The Distributional Impacts of Transportation Networks in China

The Distributional Impacts of Transportation Networks
in China
Lin Ma and Yang Tang
Singapore Management University and Nanyang Technological University
2023-10-27, Fudan TED Conference
Ma and Tang (SMU and NTU) 05-09-23, HKU 1 / 28

Introduction
Outline
1 Introduction
2 Data
Dataset Construction
Quality Matters
3 Structural Exercise
Model
Quantification
Results

Introduction
Motivation
Infrastructure investment, especially transportation networks, is
widely seen as a driver of economic prosperity.
The case of China is particularly interesting as the country has rapidly
built up its transportation networks since the 1990s.

Introduction
Motivation
However, the distributional impacts are not clear, for two reasons:
1 From a theory point of view, better roads improve the mobility of both
goods and factor:
Better goods mobility improves the relative market access of the small
and remote cities;
Better factor mobility allows people to move out from the small and
remote cities at the same time.

Introduction
Motivation
However, the distributional impacts are not clear, for two reasons:
1 From a theory point of view, better roads improve the mobility of both
goods and factor:
Better goods mobility improves the relative market access of the small
and remote cities;
Better factor mobility allows people to move out from the small and
remote cities at the same time.
2 Empirically, to do a quantitative analysis, we also need panel data on
transportation networks; this is lacking in the context of China.

Introduction
This Project: Data
We construct a new panel dataset on the transportation networks in China
between 1994 and 2017:
1 Measuring the quality of the roads and railroads using design speed.
Across time: a “highway” built in 1994 only allows for a fraction of
travel speed as compared to a highway in 2017.
Across space: roads in the rugged terrains are built for half the design
speed as those in the eastern flood plains.

Introduction
This Project: Data
We construct a new panel dataset on the transportation networks in China
between 1994 and 2017:
1 Measuring the quality of the roads and railroads using design speed.
Across time: a “highway” built in 1994 only allows for a fraction of
travel speed as compared to a highway in 2017.
Across space: roads in the rugged terrains are built for half the design
speed as those in the eastern flood plains.
2 Multiple modes: road, railroad, and waterway.
3 Distinguish the types of traffic on the railroads: passenger, freight, or
mixed.

Introduction
Main Result: Quality Matters
1 Quality-adjustments matter for travel time estimation. Without
adjustment:
Over-estimation on the road network by 22%.
Under-estimation on the rail network by 37% to 59%.

Introduction
adjustment:
2 Quality-adjusted travel time is strictly better:
External validation using bus and train timetable data.
A horse race between the quality-adjusted and unadjusted travel time
estimation against the observed travel time from timetable.
Quality-adjusted travel time has much higher explanatory power than
the unadjusted ones.

Introduction
adjustment:
3 Through a dynamic structural model:
The median bias in the real wage growth rates is 19% to 33%.
At the right tail, ignoring quality measure leads to a bias of 500%.

Introduction
adjustment:
3 Through a dynamic structural model:
The median bias in the real wage growth rates is 19% to 33%.
At the right tail, ignoring quality measure leads to a bias of 500%.
4 Biases are not random! Correlated with initial conditions.
You cannot treat them as classical measurement errors and deal with
them using IV.

Introduction
Literature
Quantitative evaluation of infrastructure projects:
Allen and Arkolakis (2014, 2022), Redding and Turner (2015),
Donaldson and Hornbeck (2016), Alder (2016), Allen and Donaldson
(2018), Donaldson (2018).
In the Chinese context:
Bai and Qian (2010), Banerjee, Duflo, and Qian (2012), Faber (2014),
Qin(2016), Lin (2017), Xu (2017), Baum-Snow, Henderson, Turner,
Zhang and Brandt (2020), Ma and Tang (2020), Fan, Lu and Luo
(2021), Alder, Song and Zhu (2021), Egger, Loumeau, and Loumeau
(2023).
Infrastructure quality
Berg et al. (2018); Asher and Novosad (2020); Jedwab and Storeygard
(2021).

Data
Outline
1 Introduction
2 Data
Quality Matters
Model
Quantification
Results

Data Dataset Construction
Measuring the Quality of Infrastructure
Existing transportation data in China are binary and ignore quality
variations:
USGS 1992 map (USGS, CDC at University of Michigan);
State Bureau of Surveying and Mapping (China) 2008 map;
ACASIAN Data Center at Griffith University from 1992 to 2010;
Baum-Snow et al. (2020), Egger, Loumeau, and Loumeau (2023).

Measuring the Quality of Infrastructure
Existing transportation data in China are binary and ignore quality
variations:
USGS 1992 map (USGS, CDC at University of Michigan);
State Bureau of Surveying and Mapping (China) 2008 map;
ACASIAN Data Center at Griffith University from 1992 to 2010;
Baum-Snow et al. (2020), Egger, Loumeau, and Loumeau (2023).
The quality of roads and railroads differ by vintage, rate, and terrain,
as stipulated by the official design codes:
Older, lower-rated, or built over rugged terrains have limited speed and
capacity.
Nine revisions in road construction codes, recent ones in 1988, 1997,
2003, and 2014.
Recent revisions in railway design codes in 1985, 1999, 2006, 2012.
Conditional on revision and rate, quality vary by terrain.

Design Speed Depends on Vintage, Rate, and Terrain
Terrains More Railway Design Codes
Revision Plains LRH Hills Mountains Plains LRH Hills Mountains
Highways First-Rate Roads
1988 120 120 100 60 100 100 60 60
1997 120 120 120 60 100 100 60 60
2003 120 120 120 80 100 100 80 80
2014 120 120 120 80 100 100 80 80
(a) Road Standards
Revision Plains LRH Hills Mountains Plains LRH Hills Mountains
National I National II
1985 120 100 80 80 100 80 80 80
1999 140 120 80 80 100 80 80 80
2006 160 140 120 120 120 100 80 80
(b) Railroad Standards, Mixed-Use
Table 1: Design Speed (km/h) of Roads and Railroads by Vintage, Rate, and
Terrain

Identify Year of Construction
For each segment, identify the Year of Construction (YOC).
identify
segment,
1996
compare
to all
segments
in 1994
new
segment
YOC =
94, 95, 96?
collect
other info.
old
segment
information
already
collected
We have national maps in 94, 96, 2000, 2003, etc.
Compare to the previous national map and determine if it is a new
construction.
If new, then determine the YOC from the other sources.
For example, we only have national maps in 1994 and 1996. If a segment
first showed up in 1996, it could be constructed in 1994, 1995, or 1996.
References: Provincial map collections,Transportation Yearbooks,Railroad
Yearbooks,Chronicles of Railroad Construction.

Example of New Construction: Cheng-Yu Highway

Structure of the Dataset
120 120 120 120 100 100 100 100 120 120 80 80 140 140 140 140
1996 (rev.1985) 2001 (rev.1999) 2004 (rev.1999)
ABC Railway, Mixed-Use, National I
Pixel Level:
Segment Level:
Path Level:
The end result is a pixel-level panel data that identify the design speed at
each pixel in every year, for every mode and usage. Three layers:
The outer-most layer is a “path” that correspond to a named road or
railroad, e.g. Beijing-Shanghai Highway, Longhai Railway, etc.
We identify the name, rate, and usage of a path.
Rate and usage are used to determine the applicable design codes.

120 120 120 120 100 100 100 100 120 120 80 80 140 140 140 140
1996 (rev.1985) 2001 (rev.1999) 2004 (rev.1999)
Pixel Level:
Segment Level:
Path Level:
A “path” contains several “segments” built over multiple years.
Within a path, segments differ in the year of construction.
In this example, one segment was built in 1996 and was subject to the
1985 revision of the design codes.
The two segments built in 2001 and 2004 were subject to the newer
1999 revision.

120 120 120 120 100 100 100 100 120 120 80 80 140 140 140 140
1996 (rev.1985) 2001 (rev.1999) 2004 (rev.1999)
Pixel Level:
Segment Level:
Path Level:
A “segment” contains many “pixels” that differ in terrain.
= “plains”, = “low-rolling hills”, and = “hills”.
Using the information on usage, rate, year of construction, and
terrain, we determine the design speed of each pixel, shown in the box
in the unit of “km/h”.

120 120 120 120 100 100 100 100 120 120 80 80 140 140 140 140
1996 (rev.1985) 2001 (rev.1999) 2004 (rev.1999)
Pixel Level:
Segment Level:
Path Level:
Within a segment, design speed differs due to terrains.
Within a terrain type, design speed differs due to the changes in
applicable technical standards.
From the design speed, we compute the distance between prefectures
i, j in the unit of travel time.
We do this for freight (Tmf
ijt ) and passengers (Tmp
ijt ) separately for
mode m at year t.

Data Quality Matters
Result: Changes in Travel Time
The median travel time decline by 31 to 59 percent.
Strongest improvements come from the passenger travel on railroads,
due to HSR.
1995 2000 2005 2010 2015
5
10
15
20
25
30
(a) Road
1995 2000 2005 2010 2015
5
10
15
20
25
30
(b) Railroad, Freight
1995 2000 2005 2010 2015
5
10
15
20
25
30
(c) Railroad, Passenger

Quality Matters: First Look at the Data
Ignoring quality measures leads to substantial bias in travel time
estimation:
over-estimating the reduction in travel time on the roads due to
spatial bias.
and under-estimating that on the railroads, due to temporal bias.
0 0.2 0.4 0.6 0.8 1
0
1
2
3
4
5
6
7
8
(d) Road
0 0.2 0.4 0.6 0.8 1
0
1
2
3
4
5
6
7
8
(e) Railway, Freight
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
(f) Railway, Passenger
Figure 1: Reduction in Prefecture-Pair Travel Time, 1994 to 2017, Baseline v.s.
Uniform Quality

External Validation with the Timetable Data
Does quality-adjustment matter?
Compare the travel time estimations with and without quality
adjustments to the observed timetable data from bus and railroad
service providers.
We do not use any timetable data in constructing our dataset, this
exercise offers an external validation test to check if quality
adjustments matter.
Long-distance bus service timetable in 1999: cross-section.
The railroad timetable in 2002 and 2010: over time.
We compute the prefecture-to-prefecture travel time from the
timetable and then compare to our measures.

External Validation with the Timetable Data
Bus Timetable Train Timetable
(1) (2) (3) (4) (5) (6)
Travel Time, Baseline 1.601*** 1.688***
(0.084) (0.292)
Travel Time, Uniform 1.555*** -0.086
(0.084) (0.283)
∆ Travel Time, Baseline 0.215*** 0.228***
(0.023) (0.030)
∆ Travel Time, Uniform 0.255*** -0.071
(0.061) (0.077)
N 1,823 1,823 1,823 534 534 534
Adj.R2 0.667 0.652 0.667 0.143 0.024 0.143
Table 2: External Validation with the Timetable Data
LHS is the observed travel time from timetables. Unit is prefecture-pair.
Both travel time estimates highly correlate with the recorded travel time of the
long-distance buses.
Column (3) runs the horse race, and quality-adjustment matters!
Similar story in Column (6)

Structural Exercise
Outline
1 Introduction
2 Data
Quality Matters
Model
Quantification
Results

Structural Exercise Model
General Environment
Dynamic migration model from CDP (2019), extended to allow for
multiple modes of transportation and route choices (AA, 2022).
J geographically segmented domestic cities, indexed by j = 1, 2...J,
and an additional ROW denoted as j = 0.
Time is discrete and infinite, indexed as t = 0, 1, 2, ....
Migration is forward-looking, subject to frictions that depend on
passenger transportation and policy barriers.

General Environment
Dynamic migration model from CDP (2019), extended to allow for
multiple modes of transportation and route choices (AA, 2022).
J geographically segmented domestic cities, indexed by j = 1, 2...J,
and an additional ROW denoted as j = 0.
Time is discrete and infinite, indexed as t = 0, 1, 2, ....
Migration is forward-looking, subject to frictions that depend on
passenger transportation and policy barriers.
The direct cost of moving from j to i at time t using mode m, for
passengers (p):
dmp
ijt = ap
· amp
·

Tmp
ijt
hp
, (1)
Tmp
ijt is the distance measured in travel time from our dataset.

The Recursive Migration Problem
vjt = log

φjt
wjt
Pjt

+ max
i,rm∈∪M
m=1Rm
ij
(
δE[vi,t+1] − ¯
dit −
K
X
k=1
dmp
rm
k−1
,rm
k
,t + κ · it,rm
)
,
(2)
Two differences relative to CDP (2019):
1 The dynamic choice set is destination-by-route: i, rm
∈ ∪M
m=1Rm
ij ,
following AA (2022).
2 Migration costs are time-varying and depends on infrastructure
changes.
φjt = φ̄jLβ
jt is the endogenous amenity, wjt/Pjt the real wage, and δ
the discount rate.
κ is the migration elasticity.
¯
dit is the entry barrier, the hukou system.

The Recursive Problem: Solution
Let Vjt ≡ E[vjt] denote the expectation of the continuation value.
The optimization problem defined in equation (2) adopts the same
solution as in CDP, conditional on λijt:
Vjt = log

φjt
wjt
Pjt

+ κ log
J
X
i=1
exp (δVi,t+1 − λijt)1/κ
!
, (3)
λijt is the expected migration costs across all possible modes and
routes from j, conditional on moving to i:
λijt = ¯
dit − κ log


M
X
m=1
X
rm∈Rm
ij
exp

−
PK
k=1 dmp
rm
k−1
,rm
k
,t

κ



 . (4)
Solving λijt
The dynamic migration probabilities follow:
µijt =
exp (δVi,t+1 − λijt)1/κ
PJ
i0=1 exp (δVi0,t+1 − λi0jt)1/κ
. (5)

Production and Trade
The production and trade side follows Allen and Arkolakis (2022):
every city is able to produce every variety, ω. Market structure is
perfect competition.
Labor is the only input:
qjt = Ajt`,
Ajt = Ājt (Ljt)α
is the endogenous productivity.
Consumers decide an origin and a route to source each variety. The
direct ad valorem cost from moving from j to i for freight
transportation (f) is:
dmf
ijt = exp

af
· amf
·

Tmf
ijt
hf

, (6)
Moving from i to j through a route rm of K steps incurs a
multiplicative cost of
QK
k=1dmf
rm
k−1,rm
k ,t ≥ 1.

Production Cont’d
Consumers in i face the following price if they source variety ω from
location j through route rm:
pijt,rm (ω) =
wjt
Ajt
QK
k=1 dmf
rm
k−1,rm
k ,t
zijt,rm (ω)
,
zijt,rm (ω) is i.i.d. Frechet shock with a shape parameter θ.
The probability that individuals in i source from j along route rm is:
πijt,rm =
(wjt/Ajt)−θ QK
k=1

dmf
rm
k−1
,rm
k
,t
−θ
PJ
j0=0 wj0t/Aj0t
−θ
τ−θ
ij0t
, (7)
τijt is the expected transportation cost between i and j at time t:
τijt =



M
X
m=1
X
rm∈Rm
ij
K
Y
k=1

dmf
rm
k−1
,rm
k
,t
−θ



− 1
θ
. (8)
Solving τijt

Trade Balance and Equilibrium
Standard trade balance condition holds every period:
wjtLjt =
J
X
i=0
(wjtτijt)−θ

AjLα
jt
θ
PJ
j0=1 wj0tτij0t
−θ

Aj0 Lα
j0t
θ
witLit (9)
The equilibrium:
the time-invariant fundamentals as Ω = {Aj, φ̄j},
the time-variant fundamentals as Ωt = { ¯
djt, Dmp
t , Dmf
t },
and the sequence of endogenous variables as Υt = {wjt, Ljt, Vjt, µjt}.
Standard concept of steady state: Υt = Ῡ, ∀t, and transition path to
a steady state, on which:
1 Individuals maximize their life-time utility by choosing a sequence of
locations so that equations (5) hold.
2 Firms maximize their profits in each period and trade balances, so that
equation (9) holds.

Structural Exercise Quantification
Overall Setup
We quantify the model to 291 prefecture cities in China plus 1 region
representing the ROW.
Calibration and estimation done on a 50-year transition path from
1995, t = 1. From 1995 to 2017, we apply the actual transportation
networks and policy.
Assume that the transportation and policy stayed the same after 2017,
and therefore the long-run steady state is the one implied by the 2017
variables.
Do not need to assume that 1995 is in steady state. Instead, year
1995 is on a transition path towards its initial steady state.
The transition path is solved in levels, which allows for flexibility in
choosing the moments in quantification.

Externally-Calibrated Parameters
Some parameters are taken from the literature:
name value source notes
α 0.1 Redding and Turner (2015) agglomeration elasticity
β -0.3 Allen and Arkolakis (2022) congestion elasticity
θ 6.2 Allen and Arkolakis (2022) trade elasticity
κ 2.02 Caliendo et al. (2019) migration elasticity
η 6.0 Anderson and van Wincoop (2004) elasticity of substitution
Table 3: External Parameters

Two-Layered Nested Estimation
All the other parameters are calibrated and estimated on the
transition path. We adopt a nested estimation strategy.
In the outer layer, we estimate the mode-specific costs and time
elasticity of trade and migration costs using the GMM. These
parameters are Θχ =

{amf
, amp
}M
m=1, hf
, hp
.
details of the outer layer
All the other parameters are jointly-calibrated in the inner layer:
Θ =
n
{Āj, φ̄j}J
j=0, af
, ap
, ψ, τ∗
o
.
details of hukou system details of the inner layer
The inner-layer calibration is done for every guess of Θχ from the
GMM estimation.

Structural Exercise Results
Counterfactuals and Outline
Compare between the “baseline” and the “uniform” counterfactual.
The baseline equilibrium uses the quality-adjusted transportation
networks.
The uniform counterfactual uses the travel time ignoring quality
differences of infrastructure across time and space.

Counterfactuals and Outline
Compare between the “baseline” and the “uniform” counterfactual.
The baseline equilibrium uses the quality-adjusted transportation
networks.
The uniform counterfactual uses the travel time ignoring quality
differences of infrastructure across time and space.
In both cases, we start from 1995 and solve the model forward 50
years towards a long-run steady state.
From 1995 to 2017, the actual data. No change after 2017.
Compare the prefecture-level growth rates to gauge the bias
introduced by ignoring quality.
We also use the model and data to evaluate the welfare impacts of
transportation networks. (details in the paper).
In short, better connectivity lowers spatial inequality.
details of baseline

Quality Matters: Magnitude of Error
Ignoring quality leads to errors in prefecture-level growth rates.
The median error in real wage growth rates is 19% to 31%.
Even with some winsorization, the error at the 99th percentile
between 175% and 542%!
Real Wage Growth Rates Output Growth Rates
50th 90th 99th Max 50th 90th 99th Max
T=10 0.33 0.67 5.42 70.30 0.07 0.34 2.68 4.12
T=20 0.21 0.35 3.56 10.69 0.05 0.28 3.59 31.14
Steady State 0.19 0.33 1.75 11.08 0.03 0.31 3.36 13.81
Relative Error ≡ |gi,uniform/gi,baseline − 1|.

Quality Matters: Errors are not Random
The errors are not random: ignoring quality systematically
over-estimates the growth rates in poor locations.
corr = -0.62
(a) Real Wage Growth Rates
corr = -0.49
(b) Output Growth Rates
Figure 2: Absolute Error Against Initial Conditions

Non-Random Errors Leads to Distributional Issues
The distributional impacts of transportation networks are particularly
sensitive to quality measures.
Why? The quality-related measurement errors correlate with the
initial conditions!
100 × β-coef. 100 × ∆ Gini coef. 100 × ∆ std(log)
Baseline Uniform Error Baseline Uniform Error Baseline Uniform Error
T=10 -0.43 -0.60 0.39 -0.18 -0.23 0.29 -0.26 -0.35 0.34
T=20 -0.96 -1.24 0.29 -0.38 -0.46 0.21 -0.57 -0.71 0.24
Steady State -1.14 -1.43 0.25 -0.43 -0.51 0.19 -0.67 -0.81 0.21
Table 4: Relative Errors in Spatial Inequality

Conclusion
We create the first panel dataset on the transportation networks that
accounts for quality differences in China between 1994 and 2017.
The expansion of transportation networks significantly lowered the
average travel time, especially for initially remote locations.
Accounting for quality differences is essential for understanding the
welfare impacts of transportation networks. Ignoring quality could
either under- or over-estimate the changes in transportation networks
and their welfare impacts.
The biases are large and persistent in both structural and reduced-form
exercises.
The biases are non-random, and therefore cannot be interpreted and
dealt with as measurement errors.

Appendix
Outline
4 Appendix

Appendix
Terrain of China
Back
3 20
200
300
Plains
Low Rolling Hills (LRH)
Hills
Mountains
Slope (degrees)
LER
(meters)
(a) Definition of Terrain (b) Terrain of China
Figure 3: Definition of Terrains, and the Terrain of China

Appendix
Railway Design Codes
Back
Codes: Code for Design of Railway Line CDRL (III,IV)1
CDSRLIF2
Revision: 1985 1999 2006 2017 2012 1987
Doc.Number: GBJ90-85 GB50090-99 GB50090-2006 TB10098-2017 GB50012-2012 GBJ12-87
National I:
National II:
National III:
National IV:
Local I:
Local II:
Local III:
Industrial I:
Industrial II:
Industrial III:
Table 5: The Mapping Between Railroad Rates and the Codes of Railroad Design
1
Code for Design of III and IV Rated Railway Line.
2
Code for Design of Standard Railway Line for Industrial Firms.

Appendix
List of Maps
Back
Year National Maps Publisher Scale Projection Provincial Map Publisher
1994 Sino Maps 1:6 million Albers, 25N, 47E Sino Maps
1995 N/A Sino Maps
1996 Sino Maps 1:4.5 million Albers, 25N, 47E Global Maps
1997 N/A Global Maps
1998 N/A Xi’an Maps
2000 Sino Maps 1:6 million Albers Xi’an Maps
2002 Sino Maps 1:4.5 million Albers, 25N, 47E Xi’an Maps
2003 Sino Maps 1:6 million Albers, 25N, 47E Xi’an Maps
2006 N/A Hunan Maps
2007 Guangdong Maps 1:6 million Lambert, 24N, 46N, 110E Dizhi
2008 N/A Dizhi
2009 Sino Maps 1:4.5 million Albers, 25N, 47E Renmin Jiaotong
2010 N/A Dizhi
2011 N/A Dizhi
2012 Sino Maps 1:4.5 million Albers, 25N, 47E Dizhi
2013 Sino Maps 1:4.6 million Albers, 25N, 47E Renmin Jiaotong
2014 N/A Sino Maps
2015 N/A Sino Maps
2016 N/A Sino Maps
2017 Sino Maps 1:6 million Albers Sino Maps
Table 6: List of References

Appendix
Example of New Construction
Back

Appendix
Solution to λijt
Back
Following the methods in Allen and Arkolakis (2022), we can further
express λijt as a function of the transportation networks and policy
barriers:
λijt = ¯
dit − κ log
M
X
m=1
bmp
ijt
!
, (10)
bmp
ijt is the (i, j)th element of the matrix Bmp
t , which is a function of
the observed transportation networks and the parameters of the
model.
Bmp
t ≡ (I − Fmp
t )
−1
=
∞
X
K=0
(Fmp
t )
K
,
Fmp
t ≡

exp −
dmp
ijt
κ
!#

Appendix
Solution of τijt
Back
Similar to λijt, we can solve τijt as:
τijt =
M
X
m=1
bmf
ijt
!− 1
θ
.
where bmf
ijt the (i, j)th element of the matrix Bmf
t , which is a function
of the observed freight transportation networks and parameters of the
model:
Bmf
t ≡
∞
X
K=1

Fmf
t
K
=

I − Fmf
t
−1
Fmf
t ≡

dmf
ijt
−θ

.

Appendix
Hukou System
Back
We map the entry barrier in the migration frictions, ¯
djt, to the hukou
reform index constructed in Fan (2019).
We extend the index to the 291 prefectures between 1994 and 2017
using the same archives.
Denote the observed hukou index in prefecture j as kjt ≥ 0, we then
model:
¯
djt = exp (ψkjt) .
ψ is a parameter to be estimated.

Appendix
Inner-Layer Joint Calibration
Back
name value target notes
{Āj } - output in 1995 fundamental productivity
{φ̄j } - population in 1995 fundamental amenity
af 0.815 internal-trade-to-GDP ratio, 2002 average internal trade costs
ap 5.823 average annual stay rate between 2000 and 2005 average migration costs
τ∗
1.614 average export-to-GDP ratio, 2000 to 2005 international trade barrier
ψ 0.039 average annual stay rate between 2010 and 2015 impacts of hukou reform on ¯
dit
{Āj, φ̄j} come from inverting the model in 1995, following Kleinman
et al. (2021).
ψ captures the impact of hukou reform on internal migration,
conditional on the changes in transportation.
Many target moments are not available in the initial year, but only on
the transition path. Solving the model in levels allows us to utilize
these moments.

Appendix
The Outer Layer GMM: Moments, Freight
Back
The first set of moments is the average share of freight traffic that
goes through the road, rail, and waterway networks each year across
prefectures.
Observed in the statistical yearbooks.
These moments capture {amf
}M
m=1. Normalize ariver,f
= 1.
The share of sales from j via mode m is the weighted average of
pair-specific shares across destinations:
smf
jt =
J
X
i=1
bmf
ijt
PM
m0=1 bm0f
ijt
!
Xijt
Xjt
. (11)
The moment condition is the vector of average shares across
prefecture each year between 1995 and 2017:
nPJ
j=1 smf
jt /J
o23
t=1
.

Appendix
The Outer Layer GMM: Moments, Passenger
Back
The second set of moments is for passenger flows. These moments
capture {amp
}M
m=1. Again, normalize ariver,p
= 1.
The share of population flow from j to all the other prefectures via
mode m can be computed similarly as:
smp
jt =
J
X
i=1
bmp
ijt
PM
m0=1 bm0p
ijt
!
Lijt
Ljt
, (12)
The associated moment condition is
nPJ
j=1 smp
jt /J
o23
t=1
.
The asymmetry in the migration costs and entry barriers do not affect
the choice of transportation mode because the route/mode choice is
conditional on a destination location.

Appendix
The Outer Layer GMM: Moments, CV
Back
The last set of moments is the coefficient of variations (CV) for
traffic volumes. They capture the time elasticities {hf
, hp
}, the
parameters that control the variations in τijt and λijt conditional on
{Tmχ
t }.
As hχ declines, the trade or migration costs matrix becomes uniform
and less dependent on the underlying geography summarized in
{Tmχ
t }.
The resulting variations in trade or migration traffic across prefecture
will decline, leading to a smaller CV.

Appendix
The Outer Layer GMM: Moments, CV
Back
The last set of moments is the coefficient of variations (CV) for
traffic volumes. They capture the time elasticities {hf
, hp
}, the
parameters that control the variations in τijt and λijt conditional on
{Tmχ
t }.
As hχ declines, the trade or migration costs matrix becomes uniform
and less dependent on the underlying geography summarized in
{Tmχ
t }.
The resulting variations in trade or migration traffic across prefecture
will decline, leading to a smaller CV.
The moment conditions are the vectors of the average CV across
prefecture each year between 1995 and 2017:
PJ
j=1 CVf
jt
J
23
t=1
and
PJ
j=1 CVp
jt
J
23
t=1
.

Appendix
GMM
Back
Denote the 92 moment conditions in the data as the vector S̄ to
estimate 6 parameters, and the counter-parts in the model as:
S (Θχ
) =
(PJ
j=1 smf
jt
J
,
PJ
j=1 smp
jt
J
,
PJ
j=1 CVf
jt
J
,
PJ
j=1 CVp
jt
J
)23
t=1
.
The GMM estimator finds Θχ to minimize the distance between the
data and the model moments:
min
Θχ

S̄ − S (Θχ
)

W

S̄ − S (Θχ
)
0
,
W is the optimal weighting matrix, which is the inverse of the
variance-covariance matrix of the data moments S̄, computed by
bootstrapping the data.

Appendix
Estimation Results
Back
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5
-3
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
Figure 4: Goodness-of-Fit, GMM

Appendix
Estimation Results
Back
name value s.e. notes
a1,f
0.875 0.003 road costs, freight
a2,f
1.052 0.020 rail costs, freight
hf
0.050 0.010 time elasticity, freight
a1,p
0.889 0.028 road costs, passenger
a2,p
1.208 0.039 rail costs, passenger
hp
0.200 0.016 time elasticity, passenger
Table 7: Parameters, Estimated

Appendix
Baseline and No-Change Equilibrium
Back
Compare the steady states and transition paths between the
“baseline” and the “no-change” counterfactual.

Appendix
Back
The baseline equilibrium is the same transition path that we used to
estimate the model.
From 1995 to 2017, the actual data. No change in transportation and
policy after 2017.
The steady state implied by the variables in 2017.

Appendix
Back
The baseline equilibrium is the same transition path that we used to
estimate the model.
From 1995 to 2017, the actual data. No change in transportation and
policy after 2017.
The no-change counter-factual:
Fix the transportation networks and policy to those in 1995.

Appendix
Results: Aggregate and Dynamic Impacts
Back
The expansion of transportation networks increases long-run output
by about 60%.
1995 2000 2005 2010 2015 2020 2025 2030 2035 2040
1
1.01
1.02
1.03
1.04
1.05
1.06
1.07
1.08

Appendix
Back
Return to trade liberalization is immediate but wanes in the long-run.
By 2017, almost all the benefit comes from goods transportation.
1995 2000 2005 2010 2015 2020 2025 2030 2035 2040
1
1.01
1.02
1.03
1.04
1.05
1.06
1.07
1.08

Appendix
Back
Return to migration liberalization could be negative in the short-run
but grows over time.
1995 2000 2005 2010 2015 2020 2025 2030 2035 2040
1
1.01
1.02
1.03
1.04
1.05
1.06
1.07
1.08

Appendix
Delayed Return to Migration Liberalization
Back
The short-run negative return to migration liberalization comes from
the delayed movements towards more productive locations.
People expect better infrastructure and policy in the future, and thus
they prolong their stay in the less productive prefectures.
1995 2000 2005 2010 2015 2020 2025 2030 2035 2040
8.5
8.6
8.7
8.8
8.9
9
9.1
9.2
9.3
9.4
9.5
(a) Top 10% Productive Locations
1995 2000 2005 2010 2015 2020 2025 2030 2035 2040
4.5
5
5.5
6
6.5
7
(b) Bottom 10% Productive Locations
Figure 5: The Dynamic Response to Migration Liberalization

Appendix
Results: Distributional Impacts
Back
β-convergence between 1995 and the long-run steady state in 2044.
Without the expansion of the network, the spatial inequality would
have stayed the same over time.
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
-0.04
-0.02
0
0.02
0.04
0.06
0.08
Baseline beta = -0.0114
No Change beta = -0.0023

Appendix
Internal Trade as an Equalizer
Back
Reduction in τ explains about 80% of the regional convergence.
Changes λ is responsible for about 9%, and statistically insignificant.
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
-0.02
0
0.02
0.04
0.06
0.08
0.1
beta = -0.0109
s.e. = 0.0039
(a) τ only
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
beta = -0.0027
s.e. = 0.0014
(b) λ only
Figure 6: The distributional impacts of trade and migration liberalization

References
Alder, Simon, Michael Zheng Song, and Zhitao Zhu, “Unequal
Returns to China’s Intercity Road Network,” 2021. working paper.
Allen, Treb and Costas Arkolakis, “The Welfare Effects of
Transportation Infrastructure Improvements,” The Review of Economic
Studies, 02 2022.
and Dave Donaldson, “The geography of path dependence,”
Technical Report, working paper 2018.
Anderson, James E. and Eric van Wincoop, “Trade Costs,” Journal of
Economic Literature, September 2004, 42 (3), 691–751.
Artuc, Erhan, Shubham Chaudhuri, and John McLaren, “Trade
Shocks and Labor Adjustment: A Structural Empirical Approach,” The
American Economic Review, 2010, 100 (3), 1008–1045.
Asher, Sam and Paul Novosad, “Rural Roads and Local Economic
Development,” American Economic Review, March 2020, 110 (3),
797–823.
Baum-Snow, Nathaniel, J. Vernon Henderson, Matthew A. Turner,
Qinghua Zhang, and Loren Brandt, “Does investment in national

References
highways help or hurt hinterland city growth?,” Journal of Urban
Economics, 2020.
Berg, Claudia N., Brian Blankespoor, and Harris Selod, “Roads and
Rural Development in Sub-Saharan Africa,” The Journal of
Development Studies, 2018, 54 (5), 856–874.
Caliendo, Lorenzo, Maximiliano Dvorkin, and Fernando Parro,
“Trade and Labor Market Dynamics: General Equilibrium Analysis of
the China Trade Shock,” Econometrica, 2019, pp. 741–835.
Donaldson, Dave and Richard Hornbeck, “ Railroads and American
Economic Growth: A “Market Access” Approach *,” The Quarterly
Journal of Economics, 02 2016, 131 (2), 799–858.
Fan, Jingting, “Internal Geography, Labor Mobility, and the Distributional
Impacts of Trade,” American Economic Journal: Macroeconomics,
2019, 11 (3), 252–88.
, Yi Lu, and Wenlan Luo, “Valuing Domestic Transport Infrastructure:
A View from the Route Choice of Exporters,” The Review of Economics
and Statistics, 08 2021, pp. 1–46.

References
Jedwab, Rémi and Adam Storeygard, “The Average and
Heterogeneous Effects of Transportation Investments: Evidence from
Sub-Saharan Africa 1960–2010,” Journal of the European Economic
Association, 06 2021, 20 (1), 1–38.
Kleinman, Benny, Ernest Liu, and Stephen J Redding, “Dynamic
Spatial General Equilibrium,” Working Paper 29101, National Bureau of
Economic Research July 2021.
Ma, Lin and Yang Tang, “Geography, trade, and internal migration in
China,” Journal of Urban Economics, 2020, 115, 103181.
Redding, Stephen J. and Matthew A. Turner, “Chapter 20 -
Transportation Costs and the Spatial Organization of Economic
Activity,” in Gilles Duranton, J. Vernon Henderson, and William C.
Strange, eds., Handbook of Regional and Urban Economics, Vol. 5 of
Handbook of Regional and Urban Economics, Elsevier, 2015, pp. 1339 –
1398.
Tombe, Trevor and Xiaodong Zhu, “Trade, Migration, and
Productivity: A Quantitative Analysis of China,” American Economic
Review, May 2019, 109 (5), 1843–72.

Appendix
Xu, Mingzhi, “Riding on the new silk road: Quantifying the welfare gains
from high-speed railways,” 2017.

The Distributional Impacts of Transportation Networks in China

Recommended

Recommended

More Related Content

Similar to The Distributional Impacts of Transportation Networks in China

Similar to The Distributional Impacts of Transportation Networks in China (20)

Recently uploaded

Recently uploaded (20)

The Distributional Impacts of Transportation Networks in China