1. travel
2. tourist destinations
3. japan

Production Networks, Geography and Firm Performance
Andrew B. Bernard
Andreas Moxnes
Yukiko U. Saito
Tuck @ Dartmouth
NBER, CEPR and CEP
University of Oslo
NBER and CEPR
RIETI
Becker Friedman Institute, April 4, 2015
1 / 36
Motivation and Questions
What determines buyer-supplier (firm-to-firm) connections?
What are the consequences for firm performance?
Mechanism:
Firms have comparative advantage (CA) in producing a given task.
But searching for suppliers (observing price/quality) is costly.
Trade-off between benefits from exploiting CA and cost of search.
2 / 36
Motivation and Questions
What determines buyer-supplier (firm-to-firm) connections?
What are the consequences for firm performance?
Mechanism:
Firms have comparative advantage (CA) in producing a given task.
But searching for suppliers (observing price/quality) is costly.
Trade-off between benefits from exploiting CA and cost of search.
2 / 36
Implications
1) Variation in firm output and productivity across space (Sveikauskas
1975, Glaeser and Mare 2001, Combes et al 2012).)
Central locations have lower costs of finding high quality suppliers & outsource
more.
2) Substantial heterogeneity in firm sales (w/in localities and industries).
High productivity firms have an incentive to search harder for good suppliers.
3) Effect of infrastructure on firm performance.
Lowers the cost of searching for suppliers.
3 / 36
Implications
1) Variation in firm output and productivity across space (Sveikauskas
1975, Glaeser and Mare 2001, Combes et al 2012).)
Central locations have lower costs of finding high quality suppliers & outsource
more.
2) Substantial heterogeneity in firm sales (w/in localities and industries).
High productivity firms have an incentive to search harder for good suppliers.
3) Effect of infrastructure on firm performance.
Lowers the cost of searching for suppliers.
3 / 36
Implications
1) Variation in firm output and productivity across space (Sveikauskas
1975, Glaeser and Mare 2001, Combes et al 2012).)
Central locations have lower costs of finding high quality suppliers & outsource
more.
2) Substantial heterogeneity in firm sales (w/in localities and industries).
High productivity firms have an incentive to search harder for good suppliers.
3) Effect of infrastructure on firm performance.
Lowers the cost of searching for suppliers.
3 / 36
Three Components of the Paper
Facts about (Japanese) production network.
I
Comprehensive data on (nearly) complete production network.
Model of producers and domestic sourcing.
I
Building on Antras, Fort and Tintelnot (2014).
‘Natural’ experiment testing predictions of model (and effects of
infrastructure).
I
I
Kyushu Shinkansen expansion in 2004.
Up to 75% fall in travel time for persons, 0% for goods.
Disclaimer: This paper is not about relocation of inputs or firms within-firm identification only.
4 / 36
Data Sources
TSR (Tokyo Shoko Research):
Credit reporting agency (1 of 2 in Japan)
Buyer-supplier linkages in 2005 & 2010 + firm sales & geolocation.
950,000+ firms in private sector.
I
Close to complete coverage of firms with 5+ employees.
F
F
Not limited to a particular sector.
More than 50% of all firms in Japan (relative to census).
Kikatsu:
1998-2008 data from “The Results of the Basic Survey of Japanese
Business Structure and Activities” - balance sheet plus much more.
All firms with 50+ employees & capital of more than 30 million yen
(USD 300,000).
5 / 36
TSR Data - Network
Each firm provides a rank ordered list of suppliers & customers (max
24).
We use union of own-reported and other-reported information.
I
A supplies B if both firms are in the TSR data and
F
F
A reports B as customer or
B reports A as supplier.
6 / 36
Reported Links Example
(TSR) In-degree = 2 (1 own+reported + 1 other-reported)
(TSR) Out-degree = 2 (1 own+reported + 1 other-reported)
7 / 36
The Production Network : Facts
Key relationships that inform the model:
1
Larger firms have more suppliers.
2
The majority of connections is formed locally.
3
Larger firms have suppliers in more locations and their distance to
suppliers is higher.
1
4
Within firms, distance is positively correlated with supplier/customer
performance measures.
Negative degree assortativity among sellers and buyers.
8 / 36
Fact I : Larger firms have more suppliers
Indegree
100
10
1
1
10
100
1,000
10,000
Sales, mill yen
kernel = epanechnikov, degree = 0, bandwidth = .2, pwidth = .3
9 / 36
Fact II : The majority of connections is formed locally
.02
Data
Random
Density
.015
.01
.005
0
0
500
1000
1500
2000
Distance, km
Median (mean) distance to connections: 30 (172) km.
10 / 36
Fact III : Larger firms have suppliers in more locations
# municipalities
100
10
1
1
10
100
1,000
10,000
Sales, mill yen
kernel = epanechnikov, degree = 0, bandwidth = .21, pwidth = .32
Slope=0.27
11 / 36
Fact III : Larger firms have higher distance to suppliers
Median distance to suppliers
100
10
1
10
100
1,000
10,000
Sales, mill yen
kernel = epanechnikov, degree = 0, bandwidth = .29, pwidth = .44
Slope=0.32
12 / 36
Fact III: Distant suppliers/buyers have better performance
Suppliers
Within-buyer regressions
In-degree Out-degree Employment
Sales/worker
Distance
0.12a
0.04a
0.17a
0.15a
includes industry, island, supplier-prefecture and buyer fixed effects
Customers
Within-supplier regressions
In-degree Out-degree Employment
Sales/worker
Distance
0.07a
0.01a
0.06a
0.06a
includes industry, island, buyer-prefecture and supplier fixed effects
Within buyers, supplier quality is increasing in distance.
Within suppliers, buyer quality is increasing in distance.
13 / 36
Fact IV : Negative degree assortativity
Mean outdegree of suppliers
1000
100
10
1
1
10
100
1000
Indegree
A firm with more suppliers - those suppliers have fewer customers.
A firm with fewer suppliers - those suppliers have more customers.
14 / 36
The Model
We build on sourcing model of Antras et al (2014) and introduce:
In-house production vs outsourcing margin.
Continuum of locations.
15 / 36
The Model : Upstream
Upstream stage:
Unit continuum of tasks ω produced in location i.
PF yU (ω) = zU (ω) lU (ω).
Task productivity zU (ω) from Frechet (T , θ ).
Iceberg trade costs: τ (i, j) ≥ 1.
Perfect competition.
16 / 36
The Model : Downstream
Downstream firm in j use labor and all tasks ω.
PF y (z, j) = zl α v (z, j)1−α . v (z, j) is CES task composite, z is
efficiency.
ω produced in-house or outsourced:
In-house: PF yI (ω) = zI (ω) lI (ω).
I
I
Task productivity zI (ω) from Frechet (T0 , θ ).
No trade costs.
Outsourced: Firm sees price CDF in i but not individual prices
p (ω, i).
I
Firm in j must pay f (j) to observe individual p (ω, i) (paid in terms of
labor).
Monopolistic competition & CES final demand.
17 / 36
The Model : Assumptions
For tractability:
T0 and T the same everywhere.
Perfect labor mobility −→ wages same everywhere.
No trade costs on final good.
Positive measure of downstream firms in each location j.
Restrict to interior solution.
18 / 36
The Firm’s Problem
Solve by backwards induction:
Conditional on locations searched, firm chooses in-house / outsourcing
location for each task ω.
Firm chooses locations to search, characterized by cutoff τ¯ (z, j):
highest trade cost location.
I
τ¯ (z, j) chosen to balance the benefit of lower MC against the cost of
search.
19 / 36
Model and Data
1
2
More productive firms outsource more tasks and therefore have more
suppliers:
∂ ln o (z, j)
> 0,
∂ ln z
More productive firms search more and costlier locations:
∂ ln τ¯
>0
∂ ln z
3
Negative degree assortivity: Higher z (high indegree) −→ firm reaches
costlier locations −→ suppliers there are on average not very
competitive in z’s home market (low avg. outdegree).
20 / 36
A Distributional Assumption
Every location faces a density of trade costs g (τ, j).
Assume g () inverse Pareto with shape γ > θ and support [1, τH ].
I
A location has few nearby markets and many remote ones.
Density fits empirical distance distribution well.
21 / 36
Two Propositions
Proposition
(i) Lower search costs f (j) lead to growth in sales among downstream
firms in j. (ii) Sales growth is stronger in input-intensive (low α) relative to
labor intensive (high α) industries.
∂ ln r (z, j)
< 0 and
∂ ln f (j)
∂ 2 ln r (z, j)
>0
∂ ln f (j) ∂ α
Two channels:
1
2
Direct: low α firms grow more bc inputs large share of total costs.
Indirect: low α firms search more markets when f (j) ↓ (|∂ τ¯/∂ f |
decreasing in α).
22 / 36
Two Propositions
Proposition
Lower search costs f (j) lead to more outsourcing and suppliers from new
locations (higher τ¯) among downstream firms in j.
∂ o (z, j)
< 0 and
∂ f (j)
∂ τ¯ (z, j)
< 0.
∂ f (j)
23 / 36
Shinkansen - A Natural Experiment
Does cheaper search improve firm performance by facilitating linkages in
the production network?
Use opening of high-speed train network (Shinkansen) in southern
Japan in 2004. Key advantages:
1
Dramatic reduction in travel time between stations.
I
2
Goods do not travel by Shinkansen, just people.
I
3
75% reduction for many city pairs.
No contemporaneous reduction in travel time for goods along this
southern route.
Likely exogenous.
I
Planned decades in advance. Timing of completion was subject to
substantial uncertainty.
24 / 36
45
40
Latitude
35
30
130
135
140
145
Longitude
Rail line connecting two prefectures (Kagoshima + Kumamoto) with
total population of 3.5 million (~CT).
Travel time Kagoshima −→ Shin-Yatsushiro: 130 −→ 35 min.
Kagoshima −→ Hakata: 4 −→ 2 hours.
25 / 36
800 series Shinkansen
Operating speed 260 km/h.
2-3 departures / hour. Capacity 392 passengers per trainset.
26 / 36
Empirical Methodology
Lower travel time should benefit input-intensive firms more than labor
intensive firms (Proposition 1).
I
Lower f (j) has no impact on MC of firms belonging to α = 1 industries.
Classify industries k according to their 2003 intermediate input use:
Hk = 1 − labor share of industryk
Define Treatf = 1 if firm f is < 30 km from new Shinkansen station
(stations between Kagoshima and Shin-Yatsushiro).
Dependent variables: lnSales, ln(sales/employee), TFP (Olley-Pakes);
relative to industry-year means.
27 / 36
Empirical Methodology
Estimate for 2000-2008 period
ln yfkrt = αf1 + αrt2 + β1 Treatf × Hk × Post2004t + γXfkrt + εfkrt ,
where αf1 and αrt2 are firm and prefecture-year fixed effects.
Triple differences:
I
I
I
Pre to post shock (1st diff)
Firm near station relative to firm not near station (2nd diff).
High Hk relative to low Hk firms (3rd diff).
Positive β1 if high Hk firms are growing faster relative to low-Hk firms
near new stations relative to elsewhere.
More controls:
I
I
Time-varying geographic controls by using average performance in f ’s
municipality (≈ 1400 municipalities).
Remaining interactions (Treatf × Hk , etc.).
28 / 36
Potential Concerns
1
We pick up market access (demand side) effects:
I
2
We pick up different trends for input- and labor-intensive firms:
I
3
No, industry trends are differenced out.
The location of the stations are endogenous:
I
4
No, because demand ↑ should affect both input- and labor-intensive
firms.
Not a problem as long as locations are not determined based on
differential growth for input/labor intensive industries.
Pre-trends; input-intensive firms near new stations always grow faster
relative to labor-intensive firms:
I
No evidence of this in placebo test.
29 / 36
Results
Table: Firm Performance
Treatf × Hj × Post2004t
Firm and city controls
Prefecture-year FE
Firm FE
N
R-sq
Sales
Sales/employee
TFP
0.47∗∗
(2.12)
Yes
Yes
Yes
148,264
0.97
0.42∗
(1.76)
Yes
Yes
Yes
146,466
0.92
0.29∗∗
(2.44)
Yes
Yes
Yes
145,058
0.94
Note: Robust t-statistics in parentheses. Dependent variables in logs.
A Shinkansen station increases sales by 0.47 log points more for a firm
with Hk = 1 relative to a firm with Hk = 0.
A firm in the 9th decile of the Hk distribution (industrial plastic
products) increased sales by 0.10 log points more than a firm in the
1st decile of the Hk distribution (general goods rental and leasing).
30 / 36
Robustness: Placebo
Use 1998-2002 data and Post2000t dummy.
Treatf × Hj × Post2000t
Firm and city controls
Prefecture-year FE
Firm FE
N
R-sq
Sales
Sales/employee
TFP
-0.30
(1.05)
Yes
Yes
Yes
66,756
0.99
-0.05
(0.22)
Yes
Yes
Yes
66,756
0.94
0.02
(0.17)
Yes
Yes
Yes
66,487
0.95
Note: Robust t-statistics in parentheses. Dependent variables in logs.
31 / 36
More Robustness
1
Labor supply - Recruiting now easier for knowledge intensive industries
(which may happen to be input intensive).
I
2
The ’straw effect’ - less economc activity in nearby locations.
I
3
Calculate R&D intensity of industries, add additional interactions −→
no change in results.
Add interactions for firms 30-60km from new station −→ small
negative effect for these firms & no change in main results.
Demand side again - Input intensive industries may have more remote
customers.
I
I
Should not see TFP effects.
corr avg distance to customers,Hj = −0.02.
4
Drop construction industry.
5
Change 30 km threshold.
32 / 36
Shinkansen - New Connections
Mechanism: Should see more supplier linkages in treated regions.
Divide Japan into a grid consisting of 500 × 500 locations (5.62 km2 ).
Number of connections from i to j at time t is Cijt , t = {2005, 2010}.
Regress
∆ ln Cij = ξi1 + ξj2 + β1 Bothij + β2 Oneij + γXij + εij ,
where ξi1 and ξj2 are source and destination FE, Bothij = 1 if both
locations i and j get a new station, Oneij = 1 if one of them gets a
new station.
Notes:
I
I
Using 2005-2010 changes at the firm level is (currently) problematic
because of cross-year matching issues.
Lower bound bc data starts in 2005.
33 / 36
Shinkansen - New Connections
Table: Connections: Descriptive statistics.
Cij2005
Cij2010
∆ ln Cij
Bothij
Oneij
Firms per cell
# sources
# destinations
# obs
Mean
Median
7.05
7.51
0.08
0.01
0.02
104.21
7,613
8,054
386,294
2
2
0
0
0
19
Std. dev.
62.35
59.93
0.56
0.08
0.14
463.30
min
max
1
1
-3.47
0
0
1
16507
14808
3.78
1
1
21,207
34 / 36
Shinkansen - New Connections
(1)
Bothij
Oneij
(2)
(3)
(4)
0.07∗∗∗
(5.91)
-0.02∗∗∗
(3.56)
0.12∗∗∗
(7.91)
-0.01
(0.74)
0.39∗∗∗
(20.12)
0.19∗∗∗
(19.87)
-0.06∗∗∗
(71.32)
No
No
386,294
Yes
Yes
386,294
7,613
8,054
0.17
Yes
Yes
386,294
7,613
8,054
0.18
0.42∗∗∗
(7.93)
0.15∗∗∗
(6.42)
-0.06∗∗∗
(81.98)
-0.01
(0.86)
0.01∗
(1.87)
Yes
Yes
386,294
7,613
8,054
0.18
ln Distij
Bothij × ln Distij
Oneij × ln Distij
Destination FE
Source FE
# obs
# sources
# destinations
R-sq
0.00
Note: Bootstrapped t-statistics in parentheses with 200 replications. Dependent variable is ∆ ln Cij = ln Cij2010 −
ln Cij2005 . ∗∗∗ significant at the 0.01 level, ∗∗ significant at the 0.05 level, ∗ significant at the 0.1 level.
35 / 36
Conclusions
The supply network matters for firm performance:
I
I
Infrastructure shock generates significant performance gains.
Evidence that gains are related to new buyer-seller linkages, as
suggested by model.
Directions for future work:
I
I
The entry margin - how heterogeneous firms sort into different markets.
Identifying additional firm performance channels.
36 / 36