S TRUCTURE AND C HANGE IN P RODUCTION
N ETWORKS : E VIDENCE FROM US F IRM
LEVEL D ATA
Vasco M Carvalho and Michael Gofman
CREi/UPF/CEPR & UW - Madison
Tokyo, November 2012
I NTRO
I
Value intermediate input transactions is as large as GDP
I
I
I
these ‡ows take place along supplier-customer networks
emerging literature stresses how these production networks
facilitate the propagation of shocks and a¤ect aggregates
Yet, limited evidence about the network structure of
production and how it changes over time.
I
Supplier-customer relationships are missing in time-series or
cross-sectional analysis of …rms.
I
Our goal is to document important empirical facts about
production networks that can help to evaluate existing models
and inspire new models.
I
Today I will present our preliminary results based on a new
proprietary database of supplier-customer relationships.
W HAT WE DO
I
Use a new database of supplier-customer relationships to
construct production chains
I
Study the relationship between …rm characteristics and
existence of an supplier-customer relationship
I
Search for evidence of positive assortative matching in
production chains
I
Study whether …rm characteristics help understanding the
duration of supplier-customer relationships
I
Investigate whether matching with more productive suppliers
has an impact on the customer
R ELATED L ITERATURE
I
Firm-level network: Cohen and Frazzini (2008), Atalay et al
(2011), Gofman (2011), Buraschi and Porchia (2012),
Ober…eld (2012)
I
Industry-level network: Carvalho (2010), Menzly and Ozbas
(2010), Acemoglu et al (2012), Antras et al (2012)
I
Country-level network: Rizova (2011), Chaney (2012)
D ESCRIPTION OF OUR DATASET
I
New proprietary panel dataset of customer-supplier
relationships
I
Information is collected from SEC …lings, press releases,
websites, interviews, and earnings transcripts as well as
primary research by provider’s analysts
I
The data is collected and sold to hedge funds and
corporations for portfolio construction, risk management,
competitive and supply chain analysis.
I
The dataset is similar to Compustat segment data, but it has
more relationships because it uses additional sources and is
updated daily.
D ATA
I
Quick facts about the data:
I
I
I
I
I
I
180,000+ unique relationships (supplier, customer,
competitors, partners) reported by for 15,000+ US-traded
companies
Relationships reported about public, private, foreign, domestic,
government organizations, and educational institutions
34,000 supplier-customer relationships with percent of sales to
this customer
Spanning 2003-2011 (daily frequency)
Observe start and end dates of customer-supplier relationship
We use:
I
I
I
38,725 unique supplier-customer relationships for publicly
listed 5,260 …rms
US …rms, relationships longer than 90 days
Able to match with Compustat
S TRUCTURE OF P RODUCTION C HAINS
F IGURE : Core of the Production Chains in 2010.
I-O N ETWORK D ATA : F IRMS
I
Proprietary data: for each
…rm gives most important
suppliers & customers
I
Updates it on a weekly
basis; Jan 2003 December 2011
I
Hundreds of thousands of
I-O relationships
I
Over 15000 …rms (>6000
publicly traded …rms in
US)
I
Source: SEC …lings, press
releases, websites,
interviews, earnings
transcripts
I
& primary research by the
I-O N ETWORK D ATA : S ECTORS
403
37
72
5
34
71
61
2
34
81
93
367
402
404
405
385
382
359
361
360
377
373
396
386
383
398
395
321
323
264
397
3 2 43 2 2 2 8 9 2 9 8
362
401
393
391
354
319
91
400
399
284
3 0 72 7 8
394
2
7
5
3
3
0
372
390
3 0 63 1 4
340
291
366
2 9 53 1 6
356
3 2 03 1 3
311 276
342
3 0 52 6 2 2 8 0
312
365
274
378
341
345
3 2 82 8 3
302
350
3 3 2 1 7 14 7 3 3 0 93 2 9 3 2 7
88
304
268
50
279
339
282
4 8 3 0 38 1 5
30383 392
51
265
287
285
85
288 310 301
270
89
348 334
266
54
293 49
269
3 3 8 2 9 63 3 5 2 8 1
352
326 86
347
443
343 346
70
55
442
297
56
415
292 47
66 69
196
241 277
52
73 379 434
410 344
349
441
84
325
363
318
58
3 3 31 5 0 3 3 13 0 0
351409
64
87
68
233
421
256
294
147
9 02 2 92 5 32 3 6
364
63 67
62
201 250
65
53
267
4
2
7
246
408
81
271
112
424
227
249200 251
106 135
2 2 82 5 5
57
71
213
299
60
425 426
75
59
74 221
420
423
61
238 231
78
248
198 273
92
79
252
433
217
239 234
430
1 6 42 4 7
2 2 31 9492 2
2 4 2 4 3 54 1 8
243
432
232
82
148 211
240
2 2 6 2 1 4 2 3 54 1 2
2 5 73 8 0
215
80
417 431
225
195
2 0 82 3 7
286
4
3
6
4
1
429
222
358
42
152
134
118 212
230
28 428 416
133
72
35
77
20637
245
129 224220
2 0 92 6 1
219
76
254
181 357
151 125
194
130
39
244
29
204 44 218
116
202
149
138
191
437
145
2 0 31 8 4
216
114
95
43 205
172
115
176
38
107
140139
98
414
141
439
207
185
210
197
128
1 8 61 8 8
137 20
193
144
440
131
108
31
438 111
163
16
146
179
32
142 153
17
187
127 19 117
182
110
33
124
27
162
136
96
12
161
159
419
30
1 2 31 2 6
143
374
192
40
122
18
160
109
34
8
167
189
444
45
120 119
190
14
154 260
132
15 23
183
10
171
11
97 13
121
177
175
155
258 259
24
102
180 178
156
26
7
157
25
158
46
22
168
94 93
169 165
9
1
3
170 166
21
6
5
2
4
101
99
353
368
369
406
381
336
371
370
407
384
387
388
290
355 263
317
337
100
BEA Detailed Input-Use Data 1997.
(5% Threshold)
36
I
Detailed Sector I-O data
from BEA
I
Every …ve years:
1967-2002
I
' 500 sectors
Consistent with US
National Accounts
113
103
104
105
I
I-O N ETWORK D ATA : S ECTORS -C OUNTRIES
0
200
400
I
Sector-Country I-O data
from WIOD
I
Every year: 1995-2009
I
35 sectors in 40 countries
I
Most of world trade
included.
600
800
1000
1200
1400
0
200
400
600
800
nz = 88126
1000
1200
1400
N ETWORK T OPOLOGY OF I-O F LOWS
IO/Year
Firms (06)
US Sectors (02)
Int. Trade (06)
I
n
8961
422
1485
n_scc
1709
259
964
d
5
11
11
All Directed, Unweighted Networks
`
5.88
4
6.51
dm
19
10
19
r (out, in )
-0.04
-0.11
-0.05
c
0.08
0.32
0.42
N ETWORK T OPOLOGY OF I-O F LOWS
IO/Year
Firms (06)
US Sectors (02)
Int. Trade (06)
I
n
8961
422
1485
n_scc
1709
259
964
d
5
11
11
All Directed, Unweighted Networks
I
Short Average Path Length
`
5.88
4
6.51
dm
19
10
19
r (out, in )
-0.04
-0.11
-0.05
c
0.08
0.32
0.42
N ETWORK T OPOLOGY OF I-O F LOWS
IO/Year
Firms (06)
US Sectors (02)
Int. Trade (06)
I
n
8961
422
1485
n_scc
1709
259
964
d
5
11
11
All Directed, Unweighted Networks
I
I
Short Average Path Length
Short Diameter
`
5.88
4
6.51
dm
19
10
19
r (out, in )
-0.04
-0.11
-0.05
c
0.08
0.32
0.42
N ETWORK T OPOLOGY OF I-O F LOWS
IO/Year
Firms (06)
US Sectors (02)
Int. Trade (06)
I
n
8961
422
1485
n_scc
1709
259
964
d
5
11
11
`
5.88
4
6.51
All Directed, Unweighted Networks
I
I
I
Short Average Path Length
Short Diameter
Some Evidence for Negative Assortativity
dm
19
10
19
r (out, in )
-0.04
-0.11
-0.05
c
0.08
0.32
0.42
N ETWORK T OPOLOGY OF I-O F LOWS
IO/Year
Firms (06)
US Sectors (02)
Int. Trade (06)
I
n
8961
422
1485
n_scc
1709
259
964
d
5
11
11
`
5.88
4
6.51
dm
19
10
19
r (out, in )
-0.04
-0.11
-0.05
All Directed, Unweighted Networks
I
I
I
I
Short Average Path Length
Short Diameter
Some Evidence for Negative Assortativity
Little Clustering in Firm Net relative to IO or International
Nets
c
0.08
0.32
0.42
P ROBABILITY OF I NPUT-S UPPLY R ELATIONS
I
Do …rm characteristics help predict the existence of
supplier-customer relationship?
I
Logit speci…cation:
Pr (i ! j ) = F ( β0 + β1 Sizei + β2 Sizej + β3 L.Prod.i + β4 L.Prod.j + β5 SICij )
I
where:
I
I
I
I
I
F (.) is the cumulative logistic distribution
Size Proxies = log fSales g or log fEmployees g
Labor Productivity log (Sales ) log (Employees )
SICij =1 if i and j are in the same 4 digit SIC sector
We implement logit for all input-supply diads in 2005
P ROBABILITY OF I NPUT-S UPPLY R ELATIONS
I Large, productive …rms are more
likely to demand from more …rms
Indep. variable
Log employees of supplier
Log employees of customer
Labor productivity supplier
Labor productivity customer
Same 4 digit SIC
3676158
0.134
N
Pseudo R 2
Robust Standard Errors;
Pr (i ! j )
0.04
0.75
0.02
0.59
2.24
signi…cant at 1%
I
1 s.d. of size (labor prod.)
associated to 75% (59%)
higher prob. that a link exists
I Large, productive …rms are more
likely to supply to more …rms
I
1 s.d. of size (labor prod.)
associated to 4% (2%) higher
prob. that a link exists
I Most trade is within narrowly
de…ned sectors
I
Firms in same 4 digit SIC are
twice as likely to trade with
each other
D ATA
D ISPERSION IN SIZE OF FIRMS AS INPUT DEMANDERS
10
Pr(X ≥ x)
10
10
10
10
0
-1
-2
-3
-4
10
0
10
1
10
x
2
3
10
Top 10 Input Demanders, 2005
Firm
# Suppliers
Wal-Mart
204
IBM
203
General Electric
179
Hewlett-Packard
178
AT&T
167
Sprint Nextel
149
Verizon
131
Boeing
129
Ford
128
Lockheed Martin
120
P ROBABILITY OF I NPUT-S UPPLY RELATIONSHIPS
P OSITIVE A SSORTATIVE M ATCHING ?
Indep. variable
Log employees of supplier
Log employees of customer
Interaction of Log employees
Labor productivity supplier
Labor productivity customer
Interaction productivity
Same 4 digit SIC
Pr (i ! j )
0.04
0.75
0.03
0.03
0.33
0.05
2.24
N
Pseudo R 2
3676158
0.135
I Large, productive …rms are more
likely to supply to other large,
productive …rms
I
I
1 s.d. of size interaction
associated to 3% higher prob.
that a link exists
1 s.d. of productivity
interaction associated to 5%
higher prob. that a link exists
I Rank correlation also points to
Robust Standard Errors;
signi…cant at 1%
(modest) PAM:
I
I
Corr(Sizei ,Sizej )=0.07
Corr(L. Prodi ,L.
Prodj )=0.06
T URNOVER IN I NPUT-S UPPLY RELATIONSHIPS
We observe high turnover in input-supply relationships
.0025
.002
Density
.001
.0015
I
For the average year, 48% of all input-supply relations are
either formed (23%) or will cease to exist (25%) during that
year
Mean duration of an input supply relation is 2.4 years but this
conceals a lot of heterogeneity:
5.0e-04
I
0
I
0
1000
2000
3000
duration
Histogram of Duration of input-supply relations (in days)
T URNOVER IN I NPUT-S UPPLY RELATIONSHIPS
I
Do …rm characteristics help understanding the duration of
supplier-customer relationships?
I
Duration analysis (accounting for right-censoring):
I
Let θ (t, Xt ) be the hazard rate at time t for input-supply
relation characterized by covariates Xt
I
I
I
Recall: hazard rate gives the probability of terminating an
input-supply relation at t conditional on surviving till time t
We implement a simple Weibull parametric speci…cation
(robust to other speci…cations)
Look at how the hazard rate depends on Xt covariates:
I
I
I
I
Size and productivity of supplier and customer in the relation
Productivity of supplier relative to (simple) average
productivity other suppliers of customer i
Productivity of customer relative to (simple) average
productivity other customers of supplier j
Note: all variables given by average observed during
input-supply relation
T URNOVER IN I NPUT-S UPPLY RELATIONSHIPS
D ETERMINANTS OF DURATION OF INPUT- SUPPLY RELATIONSHIPS
Indep. variable
Log employees of supplier
Log employees of customer
Interaction of Log employees
Labor productivity supplier
Labor productivity customer
Interaction of labor productivity
Productivity of supplier w.r.t. other suppliers of j
Productivity of customer w.r.t. other customers of i
Interaction of relative productivities
N
Robust Standard Errors;
Hazard rate
1
0.99
1.01
0.65
0.63
1.01
0.82
0.81
0.92
22081
signi…cant at 1%
Not reported: Evidence for increasing hazard rate over duration of an
input-supply relation
T URNOVER IN I NPUT-S UPPLY RELATIONSHIPS
D ETERMINANTS OF DURATION OF SUPPLIER – CUSTOMER RELATIONSHIPS
I
More productive …rms are more likely to engage in longer
input-supply relationships
I
I
For a given customer, its most productive suppliers tend to
supply them for longer
I
I
a 1 s.d. increase of labor productivity of supplier (customer)
leads to a 35% (37%) lower hazard rate
a 1 s.d. increase in the gap between a supplier’s productivity
and the average productivity of suppliers of a given customer is
associated with 18% lower hazard rate
For a given supplier, its most productive customers tend to
demand from them for longer times
I
a 1 s.d. increase in the gap between a customer’s productivity
and the average productivity of customers of a given supplier is
associated with 19% lower hazard rate
F IRM - LEVEL PERFORMANCE AND TURNOVER OF
SUPPLIERS
I
Above …ndings suggest selection of suppliers (on productivity):
I
I
a high turnover of suppliers (increase in the number of
suppliers)
more productive suppliers tend to supply inputs for longer
times
I
Does matching with more productive suppliers have an impact
on the customer?
I
Panel analysis:
I
I
I
Dependent variable: customer’s growth rates of sales,
employment and labor productivity in year t
Independent variables: turnover rate of a …rm’s suppliers in
year t, growth rate of average labor productivity of input
suppliers in year t and interaction term
Firm and year …xed e¤ects throughout + lagged growth rates
of dependent variables
F IRM - LEVEL PERFORMANCE AND TURNOVER OF
SUPPLIERS
Indep. variable / Dep. variable
Turnover of suppliers
∆ Avg l.p. of suppliers
Turnover
∆Avg l.p. of suppliers
Customer & year …xed e¤ects
Lagged dependent variable
N
I
∆Salesjt
0.06
0.05
0.06
∆Employmentjt
0.03
0.01
0.01
∆L.P.jt
0.03
0.05
0.05
Yes
Yes
Yes
Yes
Yes
Yes
9689
9689
9689
Higher supplier turnover rates are associated with customer’s
growth (in sales, employment and labor productivity)
I
a 1 s.d. increase in the turnover rate of a …rm is associated
with a 3p.p. increase in labor productivity growth
F IRM - LEVEL PERFORMANCE AND TURNOVER OF
SUPPLIERS
Indep. variable / Dep. variable
Turnover of suppliers
∆ Avg l.p. of suppliers
Turnover
∆Avg l.p. of suppliers
Customer & year …xed e¤ects
Lagged dependent variable
N
I
∆Salesjt
0.06
0.05
0.06
∆Employmentjt
0.03
0.01
0.01
∆L.P.jt
0.03
0.05
0.05
Yes
Yes
Yes
Yes
Yes
Yes
9689
9689
9689
Growth in the average productivity of suppliers is associated
with growth in sales and productivity of customers
I
a 1 s.d. increase in the average productivity growth of suppliers
is associated with a 5p.p. increase in labor productivity growth
F IRM - LEVEL PERFORMANCE AND TURNOVER OF
SUPPLIERS
Indep. variable / Dep. variable
Turnover of suppliers
∆ Avg l.p. of suppliers
Turnover
∆Avg l.p. of suppliers
Customer & year …xed e¤ects
Lagged dependent variable
N
I
∆Salesjt
0.06
0.05
0.06
∆Employmentjt
0.03
0.01
0.01
∆L.P.jt
0.03
0.05
0.05
Yes
Yes
Yes
Yes
Yes
Yes
9689
9689
9689
Growth in average productivity of suppliers through adding
and dropping suppliers is associated with growth in sales and
productivity of customers
I
a 1 s.d. increase in the interaction term is associated with a
further 5p.p. increase in labor productivity growth
R ECAP
I
Preliminary set of stylized facts on …rm level input-supply
relationships:
I
I
I
I
I
I
Large, productive …rms are more likely to supply to (and
demand from) more …rms
Large, productive …rms are more likely to supply to other large
and productive …rms (PAM)
There is high turnover of input-supply links
More productive …rms tend to engage in longer input-supply
relationships
For a given customer (supplier), its most productive suppliers
(customers) tend to supply (demand from) them for a longer
time
Growth in average productivity of suppliers through adding and
dropping suppliers is associated with growth in sales and
productivity of customers
F UTURE WORK
I
Further study of supplier-customer relationships with focus on
network formation models
I
Using our descriptive results as moments for testing
theoretical models that generate those moments
I
Compare structure of production networks with other
networks (e.g. social networks)