect of decreases in
inequality.

14

WHAT TO MAKE OF IT?

What we have seen since the beginning of
the semester suggests that the aggregate relationship at the heart of cross
country growth regression models has problems...
A specification that just adds inequality linearly to a cross country growth regression is
not seriously grounded in any model (all of
these papers start with a general discussion
or reason why inequality enter, and just stick
it in...). Theories underlying the empirical exercise do not predict simple linear causal relationship between level of inequality and growth
Cannot estimating aggregate production function on cross-country data and fix them by
adding inequality in there.
Go back to the micro foundations, estimate
micro parameters, calibrate growth models.
15

Wt +1
[h(k * ) + ( wt k * )r ]

45
Diagram 1

wt

k*
+1

w*

Wt +1

45
w* w3

Diagram 2

w* w2

w* w1

w*

w* + w1

w* + w2

k
+1

w**
w* + w3

wt

Figure 1: Relationship between income growth and lagged gini growth: partially linear model (Perotti variables)
-0.08

Residual income growth (y(t+a)-y(t))

-0.09
-0.1
-0.11
-0.12
-0.13
kernel regression
quadratic fitting
quartic fitting

-0.14
-0.15
-0.16
-0.17
-0.18

-0.06

-0.04

-0.02
0
0.02
Lagged gini growth (gini(t)-gini(t-a))

0.04

0.06

Figure 2: Relationship between income growth and lagged gini growth: partially linear model (Barro variables)
0.305

Residual income growth (y(t+a)-y(t))

0.3

0.295
kernel regression
quadratic fitting
quartic fitting

0.29

0.285

0.28

0.275

-0.06

-0.04

-0.02
0
0.02
Lagged gini growth (gini(t)-gini(t-a))

0.04

0.06

-3

2.5

x 10

Figure 3: Relationship between gini and square of gini changes

(gini(t)-gini(t-a))2)

1.5

0.5
0.3

0.35

0.4

0.45

0.5

0.55

0.5

0.55

Gini(t-a)

Figure 4: Reduced form, with Perotti variables

-0.01

Residual income growth (y(t+a)-y(t))

-0.012
-0.014
-0.016
-0.018
-0.02
-0.022
-0.024
-0.026
-0.028
-0.03
0.3

0.35

0.4

0.45
Gini(t-a)

Figure 5: Relationship between income growth and lagged Gini growth: partially linear model
0.03

Residual income growth (y(t+a)-y(t))

0.02
0.01
0
-0.01
kernel regression
quartic fitting

-0.02
-0.03
-0.04
-0.05

-0.06

-0.04

-0.02
0
0.02
Lagged Gini growth (Gini(t)-Gini(t-a))

0.04

0.06

Figure 6: Relationship between income growth and lagged Gini growth: using Arellano and Bond coefficients

Residual income growth (y(t+a)-y(t))*

0.02

-0.02

-0.04
kernel regression
quartic fitting
-0.06

-0.08

-0.1

-0.06

-0.04

-0.02
0
0.02
Lagged gini growth (Gini(t)-Gini(t-a))

0.04

0.06

Table 1
Relationship between growth and changes in Gini, linear specifications

Random
Effects
(1)
Gini(t)

0.021
(0.09)

Dependent variable: (y(t+a)-y(t))/a

Barro
Perotti
Specification
Specification
First
Fixed
Arellano
Random
First
Fixed
Difference
effect
& Bond
Effects Difference
effect
(2)
(3)
(4)
(5)
(6)
(7)
0.298
(0.18)

0.297
(0.16)

0.56
(0.039)

-0.03
(0.043)

0.158
(0.068)

0.155
(0.063)

Arellano
& Bond
(8)
0.27
(0.016)

N
128
128
128
128
98
98
98
98
Note:
Standard errors in parentheses; a is equal to 5 (Five-year periods)
Control variables:
Perotti specification: Log(GDP(t),PPP I (t), male education (t), female education (t)
Barro's specification: Log(GDP(t-1)), log(GDP(t-1)) squared, government consumption(t-1), secondary education(t),
higher education(t), fertility(t), (term of trade(t+1)-terms of trade(t)), rule of law, democ(t),
democt(t) squared, spanish or portuguese colony, other colony, investment share (t-1)

Table 2
Countries with large changes in gini coefficients
Decrease in gini coefficient larger than 3 percentage points
Country
Period Beginning of period
Change in gini
gini (in %)
(percentage points)
(1)
(2)
(3)
(4)
Bangladesh
65-70
Bulgaria
70-75
Brazil
75-80
Canada
85-90
Colombia
70-75
Spain
75-80
Finland
70-75
Finland
85-90
France
75-80
Hong Kong
85-90
Hungary
65-70
Indonesia
80-85
Ireland
75-80
Italy
75-80
Korea, Republic80-85
Sri Lanka
85-90
Sri Lanka
65-70
Mexico
75-80
Norway
75-80
Portugal
75-80
Sweden
70-75
Trinidad and To 75-80
Trinidad and To 80-85
Turkey
70-75
Venezuela
75-80
Deininger and Squire high quality sample

37.3
21.5
61.9
32.8
52.0
37.1
31.8
30.8
43.0
45.2
25.9
42.2
38.7
39.0
38.6
45.3
47.0
57.9
37.5
40.6
0.4
51.0
46.1
56.0
47.7

-3.1
-3.7
-4.2
-5.3
-6.0
-3.7
-4.8
-4.7
-8.1
-3.2
-3.0
-3.2
-3.0
-4.7
-4.1
-8.6
-9.3
-7.9
-6.3
-3.8
-6.1
-4.9
-4.4
-5.0
-8.2

Increase in gini coefficient larger than 3 percentage points

Country
Period Beginning of period Change in gini
gini (in %)
(percentage points)
(5)
(6)
(7)
(8)
Australia
Bulgaria
Brazil
Brazil
Chile
China
Colombia
Germany
Dominican Republic
Finland
United Kingdom
Hong Kong
Sri Lanka
Sri Lanka
Mexico
New Zealand
New Zealand
Sweden
Thailand
Venezuela
Venezuela

85-90
75-80
80-85
70-75
75-80
85-90
75-80
65-70
85-90
75-80
85-90
80-85
75-80
80-85
85-90
85-90
75-80
75-80
85-90
80-85
85-90

37.6
17.8
57.8
57.6
46.0
31.4
46.0
28.1
43.3
27.0
27.1
37.3
35.3
42.0
50.6
35.8
30.0
27.3
43.1
39.4
42.8

4.1
7.2
4.0
4.3
7.2
3.2
8.5
5.4
7.2
3.9
5.2
7.9
6.7
3.3
4.4
4.4
4.8
5.1
5.7
3.4
11.0

Table 3
Relationship between inequality and changes in inequality and growth

gini (t)
gini(t)-gini(t-a)
(gini(t)-gini(t-a))2
gini(t)-gini(t-a)* 1(gini(t)-gini(t-a))<0
gini(t)-gini(t-a)* 1(gini(t)-gini(t-a>)=0
F test for (gini(t)-gini(t-a))2,
(gini(t)-gini(t-a))3, (gini(t)-gini(t-a))4
(p value in parentheses)

Dependent variable:(y(t)-y(t-a))/a
Perotti Specification
Barro Specification
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
0.05 0.064 0.094
-0.042
-0.039
(0.10) (0.099) (0.11)
(0.045)
(0.043)
0.065
0.36
0.053 0.073
(0.16) (0.17)
(0.063) (0.066)
-5.09
-5.37
-2.47
-2.33
(2.95) (3.06)
(1.16) (1.17)
0.63
0.27
(0.30)
(0.10)
-0.59
-0.11
(0.33)
(0.13)
9.02
5.72
(0.029)
(0.12)

Number of observations
128
128
128
Note: Coefficient obtained using random effect specifications.
Standard errors in parentheses; a is equal to 5 (Five-year periods)
Control variables: see note to Table 2

128

98

98

98

98

Table 4
Estimation of the reduced form model
Dependent variable: (y(t+a)-y(t))/a
Perotti
Barro
(1)
g(t-a)

-0.047
(0.076)

(2)

(3)

(4)

(5)

0.77
-0.033
-0.043
-0.21
(0.66)
(0.082)
(0.039)
(0.21)
g(t-a)^2
-0.94
0.26
(0.81)
(0.27)
Control variables
X(t)
X(t)
X(t-a)
X(t)
X(t)
Note: Coefficient obtained using random effect specifications.
Standard errors in parentheses; a is equal to 5 (Five-year periods)
Control variables: X(t) stands for control variable not lagged.
X(t-a) stands for control variables lagged one period (5 years).
For a list of control variables see note to Table2.

Dependent variable: change in gini coefficient

Perotti
Barro
(6)
-0.10
(0.043)

X(t-a)

g(t)-g(t-a)
(7)

(g(t)-g(t-a))2
(8)

g(t)-g(t-a)
(9)

-0.087
(0.038)

0.0067
(0.0025)

-0.25
(0.066)

X(t-a)

X(t-a)

X(t-a)

(g(t)-g(t-a))2
(10)
0.0076
(0.0038)

X(t-a)

Table 5
Non-linearity of the relationship between change in gini and growth in models based on first differences

Control variables

(y(t+a)-y(t))/a
Perotti
Barro
(1)
(2)

Dependent variable
(1/a*[y(t+a)-y(gdp(t)]*)
Perotti
Barro
(3)
(4)

A. Linear assumption: OLS coefficient of (Gini(t)-Gini(t-a) )

Gini(t)-Gini(t-a)
0.298
0.158
0.36
(0.18)
(0.068)
(0.18)
B. Piecewise linear assumption: OLS coefficients of (Gini(t)-Gini(t-a) )
if Gini(t)-Gini(t-a) <=0
0.79
0.39
0.69
(0.30)
(0.13)
(0.38)
if (Gini(t)-Gini(t-a) )>=0
-0.3
-0.13
-0.49
(0.35)
(0.11)
(0.38)
C. Quartic specification
F-test for non-linear
2.21
3.37
2.55
terms jointly significant
(0.09)
(0.02)
(0.059)
D. Quadratic specification
Gini(t)-Gini(t-a)
(Gini(t)-Gini(t-a) )2

0.17
(0.07)
0.4
(0.13)
-0.11
(0.14)
3.3
(0.02)

0.23
(0.18)

0.13
(0.067)

0.311
(0.19)

0.15
(0.66)

-5.88
(3.39)
128

-3.24
(1.26)
98

-5.94
(3.43)
128

-3.28
(1.23)
98

Number of observations
Note:
Standard errors in parentheses; a is equal to 5 (Five-year periods)
For a list of control variables see note to Table2.
For a definition of residual growth, see the text

Inequality and
Growth: Empirics

A simulation-based approach to inequality

and growth
Based on Townsend-Ueda (2004)
They set up a model where there are
forward looking risk-averse individuals are
making a decision about how to invest
their savings.
They can either invest it in an asset that
faces both idiosyncratic and aggregate risk
(their own business) or pay a fixed cost
and invest it in the "bank" where there is
no idiosyncratic risk.
Those who have enough to invest will
clearly choose the latter.
The parameter values are chosen so that
given the 1976 distribution of wealth in
Thailand and the actual growth path of
returns, the fraction of people in the
banking sector is 6%, which is what was.
However given the assumed parameter
values everyone does make into the stock
market in the long run.

 They then start with an economy with

1002 people drawn from the Thai wealth
distribution in 1976 and then draw both
idiosyncratic shocks and aggregate shocks
for every year from a fixed distribution.
They then let the model run forward,
generating the evolution of wealth and
financial deepening.
The key point is that in this model
inequality at any point of time is
co-determined with growth and financial
sector participation.
They simulate the model 1000 times,
which gives them panel data from 1000
identical countries that differ only in terms
of shocks.
They then run cross-country regressions
with this data: 20 year growth rate on 1985
and 1980 initial levels of inequality and
financial deepening.
1980 and 1985 give very different results.
Financial deepening comes in with the
wrong sign when inequality is included.

 Then they run the panel regressions with

country fixed effects: the sign of the
inequality effect depends on which initial
condition is used.
Nothing is significant.