Chapter 4

Quantitative Spatial Economics

In this chapter we introduce a new breed of urban economics models, labelled by

Redding and Rossi-Hansberg (2017) as quantitative spatial economics models. The
models we have analyzed so far are not very amenable for estimation, starting from
their assumptions and their mathematical complexity. Moreover, at least in our study
of the empirics of agglomeration, there seems to be a disconnect between the models
and the data.
Redding and Rossi-Hansberg (2017) highlight several advantages of this new
breed of models, which were developed in international trade and then adapted to
spatial economics:
• They easily accommodate many regions and a rich mobility structure
• They can rationalize data as equilibria of the model
• They are usually exactly identified
• They can be used to carry out counterfactual decomposition and welfare analyses
We will start by tackling the question of the extent of agglomeration forces once
again. Ahlfeldt et al. (2015) approximate the ideal experiment of density, and use the
variation from the experiment to estimate the strength and extent of agglomeration
within an urban model. To do this, they use a singular accident of nature: The rise
and fall of the Berlin Wall.
To call this paper anything less than a classic would be an understatement. Re-
cently, in 2018, this paper won the Frisch Medal for the best paper published in
Econometrica in the last five years. With this rather grandious introduction, let us
dig in the basics of this paper.
Ahlfeldt et al. (2015) attack the same question we tackled in the last chapter. How
strong are agglomeration and dispersion forces in a city? We already know that this
is a difficult question, and they point out two reasons why this is the case. First is
the endogeneity: it is hard to disentangle the strength of agglomeration forces from
simple differences in location fundamentals. For this you would need exogenous
variation in agglomeration. Second, it is hard to bring the traditional urban models
to data on cities.

41
42 4 Quantitative Spatial Economics

Ahlfeldt et al. (2015) find a fantastic source of exogenous variation to identify the
effects of agglomeration: the Berlin Wall. A little historical background is in order.
In July of 1945, after the Second World War, Berlin was split into three sectors, for
the Americans, the British and the Soviet. A remainder French sector would lately
be created from the British sector. These sectors were intended to be approximately
equal in population. Although the city was intended to be governed jointly, rela-
tionships between the Western and Eastern factions deteriorated and in 1961 East
Germany built the Berlin Wall.
Figure 4.1 shows a map of Berlin in 1936 along with the boundary of the Berlin
Wall. The wall cut off West Berlin from the CBD, cutting through the subway and
rail lines.

Fig. 4.1 Ahlfeldt et al. (2015) Figure 1

Towards 1989, with the beginning of the fall of the Soviet Union, the Berlin Wall
fell on November 9th, 1989, and Germany was reunified in October of 1990.
Ahlfeldt et al. (2015) have data on area, employment, land prices and distances
for city blocks in Berlin, for 1936 (pre Berlin-Wall), 1986 (Berlin Wall) and 2006
(Post Berlin Wall). Once the wall goes down, you can think that the potential func-
tion jumps in the center of the city.
So how does equilibrium in this city change when the Berlin Wall rises and falls?
The issue is that while we can know a lot about equilibrium, we can not know it all
from land prices and employment data, because we lack commuting data and wages.
Consider the simple example in table 4.
Here we have data on residence and workplace. We do not observe commuting,
or amenities. The issue here is that there are many commuting patterns consistent
4.1 Model 43

Block 1 Block 2
Workers 1 2
Residents 2 1
Workplace amenities B1 B2
Residential amenities A1 A2

with equilibrium, each of which is going to be consistent with a certain pattern of

amenities. For example, an equilibrium could feature commuting only from block 1
to block 2, wich would be consistent with high workplace amenities in 2 and high
residential amenities in 1. Another equilibrium could have commuting going in both
directions. We will model this in a particular way to narrow down these possibilities.

4.1 Model

This section follows Ahlfeldt et al. (2015) closely. If you want a simplified version
along with an economic geography model and an overview of the versatility of these
models, see Redding and Rossi-Hansberg (2017).

4.1.1 Basic assumptions

We consider an open city. In the agricultural area utility is Ū. The city is in discrete
space, and there are S blocks indexed by i = 1, . . . , S. There is Li floor space in every
block, and these blocks can be assigned to residential or commercial usage, much
as in Fujita and Ogawa (1982). θi is the endogenous fraction of every block that
is dedicated to commercial use. The city produces a single numéraire good that is
costlessly traded. The blocks are connected, and there are H (endogenous) workers
that can freely move within the city.
Workers get utility from living in a particular block i and working in a particular
block j. The utility worker o gets from living in i and working in j is:
 β  1−β
Bi zi jo ci jo `i jo
Ci jo = (4.1)
di j β 1−β
.
Here, Bi are residential amenities, di j is the commuting cost (in terms of utility)
from i to j, c is consumption, ` is land and β governs the housing share of expendi-
ture. zi jo is a preference shock, that captures the fact that some workers may prefer
to live or work in some locations. This is where all the magic happens, so we will
explain this in detail.
44 4 Quantitative Spatial Economics

4.1.2 Fréchet “magic”

We are going to assume a particular extreme-value distribution for this preference

shock parameter. This distribution is going to be a Fréchet distribution:
−ε
F(zi jo ) = e−Ti E j zi jo (4.2)
Ti is the average utility from living in block i, E j is the average utility from work-
ing in block j, and ε is a parameter that measures the dispersion of the distribution.
Eaton and Kortum (2002) show that if worker productivity is the product of in-
ventions that happen over time drawn from a Pareto distribution, then the worker
productivity for the most productive technology has a Fréchet distribution. They use
this to model worker productivities in a Ricardian model of trade. We are going to
use it in a Ricardian model of commuting and migration.
Mathematically, there are two reasons why the Fréchet distribution is useful in
this setting. First, the Fréchet is constant under linear transformations. That is, if
z is Fréchet, e.g F(z) = exp(−T z−θ ), then y = kz + b is Fréchet, with F(y) =
exp(−T kε (Y − b)−ε ).
Second, the maximum of Fréchets is Fréchet distributed, and there are simple
expressions for the fractions when a Fréchet variable is larger than other. Consider
two Fréchet variables z1 and z2 , with shape parameters T1 and T2 and the same
dispersion parameter ε. Then

Z ∞
Pr(z1 < z2 ) = F(z1 )dF(z2 )
Z0 ∞
−ε −ε
= e−T1 z e−T2 z T2 εz−ε−1 dz
0
T2
Z ∞
−ε
= e−(T1 +T2 )z ε(T1 + T2 )dz
T1 + T2 0
T1
= .
T1 + T2

4.1.3 Household choice

We now turn to solving the household’s problem. We adopt the same strategy used
in the previous chapters: first, we solve for the optimal consumption and land use
given a location, then we examine location choice. If wages in location i are wi and
rents are Qi , then the indirect utility of living in i and working in j from the first part
of the optimal choice is:
β −1
zi jo Bi w j Qi
ui jo = . (4.3)
di j
4.1 Model 45

We model the commuting costs as a function of travel times:

di j = eκτi j (4.4)
Now we’ll use our two Fréchet distribution facts. Since zi jo is Fréchet, then ui jo
is Fréchet, with shape parameter

1−β −ε
 
Φi j ≡ Ti E j di j Qi (Bi w j )ε . (4.5)

The individuals will choose to live in i and work in j with some probability, which
is the probability of ui jo = maxr,s urso . From our second Fréchet useful property, this
probability is
Φi j Φi j
πi j = ≡ . (4.6)
∑r=1 ∑Ss=1 Φi j
S Φ
This completely characterizes the solution of the household’s problem and the
spatial distribution of households. We can now find the fractions of people who
reside in i, πRi and the fractions of people who work in j, πM j :

πRi = ∑ πi j ; πM j = ∑ πi j . (4.7)
j i

Notice that the probability of living in a place increases with residential amenities
and decreases with rent. The probability of working in a place increases with wages.
ε is important in determining these elasticities. Higher commuting costs reduce the
probability of choosing a particular pair.
The only remaining condition is that everything adds up. The probability of work-
ing on a particular place j conditional on living in i is

πi j E j (w j /d j )ε
πi j|i = = . (4.8)
∑s πis ∑s Es (ws /ds )ε
Adding up then requires that the workers in each place j, HM j , equal the sum of
the residents in each place, HRi , times the probability that they work in j:

HM j = ∑ πi j|i HRi (4.9)

i

4.1.4 Firm Choice

Firms are assumed to have CRS Cobb-Douglas production functions over workers
HM j and commercially-used land LM j , for which they pay wages w j and rents q j :
α L1−α
y j = A j HM j Mj
From the first-order conditions of the firm’s problem:
46 4 Quantitative Spatial Economics
  1
αA j 1−α
HM j = (4.10)
wj
  α
α 1−α 1
q j = (1 − α) A j1−α (4.11)
wj

The first condition determines equilibrium employment in each location, and the
second condition determines equilibrium commercial land prices.

4.1.5 Land Markets

Here we simply assume that floor space is supplied proportionally to land K at each
location
1−µ
Li = ϕi Ki (4.12)
Condition (4.11) determines commercial land demand. For residential land de-
mand, use the fact that households spend a constant fraction of their income on
commercial land, to get the expected demand of land from each individual li :

E[ws |i]HRi
E[li ]HRi ] = (1 − β ) (4.13)
Qi
Together with the supply function 4.12 these determine rents in equilibrium qi
and Qi . If a block has both residents and firms, Qi = qi . In all other cases the block
is either completely populated by residents or firms.

4.2 Equilibrium and agglomeration

An equilibrium here requires:

• Firms maximize profits and choose optimal locations
• Households maximize utility and choose optimal locations
• Land markets clear
• Labor markets clear
The endogenous variables are population H, the fractions living and working in
each place πR and πM , land rents Q, q, wages w and land use θ . Ahlfeldt et al.
show that there is a unique equilibrium in this model. This contrast with the non-
uniqueness of equilibria in Fujita and Ogawa (1982), but we have not introduced
agglomeration forces yet.
To model agglomeration, Ahlfeldt et al. (2015) make residential and workplace
amenities depend on the density of employment and residency:
4.3 Rationalizing data 47

HMs λ
 

−δ τ js
Aj = aj ∑e (4.14)
s Ks
HRr η
  
Bi = bi ∑ e−ρτir
r Kr

Here, τi j is the travel time from i to j. λ and η govern the strength of agglomera-
tion forces (β in Fujita and Ogawa (1982)) and δ and ρ govern their spatial decay (α
in Fujita and Ogawa (1982)). In presence of agglomeration forces, the equilibrium
is, as expected, not unique.

4.3 Rationalizing data

Now we are in good shape to take our model to data and calculate equilibriums.
There is an issue, however. In order to determine the endogenous variables, we
need values of the residential amenities, workplace amenities and the density of
development. These are unfortunately, unobserved. However, we can assume the
data is an equilibrium of the model, and find values of these unobserved variables
such that the data is in fact an equilibrium. This is only possible if there is an unique
mapping from the data to the unobserved variables. Turns out that this is the case.
Given values of the data

• Rent Q = max(Q, q)
• Residents and workers HR , HM
• Land K
• Travel times τ

and values of the parameters: the labor share α, the housing share of expenditure
β , the elasticity of land development µ, the dispersion parameter ε, the commut-
ing cost parameter κ, and the agglomeration parameters λ , δ , η, ρ, there are unique
(normalized) Ai , Bi , ϕi that are consistent with the data being an equilibrium of the
model. The normalization is necessary because some of these unobserved parame-
ters appear isomorphically in the model.
To fix ideas, we will look at an example (from Matthew Turner) of how all this
works. Consider the simplified model in table 4.1. Forget about commercial land
rents, and about floor space development. Assume that all places have the same area
L, and that d12 = d21 and dii = 1, i = 1, 2. Commuting is 1/2 from block 2 to block
1, and we assume there is no cross-commuting from block 1 to block 2.
We observe rents, land areas, and the distribution of workers and residents. Our
goal is to get the unobserved amenities and productivities. Notice that we don’t
observe wages either.
Assume a simpler structure for utility:
w j zi jo
ui jo = bi + − Qi , (4.15)
di j
48 4 Quantitative Spatial Economics

Variable Block 1 Block 2

HR 1 2
HM 3/2 3/2
Workplace productivity a1 a2
Residential amenities b1 b2
Rents (observed) Q1 Q2
Table 4.1 Simplified model

where z is Fréchet (T, ε). Also assume a Cobb-Douglas structure for production:
1−α
Yi = ai Lα HMi . (4.16)
From the structure of commuting we can recover the conditional probabilities of
working in each place.
• π11|1 = 1 since every resident from 1 works in 1
• π12|1 = 0 since we assumed no commuting from 1 to 2
1/2
• π21|2 = 1/4 = 2 which is the fraction of block 2 residents that work in block 1
• π22|2 = 3/4
Now to recover wages and productivities. Assuming production is Cobb-Douglas,
wages must satisfy
1−α
Yi ai Lα HMi
wi = (1 − α) = (1 − α) . (4.17)
HMi HMi
The only unobserved component here is ai . From the properties of the Fréchet
distribution, we know that the fractions must satisfy:

T (w1 /d11 )ε
π11|1 =
T (w1 /d11 )ε + T (w2 /d12 )ε
T (w2 /d12 )ε
π12|1 =
T (w1 /d11 )ε + T (w2 /d12 )ε
T (w1 /d21 )ε
π21|2 =
T (w1 /d21 )ε + T (w2 /d22 )ε
T (w2 /d22 )ε
π22|2 =
T (w1 /d21 )ε + T (w2 /d22 )ε

Provided this system of 4 equations has the proper rank, you can solve here for
a1 and a2 . You can be even greedier and try to get α and ε from here, but this is not
necessary here as we assumed those were observed. With ai you also have wages
wi . h i
wz
With wages in hand, we can solve for expected income vi = E di iijjo .
4.4 Reduced-form evidence 49

With expected income in hand, spatial equilibrium implies that utility must be
equalized across locations.

b1 + v1 − Q1 = b2 + v2 − Q2 (4.18)
If we normalize b1 = 1, we can get b2 from here, and we are done!

4.4 Reduced-form evidence

Before digging into structural estimation of the model, Ahlfeldt et al. (2015) provide
some reduced-form evidence of the effects of division and reunification of Berlin.
Tables 4.2 provides difference in difference estimates of the effect of division and
reunification on land prices and employment in West Berlin, by distance to the CBD.
The estimating equation is
K
∆ ln(Oi ) = α + ∑ 1k βk + lnMi γ + ui (4.19)
k=1

Which is the first-difference version of differences-in-difference. k indexes grids

of distance to the CBD. Note that any block specific controls in this specification
translate to trends that vary by the level of this control in the levels specification.

Table 4.2 Ahlfeldt et al. (2015) CITE Table 1

The coefficient in the first row of column (1) implies that floor space prices fell by
around 55% in the blocks closest to the CBD. Note that the effects become smaller
50 4 Quantitative Spatial Economics

as the distance to the CBD increases, and that they seem robust across specifications.
Columns (6) to (9) show decreases in employment. In the paper, they also report
results for reunification, which confirm the results here, having the opposite signs
and slightly smaller magnitudes.

4.5 Structural Estimation

The last section of Ahlfeldt et al. (2015) estimates the parameters λ , δ , η, ρ that
govern the strength of agglomeration forces. They use the exogenous variation aris-
ing from the change in density from the rise and fall of the berlin wall. The basic
idea behind their estimation is the following: any change over time in the location
fundamentals a j and bi from (4.14) should not be correlated with the changes in
density. Changes in density are measured by the distance to the CBD. I will ignore
normalization of these fundamentals and write down the identifying conditions as:

E [1k × ∆ ln(ait )] = 0, k = 1, . . . , K (4.20)

E [1k × ∆ ln(bit )] = 0 k = 1, . . . , K
(4.21)

How to estimate the model now? We use this identifying condition in an analo-
gous way to the identifying conditions in OLS (X 0 e = 0) or IV (Z 0 e = 0). The issue
is that while the error term is easy to obtain in a linear model, it is not easy to obtain
them here. But it was easy to obtain them in our simpler model in section 4.3.
The estimation proceeds in two loops, it what is called a “nested-fixed point
algorithm”. For details on this, see Holmes and Sieg (2015)
1. Fix a starting value of the parameters λ , δ , η, ρ.
2. For these values, find the values of ai , bi that rationalize the data as being an
equilibrium of the model. This is the “inner loop”.
3. With estimates of the location fundamentals âi , b̂i in hand, use the identification
condition in (4.20) to obtain estimates λ̂ , δ̂ , η̂, ρ̂
4. Find new values of the location fundamentals for the new values of the parame-
ters, â0i , b̂0i . This is the “outer loop”
5. Iterate until convergence in λ̂ , δ̂ , η̂, ρ̂
Severen (2018) provides of identification strategies for these quantitative spatial
economics models. To do the outer loop steps, we will use the generalized method
of moments (GMM) to find the estimates.
4.5 Structural Estimation 51

4.5.1 GMM

This section follows Hayashi (2000). Recall OLS estimation with L variables. The
identifying assumption to estimate βL×1 in

Yn×1 = Xn×L βL×1 + en×1

is
X0 e = 0.
Define the moment condition as

1 n
m(β ) = ∑ xi · ei
n i=1
(4.22)

1 n
= ∑ xi · (yi − xi β )
n i=1

This is the sample analog of the identifying condition. We would like to set this
to 0. This is a system of L equations with L unknowns. If you solve this system, we
will get the usual OLS solution for β .
The same logic can be applied to linear instrumental variables estimation. In
that case, the identifying condition is the exclusion restriction on the instruments Z.
Suppose that we have the same number of instruments as endogenous variables, so
Z is Zn×L . The identifying condition is Z0 e = 0. The moment condition now is:

1 n
m(β ) = ∑ zi · ei
n i=1
(4.23)

1 n
= ∑ zi · (yi − xi β )
n i=1

And by setting this equal to 0 and solving, we will get the usual IV estimator.
Now consider a setting where you have more instruments that endogenous vari-
ables, e.g. Z is n × M, X is n × L and M > L. In that case, we can no longer set m(β )
to 0. But we can set it as close to 0 as possible.
Consider minimizing the following function with respect to β .

J(β ) = N[m(β ) − 0]W[m(β ) − 0]0 (4.24)

What this is doing is choosing the value of β for which the identifying condition
is as close to being satisfied as possible in the sample. We minimize a quadratic
form in m(β ), weighting by W.
If you recall the previous chapter, this looks awfully familiar to the Ciccone and
Hall setting. Their estimator minimized
52 4 Quantitative Spatial Economics

J(β ) = Z 0 [y − f (β )]W [y − f (β )]0 Z (4.25)

which as an analogous problem to the previous one. In this case, the identifying
condition was Z 0 [y − f (β )] = 0. It does not matter that f (β ) is not linear, as we can
still minimize J(β ) using numerical methods.
Back to Ahlfeldt et al. (2015), they implement this estimator using the moment
conditions from (4.20), plus additional moment conditions based on the variance of
travel times. In general, you need an instrument for every parameter you want to
identify.

4.6 Structural Estimation Results

Table 4.3 shows results of the GMM estimation. The parameters here can be com-
pared to the density elasticity parameters we have seen in previous studies. The
parameter λ is the elasticity of productivity to density. It is around 7 %, close to the
preferred parameter of 4 % of Ahlfeldt and Pietrostefani (2019). On the other hand,
residential externalities seem to be larger: doubling residential density implies an
increase in residential amenities of about 15 % according to the estimate of η. The
parameters δ and ρ measure how quickly the agglomeration externalities dissipate
in space.

Table 4.3 Ahlfeldt et al. (2015) Table 5

4.7 Summary 53

To better understand the spatial decay of agglomeration forces, table 4.4 shows
the percentage of the production and residential externalities agents get as they move
away from density. These results show that agglomeration externalities are very lo-
calized, confirming the results in Arzaghi and Henderson (CITE). After 20 minutes
of travel time, production externalities dissipate. Residential externalities dissipate
even faster.

Table 4.4 Ahlfeldt et al. (2015) Table 6

4.7 Summary

We have introduced state-of-the-art quantitative methods of spatial economics. The

combination of transparent reduced-form evidence –from an exogenous source of
variation– and structural estimation makes this paper unique. Along the way, we
learned about modelling strategies and the generalized method of moments.
54 4 Quantitative Spatial Economics

The results of Ahlfeldt et al. (2015) are one of the most transparent estimates of
the effects of density in the literature. There are many other applications of quantita-
tive spatial models. The discipline is now extending them to analyze many different
problems. Here are some examples:

• Monte et al. (2018) analyze how employment responds to local shocks in pro-
ductivity, accounting for the spatial connectedness across areas through migra-
tion and commuting. They conclude that areas that are more open to commuting
receive larger increases in employment.
• Severen (2018) examines the effects on commuting of the LA metro rail. Along
the way, he surveys estimation methods in quantitative spatial economics.
• Tsivanidis (2018) estimates the welfare and redistributive effects of building a
BRT in Colombia.
• Pérez Pérez (2017) extends quantitative spatial models to account for unemploy-
ment and uses them to analyze the effects of local minimum wages in spatial
equilibrium.
• Pérez-Cervantes (2016) uses these models to show how consumers use commut-
ing to insure against local shocks in Mexico.
In the next section we will move from inside the city to outside the city, and
consider the allocation of economic activity across regions.