Augmenting the Solow-Romer Model with

Spatial Externalities: An Application to the

Brazilian Case.

Waldery Rodrigues Júnior

Institute of Applied Economic Research,
Directorate of Urban and Regional Studies,
E-mail: waldery.rodrigues@ipea.gov.br

Pedro Henrique Melo Albuquerque*

Camilo Rey Laureto
Marina Garcia Pena
Rafael Dantas Guimarães
Gilberto Rezende de Almeida Júnior
Institute of Applied Economic Research,
Directorate of Urban and Regional Studies

Abstract: The main objective of this paper is to show how the results
obtained with traditional economic growth models will change with the
introduction of spatial elements. Specifically we show how the Arrow-
Romer externalities coupled with spatial externalities will change the
results obtained with the use of the traditional Solow-Swan Model for
economic growth. We use a spatial econometric specification based on
Ertur and Koch (2007) and Koch (2010) and applied the model to
the analysis of per capita income for the 558 micro regions in Brazil.
The results show that spatial externalities are important. Using various
ways for constructing the spatial weigh matrix we also identify what
are the most important municipalities in terms of the elasticity of per
capita income for Sao Paulo with respect to the rates of saving of its
neighbors. Explicit formulae for this Spatial version of the Solow model
are provided throughout the paper.

Keywords: Spatial Solow, Spatial Econometrics

Reference to this paper should be made as follows: Rodrigues Júnior

et. al. (2010) “Augmenting the Solow-Romer Model with Spatial
Externalities: An Application to the Brazilian Case.”, 32o Encontro
Brasileiro de Econometria.

JEL: C21, R15

*Professor at Departament of Management - University of Brasilia

Copyright c 2010. Sociedade Brasileira de Econometria
2 Rodrigues Júnior et. al.
1 Introduction

The main objective of this paper is to show how the results obtained with
traditional economic growth models will change with the introduction of spatial
elements. Specifically we show how the Arrow-Romer externalities couple with
spatial externalities will change the results obtained with the use of the traditional
Solow-Swan Model for economic growth.
The key element that drives the spatial influence is the technological
interdependence cross locations. Ertur and Koch (2007) and Koch (2010) have
developed a framework where the spatial externalities are explicitily introduced
and the technological interdependence is modeled in a study for 91 countries
(indexed by i = 1, . . . , N ). This framework serves as base for a test of the Spatial
Solow Model in Brazil.

2 Ertur and Koch (2007) and Koch (2010) Setup

The framework starts with an aggregate production function of the Cobb- Douglas
type for country i in time t and with constant returns to scale in labor and the
usual reproducible physical capital (k).

Yi (t) = Ai (t)Kiα (t)L1−α

i (t)
where the standard notation is used:Yi is the output, Ki is the level of physical
capital, Li represents labor, Ai is the aggregate level of technology for country i.
Moreover the tecnological process is written as below:

Ai (t) = Ω(t)kiφ (t) Aj γwij (t)

Q
(1)
j6=i

Note three important terms in the specification for the global technology Ai (t):

1. As it is used in standard Solow model there is an exogenous part for the

technological progress which is identical for all countries:

Ω(t) = Ω(0)eµt

where µ is the constant rate of growth for Ω and is identical for all countries.

2. The level of technology also depends on the capital-labor ratio (idiossincratic

for every single country). Note that Ai (t) increases with the level of ki and
ki = K i (t)
Li (t) . Here the parameter 0 ≤ φ < 1 is related to the local externalities.
Here is one channel through which capital accumulation influences growth.

3. Note that the Romer treatement of knowledge spillover effect is present

here. It is represented by the weighted average of knowledge present in the
neighbors (hence the representation Aj with j 6= i). The level of technological
interdepence that is brougth about by the spatial elements is represented by
 Augmenting the Solow-Romer Model with Spatial Externalities 3
the parameter 0 ≤ γ < 1. This is assumed to be identical for all countries.
However the net effect os spatial externalities is dependent on the how
connected one region is to its neighbors and this is represented by the
frictions elements wij .

3 Spatial Dependence and Economic Growth

In order to estimate the parameters of the model one needs to necessarily quantify
the neighbors influence (in a quantitative way). This is usuallly done through the
(1)
use of the contiguity matrix Wn×n also known as the neighborhood matrix ou
weight matrix.
(1)
Given a set of n polygons (P1 , . . . , Pn ) we construct the matrix Wn×n where
each element (friction) wij represents a measure of spatial proximity between
polygons Pi and Pj .
This measure can be evaluated via the following criteria:

• wij = 1 if the centroid of Pi is at a given (non zero) distance of polygon Pj ;

otherwise wij = 0.
• wij = 1 if Pi shares at least a common side with Pj ; otherwise wij = 0
(Rook).
• wij = 1 if Pi shares at least a common point with Pj ; otherwise wij = 0
(Queen).
l
• wij = liji where lij s the length of the common frontier between the polygons
Pi and Pj and lij is the perimeter of polygon Pi .

The diagonal matrix is equal to zero by definition. Since it is used in building

spatial indicators it is common to apply a normalization process in order for the
sum of the elements for any given row to be equal to 1.
We can generalize the definition of matrix W(1) for high degree orders, i.e., it
is possible to define if Pi and Pj are neighbors and Pi is a neighbor of Pk then Pj
and Pk are also neighbors (transitivity property).
t
Another idea related to continguity matrix is to take for each i 6= j, wij = tiji
where tij Pcan represent the number of links (phone calls, for example) for i to j
and ti = tij is the number of calls having i as the starting point.
j
The choice of matrix to be used in spatial analysis is not prone to a certain
degree of arbitrary selection. An analytical proposal is to choose the matrix of
neighborhood that maximizes a given measure, say the likelihood.
Cliff-Ord (1982) generalize the matrix W via a functional relation between the
relative size of the common frontier between the two polygons adjusted by the
bβ
inverse of the distance between the two observations, i.e., wij = dij
α where bij is
ij
the common frontier between regions i and j and α, β are adjusting parameters.
These weights may represent a measure of potential interaction between units i
and j. For example, the may have a direct relationship with the spatial interaction
(element) as wij = d1α or wij = e−βdij or else in a more complex way it may use
ij
4 Rodrigues Júnior et. al.
other measures of distance such as the Manhattan, Minkowsky or even stastistical
measure named Mahalanobis distance.
Tipcally the vectors of parameters(α, β) are given apriori (Example: α = 2
reflecting gravity) instead of the case where they would be given jointly with other
model parameters.
Clearly if these parameters are jointly estimated the objective function will
have a high degree of non-linearity, which led to a complex otimization problem.

Figure 1 Figure 1

Example: The figue below is represented by the following weight matrix of the
first order-Queen:
 
01000
1 0 1 0 0
(1)
 
W = 0 1 0 1 1

0 0 1 0 1
00110
Note that we still need to use the usual normalization (in order to obtain the
property that the sum of every row is one).

4 Spatial Solow Model

Given the type of technological interdependence implied by the model the

location (countries) cannot be analyzed in isolation. Instead they should take into
consideration the influence of neighbors and the role played by the matrix W. We
can rewrite the function for Ai (t) in matrix form:

A = Ω + φk + γWA
e e e
where A is a N × 1 vector of logarithms of the technology level in aggregate
terms, k is a N × 1 vector of logarithm of the physical capital per worker, W is
e
the N × N contiguity matrix . Solving for A led us with:
e

−1 −1
A = (I − γW) Ω + φ(I − γW) k (2)
e e
 Augmenting the Solow-Romer Model with Spatial Externalities 5
Rewriting the above equation and considering |γ| < 1 we obtain:
∞
(r)
n γ r wij
P
1 φ
Ai (t) = Ω 1−γ (t)kiφ (t)
Q
kj r=1 (3)
j6=i

Note that the level of technology depends on its own physical capital per worker
and the level of physical capital per worker of its neighbors.
Rewriting the above equations and using the equation for the production
function we obtain:
1 n
u
yi (t) = Ω 1−γ (t)kiuii (t)
Q
kj ij (t) (4)
j6=i

where the terms

∞
!
(r)
X
r
uii = α + φ 1 + γ wii
r=1

and
∞
(r)
X
uij = φ γ r wij
r=1

are the new elemente in the spatial model. Here the terms of friction wij are the
elements of row i and column j of matrix W when using power r. Finally,
Yi (t)
yi =
Li (t)
is the usual level of output per worker.
This model implies spatial heterogeneity in the parameters of the production
function. Note that for the special case where there is no capital externality
regarding the physical capital , i.e., φ = 0, we are left with uii = α and uij = 0.
This gives the standard production funciton of the Cobb-Douglas type.

5 Elasticities

We can evaluate the elasticity of income per worker in country i with respect to
the physical capital. When a country i increases its own physical capital per worker
N
P φ
it gets a social return of uii . This is augmented to uii + uij = α + 1−γ , if the
j6=i
countries simultaneously decide to increase their stocks of capital per worker.
In order to assure local convergence and avoid explosive or endogenous growth,
φ
we assume diminishing social returns: α + 1−γ <1
As in the Solow model, we assume that a constant fraction of the output, si ,
is saved and that work grows exogenously at a rate of ni for a country i. We also
assume an annual rate of physcal capital depreciation for all countries denoted
by δ. The evolution of the output per worker in country i is given by the Solow
fundamental dynamic equation:

ki 0 (t) = si yi (t) − (ni − δ)ki (t) (5)

6 Rodrigues Júnior et. al.
where the line over the variable denotes its derivative with respect to time.
Since the output per worker is characterized by diminishing returns, equation
5 implies that the physical capital-to-output ratio of country i, for i = 1, .h. . , iN is
0 ∗
ki ki
constant and converges to a balanced growth defined by ki = g or also yi =
si
ni +g+δ . In other words:

1 u
1
  1−u N ij
∗ 1−u
si
ki∗ = Ω (1−γ)(1−uii ) (t)
ii Q
ni +g+δ kj ii
(t) (6)
j6=i

Since production technology is characterized by externalities between countries,

we can observe that physical capital per worker in steady state depends not only
on usual technology and preference parameters, but also on the level of physical
capital per worker in neighbor countries.
In order to determine the equation that describes the real income per worker
in a country, we rewrite the production function in matrix form:

y = A + αk
e e e

Substituting A in equation 2, and multiplying both sides by (I − γW) we

e
obtain:

y = Ω + (α + φ)k − αγWk + γWy

e e e e

Rewriting the equation for the economy i and introducing the logarithm of the
output − capital ratio in the steady state, we obtain the real income per worker in
country i in steady state:

α+φ α+φ
log [yi∗ (t)] = 1−α−φ1
log [Ω(t)] + 1−α−φ log [si ] − 1−α−φ log(ni + g + δ)
N N N
αγ αγ γ(1−α) P
wij log yj∗ (t)
P P  
− 1−α−φ wij log [si ] + 1−α−φ wij log(nj + g + δ) + 1−α−φ
j6=i j6=i j6=i

This Solow model - spatially augmented - has the same qualitative

interpretations as in the traditional Solow model, especially concerning the
influence of the internal savings rate and the population growth rate on real income
per worker in steady state.
First, the real income per worker in steady state for a country depends
positively on its own savings rate and negatively on its own population growth
rate. Second, it can also be shown that the real income per worker in a country
depends positively on the savings rates of neighbor countries and negatively on
their population growth rates.
Actually, even though the sign of the coefficient for the neighbor countries’
savings rate is negative, each of these savings rates (log [sj ]) positively influences
its own real income per worker in steady state (log yj∗ (t) , which positively


influences the real income per worker in steady state for country i through
externalities and global spatial technological interdependence.
 Augmenting the Solow-Romer Model with Spatial Externalities 7
The net effect is certainly positive, as can also be shown by calculating the
income elasticity per worker in country i with respect to its own savings rate ξsi
and with respect to its neghbors’ savings rates ξsj .
We can also calculate the income elasticity per worker with respect to the
depreciation rate for the country i denoted by ξni , and for the neighbor countries
j, denoted by ξnj . This way, we obtain:
∞ h ir
α+φ φ γ(1−α)
ξsi = −ξni = r
P
1−α−φ + (1−α)(1−α−φ) wij 1−α−φ (7)
r=1

and
∞ h ir
φ γ(1−α)
ξsj = −ξnj = r
P
(1−α)(1−α−φ) wij 1−α−φ (8)
r=1

These elasticities help us better understand the effects of an increase of the

savings rate of country i or one of its neighbors j on the income per worker in
steady state.
First, we can see that an increase in the savings rate in country i leads to
a greater impact on the real income per worker in steady state than in the
traditional Solow model, due to the technological interdependence modelled as a
spatial multiplier effect representing knowledge diffusion.
Moreover, an increase in the savings rate of the neighbor country j positively
influences the real income per worker in steady state in country i.

6 Conditional convergence

As in the traditional Solow model, the Spatial Solow model predicts that the real
income per worker in a certain country converges to the value for the countries in
steady state.
Rewriting the fundamental equation of the Solow model (equation 6), including
the production function (4), we obtain
0
1 N
ki (t) −(1−uii ) Q u
ki (t) = si Ω 1−γ (t)ki (t) kj ij (t) − (ni + δ) (9)
j6=i

The main element responsible for the convergence result in this model is due
to the diminishing returns on capital. Actually,
 0 
k (t)
∂ kii (t)
<0
∂ki (t)
provided that uii < 1.
When a country increases its level of physical capital per worker, the growth
rate decreases and converges to its own steady state.
However, an increase in physical capital per worker in a neighbor country
j increases the productivity of industries in country i due to the technological
interdependence.
8 Rodrigues Júnior et. al.
We then have,
 0

ki (t)
∂ ki (t)
>0
∂ki (t)
provided that uii > 1.
Externalities on physical capital and technological interdependence only delay
the decrease in the marginal productivity of physical capital. Therefore, the
φ
convergence result is still valid under the hypothesis α + 1−γ < 1, in opposition to
the endogenous growth models, where the marginal productivity of physical capital
is constant.
Furthermore, this model allows for quantitative predictions about the
convergence speed to the steady state. As in the literature, the transition dynamics
can be quantified by means of a log-linearization of equation 9 around the steady
state, for i = 1, . . . , N :
dlog[ki (t)]
dt = g − (1 − uii )(ni + g + δ) [log (ki (t)) − log (ki∗ )]
N
uij (ni + g + δ) log (kj (t)) − log kj∗
P 

+
j6=i

We obtain a system of linear differential equations whose solution is quite

complicated. However, considering the following relations between the countries
relative to their steady states:

log (ki (t)) − log (ki∗ ) = Φ log (kj ) − log kj∗




log (yi (t)) − log (yi∗ ) = Θj [log (yi (t)) − log (yi∗ )]
With convergence speed equal to:

dlog [yi (t)]

= g − λi {log [yi (t)] − log [yi∗ ]}
dt
where:
N
uij Φ1j (nj g + δ)
P
N
j=1 X 1
λi = N
− (nj + g + δ)
Θj
uij Φ1j
P
j=1
j=1

These hypotheses state that the difference between the country i and its own
steady state is proportional to the correspondent differential for country j.
Therefore, if Θj = 1, countries i and j are at the same distance from their
steady states.
If Θj > 1 (alternatively Θj < 1) then country i is farther (or alternatively
closer) from country j’s steady state.
The relative difference between the countries with respect to their steady states
affects the convergence speed.
 Augmenting the Solow-Romer Model with Spatial Externalities 9
∂λi u (n +g+δ)
Actually, ∂Θ j
= ij Θ j
2 > 0 and the convergence speed is high, if the
j
country i is far from its own steady state.
Furthermore, the convergence speed is high if country j is close to its steady
state.
Therefore, there is a strong form of heterogeneity in this model, since the
convergence speed in country i is a function of the parameters wij representing the
distance to neighbor countries, from their own steady states.
When there are no externalities on physical capital (φ = 0), heteregeneity of the
convergence speed is reduced to the traditional Solow model: λi = −(1 − λ)(ni +
g + δ)
In this sense, we have the same relation between the externalities on physical
capital and the heterogeneity we obtained with the production function.
The solution for log (yi (t)), subtracted from log (yi (0)) - the real income per
worker in some initial moment - on both sides, we obtain:

µ 1
log (yi (t)) − log (yi (0)) = gt − (1 − e−λi t ) 1−γ λi −
−λi t −λi t ∗ (10)
−(1 − e )log (yi (0)) + (1 − e )log (yi )

The model predicts convergence, since the growth of real income per worker is a
negative function of the initial level of income per worker, but only after controlling
for the determinants of the steady state.
More specifically, the growth rate of real income per worker depends positively
on its own savings rate and negatively on its own population growth rate.
Furthermore, it also depends on the same variables on the neighbor countries,
due to technological interdependence. We can see that the growth rate is higher
when the initial level of income per worker is higher and when the growth rate of
neghbor countries is higher.
Finally, it can be seen that the growth rate of a country depends on the
growth rate of its neighbors, weighted by the speed of convergence and the spatial
neighboring conditions.

7 Analysis of Data for Brazil

Following the empirical literature on growth, we have used data from IPEADATA,
which contains information on real income, investments and population (among
other variables) for a great number of Brazilian cities.
Specifically for this work, we have used data from 558 micro-regions for Brazil.
We measure n as the average growth rate of the active population (aged 15 to
64). We also compute the number of workers.
Real income per worker is measured by dividing total income by the number
of workers. Also, the savings rate is denoted by s.
The matrix W defined before is the weight matrix commonly used in spatial
econometric models of spatial interdependence between countries or regions
(Anselin (1988); Anselin and Bera (1998);Anselin (2006)).
More precisely, each country is linked to a set of neighbor countries by means
of a purely spatial model introduced by W.
10 Rodrigues Júnior et. al.
The elements wii of the main diagonal are conventionally equal to zero, whereas
the elements wij indicate with which intensity country i is spatially connected to
country j.
In order to normalyze the external infulence on each country, the weights of
the matrix are standartized so that the elements of the rows sum to 1.
For variable x, this transformation means that the expression Wx, called
spatially lagged variable, is simply the weighted average of neighbor observations.
It is important to stress that the elements wij must be exogenous to the model.
In this sense, we consider the purely geographical concept of queen neighborhood,
according to which two countries are neighbors if their borderlines share at least
one point, which is indeed strictly exogenous.

8 Econometric Model Specification

In this section, we aim at assessing the impact of savings, population growth and
location on per capita real income.
Taking equation 10, we can see that the real income per worker along the
balanced growth path, at a certaint moment (t = 0, for the sake of simplicity), is:

h i N
Yi P
log Li = β0 + β1 logsi + β2 log(ni + g + δ) + θ1 wij logsj +
j6=i
N N h i (11)
P P Yj
θ2 wij log(nj + g + δ) + ρ wij log Lj + i
j6=i j6=i

1
Where 1−α−φ logΩ(0) = β0 + εi for i = 1, . . . , N with β0 a constant, εi a
stochastic term and the term Ω(0) reflects not only technology, but also resource,
climate and other endowments, and thus can differ from country to country.
We also assume that g + δ = 0.05, as is common in the literature of MRW
(1992) and Romer (1989). Finally, we have the following theoretical limitations for
the coefficients:
α+φ
β1 = −β2 =
1−α−φ
and
αγ
θ2 = −θ1 =
1−α−φ
γ(1 − α)
ρ=
1−α−φ
Also,

Y = Xβ + W Xθ + ρy + ε

y = ρWy + Xβ + ε
 Augmenting the Solow-Romer Model with Spatial Externalities 11
where,

ε ∼ N (0, σ 2 I)
e e

In this model, the vector y contains n observations of the dependent variable

of interest, X is the delimitation matrix that contains the covariates that explain
the event y and W is the matrix of known spatial weights.
Parameter ρ is responsible for measuring the spatial dependence of the
dapendent variable y and its respactive neighbors.
The vector of parameters β reflects the influence of the covariates on the
dependent variable (does not include the spatial component of neighboring).
The model is called mixedmodel because it combines the standard regression
with a spatial component which is the spatially lagged variable Y .
Thus, the model tries to explain the dependent variable of the region yi , with
the value of the dependent variable from its neighbors and with covariates for its
own region.
We can estimate these parameters through numeric optimization or, more
commonly, through concentrated likelihood. In both cases, first it is necessary to
build a likelihood function (joint distribution of vector y)
Note that:

y = ρWy + Xβ +  ⇒
y − ρWy = Xβ +  ⇒
(I − ρW)y = Xβ +  ⇒
(I − ρW)y − Xβ = 

Since the distribution ε is known, we can use the Jacobian transformation in

order to obtain the joint distribution of y. For that, we need the Jacobian:
 
 ∂ 
 
J =   = |I − ρW|
 ∂y 

so, the joint dostribution function of y is:

 
 ∂ 
f (y) = f ()  ∂y  ⇒
  
 ∂ 
f (y) = f (I − ρW)y − Xβ  ∂y  ⇒
(  T  )
−n 2 − 1 1 2 −1
f (y) = (2π) 2 |σ I| 2 exp − 2 (I − ρW)y − Xβ (σ I) (I − ρW)y − Xβ |I − ρW|

The equation above represents the joint distribution of y and also the likelihood
of the parameters θ = (ρ, β T )T .
Taking logarithms, we obtain the log-likelihood function:
12 Rodrigues Júnior et. al.
 T
l(ρ, β, σ 2 ) = − n2 log(2π) − 12 log|σ 2 I| − 12 (I − ρW)y − Xβ ×
 
2 −1
×(σ I) (I − ρW)y − Xβ + log|I − ρW|

To proceed with the concentrated likelihood, we need to define which

parameters are nuisance. In this case, the parameters of disturbance are β and
σ 2 . The next step is to obtain the maximum likelihood estimators of nuisance
parameters that should be a function of the parameter of interest ρ. For the
nuisance parameter β we have:
 
∂ 2 1 T
∂β l(ρ, β, σ ) = − 2σ 2 [−X] (I − ρW)y − Xβ = 0 ⇒
 
T
[−X] (I − ρW)y − Xβ = 0 ⇒
−XT (I − ρW)y + XT Xβ = 0 ⇒
XT Xβ = XT (I − ρW)y = 0 ⇒
−1
β = (XT X) XT (I − ρW)y ⇒
−1 −1
β = (XT X) XT y − ρ(XT X) XT Wy ⇒
β̂ = β̂ − ρβ̂
O L

Where β̂ is the estimator of the parameter vector β of the model y = Xβ +

O O
 by the method of ordinary least squares and β̂ is the OLS estimator of the
O L
model Wy = Xβ +  . Similarly, we estimate the parameter σ 2 :
L L

 T  
∂ 2 −n 1
∂σ 2 l(ρ, β, σ ) = − 2σ 2 + 2σ 4 (I − ρW)y − Xβ (I − ρW)y − Xβ = 0 ⇒
 T  
(I−ρW)y−Xβ (I−ρW)y−Xβ


σ2 = T  n ⇒
y−ρWy−Xβ y−ρWy−Xβ
σ2 = n ,comoβ̂ = β̂ − ρβ̂ temos:
 T  O L
y−ρWy−X(β̂ −ρβ̂ ) y−ρWy−X(β̂ −ρβ̂ )
σ2 =  n
T
O L

O L
⇒
y−Xβ̂ −ρWy−ρXβ̂ y−Xβ̂ −ρWy−ρXβ̂
σ2 = O
 n
L
T  
O L
⇒
ˆ −ρˆ
 ˆ −ρˆ

σ̂ 2 = O L
n
O L

Where ˆ is the residue of the estimated model y = Xβ +  and ˆ is the

O O O L
residue of the estimated model Wy = Xβ +  .
L L
 Augmenting the Solow-Romer Model with Spatial Externalities 13
So the concentrated likelihood is as follow:

l(ρ) = Constante− n2 log(σ̂ρ2 ) + log|I −ρW| ⇒

 T  
ˆ −ρˆ
 ˆ −ρˆ

l(ρ) = Constante − n2 log  O L
n
O L  + log|I − ρW|

Then, we maximize the concentrated likelihood function on ρ thus obtaining ρ̂,

then σ 2 and β are estimated respectively by:
1 T
σ̂ 2 = n (ˆ
 − ρˆ ) (ˆ − ρˆ ) e β̂ = (β̂ − ρ̂β̂ ).
O L O L O L
Consequently, the solution of the estimation problem can be schematically as
follows:

• Regress X in y by OLS (getting as β̂ ).

O

• Regress X in Wy by OLS (getting as β̂ ).

L

• Calculate ˆ and ˆ .
O L

• Given ˆ e ˆ find ρ that maximizes the concentrated likelihood function.

O L
 T  
ˆ −ρ̂ˆ
 ˆ −ρ̂ˆ

• Given ρ make σ̂ 2 = O L
n
O L
e β̂ = (β̂ − ρ̂β̂ ).
O L

The spatial lag model is used, in general, the following cases:

• Dependent variable is spatially correlated.

• Spatial factor directly influences the dependent variables.

• Spatial factor influences overall space.

Graphically, this model to represent the interaction between the response

variable, the covariates and error as follows:

9 Empirical Evidence

Our results for the qualitative predictions are essentially identical to those of MRW
(1992), since the coefficients of savings and population growth have the expected
signs and are significant. We have results from the Spatial Auto Regressive Model
(SAR), Spatial Error Model (SEM) and Ordinary Least Squares (OLS). The
models and data imputation (Missing data for each region was imputed with the
average from the neighbors.) were run on IpeaGEO.
14 Rodrigues Júnior et. al.
Figure 2 Figure 2

Table 1 Spatial Auto Regressive Model

Parameter Coefficient Standard t Stat P-value
Error
Intercept 5.446541 0.293504 18.556929 0.000000
LOG SI 0.238113 0.024377 9.768142 0.000000
LOG X 0.892737 0.145027 6.155638 0.000000
W LOG SI -0.022716 0.033231 -0.683595 0.494231
W LOG X 1.708643 0.181219 9.428597 0.000000

Table 2 Spatial Error Model

Parameter Coefficient Standard t Stat P-value
Error
Intercept 7.817188 0.314011 24.894615 0.000000
LOG SI 0.239212 0.022426 10.666562 0.000000
LOG X 1.167668 0.128494 9.087370 0.000000
W LOG SI -0.124858 0.061760 -2.021655 0.043212
W LOG X 0.736193 0.213608 3.446463 0.000568

Considering the results from the SAR, as for the spatial lag parameter, we
obtain ρ = 0.687192 with standard error equal to 0.023957 and p-value of 0.000000.
These results imply γ = 0.573379, φ = 0.160321 and α = 0.031999.
Considering the results from the SEM, as for the spatial error lag parameter, we
obtain λ = 0.877103 with standard error equal to 0.019836 and p-value of 0.000000.
These results imply γ = 0.808547, φ = 0.068422 and α = 0.124614.
The results from the OLS imply γ = 0.633265, φ = 0.137914 and α = 0.052577.
Then there is evidence that spatial dependence is significant, this way, the usual
Solow model cannot be used for that data set.
Our model predicts that the savings rate and population growth have larger
effects on real income per worker because of physical capital and technological
interdependence externalities.
 Augmenting the Solow-Romer Model with Spatial Externalities 15
Table 3 Ordinary Least Squares
Parameter Coefficient Standard t Stat P-value
Error
Intercept 5.130454 0.299209 17.146712 0.000000
LOG SI 0.235317 0.024417 9.637324 0.000000
LOG X 0.834924 0.145600 5.734382 0.000000
W LOG SI -0.041130 0.033438 -1.230036 0.218684
W LOG X 1.785757 0.181997 9.812022 0.000568
W LOG INCOME 0.741153 0.025822 28.702054 0.000000

To calculate these elasticities of real income per worker in steady state in

relation to the savings rate and the effective rate of depreciation, we can rewrite
the equations 7 and 8 in matrix form:

−1
Ξ = β1 I + (β1 ρ + θ1 )W(I − ρW)

Therefore, from estimates reported in Table I, we obtain a matrix Ξ 558 x 558

with direct elasticities on the main diagonal and cross-elasticity off-diagonal.
In column j, we have the effect of an increase in the savings rate sj of the
micro-regions of Brazil. Note that once the terms wij are considered the effect is
larger for the closest micro-region. In row i, we have the effect of an increase in
savings rate in each micro-region, to the micro-region i in real income per worker.
We also note that the sum of each row is identical for all micro-regions. This
property is due to the neighborhood matrix W, which means a similar increase in
the savings rate in all the micro-regions will have the same effect on real income
per worker in steady state.
On average, the elasticity of real income per worker in relation to the savings
rate is approximately 0.001236 for the SAR, 0.001670 for the SEM and 0.001347
for the OLS.
We also obtain the complete results of cross-elasticities, indicating the influence
of savings rates or rates of population growth in neighboring micro-regions on the
real income per worker.
Therefore, these values of elasticities can provide a much better explanation of
the differences between micro-regions in terms of real income per worker.
Indeed, physical capital externalities, technological interdependence and, more
generally spillover explain these income inequalities between micro-regions, since
they involve greater elasticity.

10 Conclusion

This paper presented an enrichment to models of economic growth from the

explicitness of spatial elements. The way that spatial externalities are introduced
in the empirical results for the analyzed case of micro-regions of Brazil, show the
importance of spatial elements. Additionally this work presents the results which
are unprecedented in terms of application of the spatial Solow model for Brazil.
16 Rodrigues Júnior et. al.
References

L. Anselin. Spatial Econometrics: Methods and Models, Studies in Operational

Regional Science, 1988.
L. Anselin. How (Not) to Lie with Spatial Statistics, American Journal of
Preventive Medicine, 2, 2006.

L. Anselin and A. Bera, Spatial dependence in linear regression models with an

introduction to spatial econometrics, New York: Marcel Dekker, 1998.
A. Cliff and J. Ord, Spatial Processes Models and Applications, London: Pion, 1981.
W. Koch and C. Ertur Growth, technological interdependence and spatial
externalities: theory and evidence., J Appl Econom, 2006.
Instituto de Pesquisa Econômica e Aplicada (IPEA), IpeaGEO 1.0.4,
http://www.ipea.gov.br/ipeageo/, 2010.
Mankiw, N. Gregory and Romer, David and Weil, David N., A Contribution to the
Empirics of Economic Growth, The Quarterly Journal of Economics, 2, 1992.
D. Romer Staggered Price Setting with Endogenous Frequency of Adjustment.,
Economics Working Papers, 1989.
 Augmenting the Solow-Romer Model with Spatial Externalities 17
11 Annex I: Map of Income-Savings Rate Elasticities

Here we present maps of the Income-Savings Rate Elasticities for the 9

Metropolitan Regions of Brazil for the SAR, SEM and OLS.

Figure 3 Micro region for Belém. Figure 4 Micro region for Fortaleza.

Figure 5 Micro region for Recife. Figure 6 Micro region for Salvador.
18 Rodrigues Júnior et. al.

Figure 7 Micro region for Belo Horizonte. Figure 8 Micro region for Rio de Janeiro.

Figure 9 Micro region for São Paulo. Figure 10 Micro region for Curitiba.
 Augmenting the Solow-Romer Model with Spatial Externalities 19

Figure 11 Micro region for Porto Alegre. Figure 12 Micro region for Belém.

Figure 13 Micro region for Porto Alegre.

20 Rodrigues Júnior et. al.

Figure 14 Micro region for Belém. Figure 15 Micro region for Fortaleza.

Figure 16 Micro region for Recife. Figure 17 Micro region for Salvador.
 Augmenting the Solow-Romer Model with Spatial Externalities 21

Figure 18 Micro region for Belo Figure 19 Micro region for Rio de
Horizonte. Janeiro.

Figure 20 Micro region for São Paulo. Figure 21 Micro region for Curitiba.
22 Rodrigues Júnior et. al.

Figure 22 Micro region for Porto Alegre.

Figure 23 Micro region for Belém. Figure 24 Micro region for Fortaleza.
 Augmenting the Solow-Romer Model with Spatial Externalities 23

Figure 25 Micro region for Recife. Figure 26 Micro region for Salvador.

Figure 27 Micro region for Belo Figure 28 Micro region for Rio de
Horizonte. Janeiro.
24 Rodrigues Júnior et. al.

Figure 29 Micro region for São Paulo. Figure 30 Micro region for Curitiba.

Figure 31 Micro region for Porto Alegre.