B
Data Appendix
In the United States, a public firm must report the name of all major customers who
represent 10% or more of the firm’s sales. This is due to the SFAS No. 176 requirement
that mandates firm disclosure of significant customers. Section B.9 discusses the mandatory
disclosure in more detail. Customer names are typically reported in footnotes of financial
statements. This data are extracted and stored in the Compustat Historical Customer
Segment database. Although some firms may voluntarily disclose their customers, I do not
exclude them from the final sample.23
The final sample in the data comprises 8,110 unique firms from 1984 through 2014,
represents 244 three-digit SIC industries (out of 276 in all Compustat), and have headquarters
in all 50 states. It is made up of public firms that are customers and/or suppliers of other
public firms. The sample only includes firms with positive total assets and sales, firms
with headquarters in the United States, and only includes non-financial firms. The state
headquarter location will be relevant for the instrumental variables approach and comes
directly from historical 10-K SEC filings available on WRDS SEC Filings data. I do not
use Compustat state information since it only reports the state headquarter from the most
recent SEC filing.
Although the final sample only contain firms with observed customers or suppliers, it represents about half of all Compustat production annually. Since firms only report important
customers, the firm-level network constructed from the disclosure requirement is subject to
truncation bias. Customer firms that are large relative to supplier firms are over-represented
in the data, and firms with many small customers are not observed. In the sample, additional customer connections through time may be mechanically observed for several reasons.
Sales to other customers might decrease, sales to the particular customer increases, or some
combination of a two that results in the fraction of sales to the customer to be higher than
23
Excluding connections where customer weights are less than 10% reduces the sample size but does not
affect the main results.
74
10%. This data censorship introduces a bias when studying the number of customers that
a firm has. Moreover, whether firms adjust on the intensive or extensive margin also affects
the bias. For example, during the 2008 - 2009 financial crisis, many firms connections decrease, but the fraction of sales to remaining customers increase. Although empirical results
using the number of customers are susceptible to this bias, Atalay et al. (2011) show that
the censorship is less of an issue when studying the number of suppliers.24
Compared to the United States industry flows in 2007, the final sample also represents at
least 80% of some flows.25 Industry input-output data for the United States comes from the
Bureau of Economic Analysis. The data contains dollar values of pair-wise industry inputoutput flows for 389 industries defined based on NAICS codes. The final sample represents
flows from Wholesale Trade to Retail, Retail to Healthcare, and Retail to Manufacturing
flows particularly well. However, the sample is not representative of all pair-wise industrylevel flows. Given the focus on firm-level production networks where the final producer
faces consumers, the model applies for the subset of firms studied. I discuss the sample
representativeness in more detail in Section B.7.
B.1
Sample Construction
The network data construction is relatively standard, following other research papers that
have previously used firm network data. Compustat provides customer segment data. This
data contains the names of the reported customers as well as a categorization of customer
type. The network data is constructed by matching customer names with company names on
using several different string matching algorithms. First, I use a typographical error-robust
24
Atalay et al. (2011) show that empirically, the probability of an edge existing between two firms depends
only on the dollar sales between two firms relative to the supplier’s total sales, and is not correlated to other
observable characteristics such as the industry, size, or physical distance between its suppliers or customers.
The analysis only uses Bayes’ Rule and relies on two assumptions: (1) the joint probability distribution is
additively separable in the fraction of sales and firm characteristics, and (2) firms actually follow the SFAS
No. 14 federal mandate to disclose customers whose purchases represents 10% or more its total sales. They
also estimate the effect on the in-degree distribution by using a kernel-weighted polynomial regression to
approximate the probability distribution of customer shares for the observed data, and find that qualitatively
the shape of the in-degree distribution is similar with the actual estimated in-degree distribution.
25
The comparison excludes international flows in the Bureau of Economic Analysis data.
75
name matching algorithm after removing extra white spaces. After the initial matches, I
follow Cohen and Frazzini (2008) and use the Soundex name-matching method. Following
this, I allow for fuzzy matches with propensity match scores above 0.995, reviewing each
pair individually. I verify all matches by hand for accuracy. I only include links between
domestic public firms.
The firm network data must be combined with firm balance sheet information. To do
this, I consider connections keeping only reported customers who fall under the “Company”
category and drop large consumers that are government agencies, private firms, or other
organizations. Balance-sheet information for publicly listed firms in Compustat range from
1950 onward. Firm-level connections data are available from 1976 onward, so there is no
time truncation introduced by this procedure.
B.2
Data Processing & Filters
Although there is not much processing or cleaning prior to the sample construction described above, I outline the steps to pull the data. I use SAS to access all the data in WRDS.
To get the fullest Compustat sample possible, I consolidate data formats. Some data are
reported with under the “Standardized Summary” format while others are in the “Standardized” format. One is not an obvious subset or duplication of the other, so I consolidate
observations by gvkey and fiscal year, replacing missing values if at least one of the formats
has a non-missing value for that column, and prioritizing the “Standardized” format over
the “Standardized Summary” format. Then, all duplicates by GVKEY and fiscal year are
removed.
To maximize the data set and network granularity, I compute network statistics using
the network data after matching names to GVKEYs in Compustat but before placing filters
on balance sheet information. All network summary statistics are based on the network
data before placing filters on balance sheet information. After calculating firm connection
measures like in-degree and out-degree, I merge the firm-level network data to balance sheet
76
information, place the appropriate filters, and proceed with the analysis.
Firm headquarter state information comes from the SEC data available on WRDS starting from 1992. To be able to preserve the sample size, I impute firm headquarters backwards
from 1992 to be able to study tax changes since 1989. Using only tax changes from 1992
onward does not change any results. Classical mis-measurement in state headquarters will
attenuate the empirical results, and I discuss this in more detail in the Empirical Results
section. Compared to this method, Heider and Ljungqvist (2015) extend the state headquarter information by supplementing the data on WRDS SEC Filings with data from Thomson
Reuters.
B.3
Aggregate Production Network Features
and Summary Statistics
Certain features of the firm-level production network provide guidelines for an appropriate
model. The dis-aggregated firm-level network reveals telling features that differs starkly from
its industry-level counterpart. The firm-level production network has low levels of reciprocity
and triadic closure. Low reciprocity means that meaning firm A supplies to firm B, but firm
B does not supply firm A. Low triadic closure means that firm B’s customer is not also
firm A’s customer, and firm B’s customer also does not supply to firm A. The observed
production network suggests a directed network with distinct production tiers, a model set
up that allows for a solution method to solve for equilibrium in the entire network.
Because firm-level production networks show little reciprocity and transitivity, a tiered
network model with no cycles captures the production flow from upstream firms to downstream firms well. Less than 1% of the connections exhibit reciprocity. All connections with
reciprocity are conglomerates that report sales between subsidiaries. Table 1 shows that a
firm and its customer share customers less than 0.5% of the time. Therefore, these network
characteristics suggest a directed production tier structure where goods flow from a primary
producer through intermediate producers to final producers.
77
The production network is stable. The aggregate network maintains an average of 75%
of the observed links from one year to another. Of the links that persist, the average cosinesimilarity between customer weights from one year to the next is over 90%. The similarity
stems from stability in actual dollar flows between firms. Pairwise firm connections also seem
sticky, with an average link lasting over 3 years. The shortest relationship observed is 1 year
and the maximum is 35 years. Around 26% of the firm connections are to firms within the
same Fama-French industry. This shows that production tiers do not line up exactly with
industry definitions.
The number of customers and suppliers are right-skewed. The average number of customers is 1.21 and number of suppliers is 1.27, while the median firm in the sample has one
customer and no suppliers. Bristol-Myers Squibb, a pharmaceutical company, voluntarily
disclosed the most number of customers - 38 in 2013 and 2014. On the other hand, Walmart
had the most suppliers, with 157 suppliers reporting it as a major customer in 2013.
Firms whose customers are large relative to their own size are more likely required to
report these business links. Figure 13 shows the distribution of customer sales relative to
supplier sales. Customers tend to be very large relative to suppliers. However, Figure 14
suggests that despite this asymmetry, there is still large variation in the dollarized sales
from customer to supplier. The amount of dollars of traded goods may be large or small,
but relative to they are still significant relative to the supplier’s overall sales. For example,
in 2009 PetroChina sells 45% of its output to China Petroleum & Chemical Corp. This
45% of sales is $67 bn. In this case, both the fraction of sales and overall dollar sales are
economically large.
78
Figure 13: Customer-Supplier Size Ratio
The plot below shows the distribution of ratios of customer sales to supplier sales.
Distribution of Sales Ratio
Frac of Links
0.3
0.2
0.1
0.0
0.01
10.00
10,000.00
10,000,000.00
Customer Sales / Supplier Sales
Figure 14: Dollar Sales in Customer-Supplier Links
The plot below shows the distribution of a firm directed sales.
Distribution of Directed Sales
0.4
Density
0.3
0.2
0.1
0.0
0.1
100.0
100,000.0
Millions $
B.4
Sample Entry and Exit
To get a sense of data availability, I show the number of firms in the network data in
Figure 15. Figure 16 shows the number of clusters in the data through time.
79
Figure 15: Size of Network Data
The plot below shows the number of firms and clusters in the Compustat firm network data. The gray bars
are NBER recession periods.
Nodes & Connections in Firm Network
Correlation = 0.94
Nodes
Connections
2000
1000
1980
1990
2000
2010
Figure 16: Number of Clusters
The plot below shows the number of firms and clusters in the Compustat firm network data. The gray bars
are NBER recession periods.
Clusters in Firm Network
200
150
100
50
1980
1990
2000
80
2010
Figure 17: Sample Cluster Distribution
The plot below shows the distribution of firm clusters in the Compustat firm network data for 2012.
Cluster Distribution for 2012
% of Observations
0.6
0.4
0.2
0.0
1
2
3
4
5
6
Cluster Size
7
8
9
10
>10
Firm entry into and exit out of the sample increase around 2000 and 2009, but otherwise
hovers around 22-36% of existing firms. Whether supply chain networks are sticky has
implications for its suitability as a driver of leverage. Overall, a large fraction of existing
links change over the years, although the net changes are small, as shown in Figure 18. The
identities of customers and suppliers in the aggregate network do not significantly change
through time.
81
Figure 18: Data Turnover
The plot below shows some measures of the Compustat firm-level network structures through time. The
gray bars are NBER recession periods.
Firm Network Data Turnover
Exit
Links
1980
1990
2000
2010
Unique CustomersUnique Suppliers
% of Observations
Entry
60
50
40
30
20
10
60
50
40
30
20
10
60
50
40
30
20
10
Firms leave the final sample either because the firm relationships are terminated or simply
due to truncation bias. Truncation is possible if the supplier firm has a major increase in sales
while the sale to the previous customer remains relatively unchanged. It is also possible if
firms decrease sales along the intensive margin and increase sales along the extensive margin,
adding customers without affecting total sales. The latter truncation is due to strategic
sourcing or sales, and is arguably part of the economic mechanism outlined in the paper.
To address the former mechanical truncation due to increases in overall sales of a supplier,
I evaluate whether the results are robust to conditioning on supplier sales increasing or
decreasing and find no effect on the empirical results.
B.5
Production Tier Classification
The classification of production tiers trades off more precise measures of distance from
final consumers with sample size. I group firms into three production types: primary producers, intermediate producers, and final producers. Primary producers are firms that have
82
supplied products to other firms at least once since 1980 but have never been important
customers of other public firms. Final producers are firms who have been customers of some
other firm at least once since 1980, but have never been suppliers of other public firms.
Intermediate producers are those who have had both customers and suppliers at least once
since 1980. Chains with two tiers consist of only primary and final producers.
This definition does not induce a look-ahead bias. It assumes that a firm’s true production
tier is sticky.26 Conceptually, the tier of production for a particular firm is related to the
technological production process, complexity of the final good, and regulations that affect
the industrial organization of all related markets. I assume that these are unchanged in the
short-term. My classification method allows firms to move from being primary and final
producers to intermediate producers, but not vice versa.
Because all intermediate producers have both suppliers and customers, the economics for
all intermediate producers are similar even if in reality they do not have the same distance
to final consumers. For example, suppose a mining firm supplies a silicon chip manufacturer,
the silicon chip manufacturer supplies a phone assembly firm, and the phone assembly firm
supplies a telecommunications firm. In this production chain, although the silicon chip
manufacturer is further away from the final consumer compared to phone assembly firm, they
are both classified as intermediate producers. In the data, the longest observed production
chain has 6 tiers of production. Supply chains of such lengths are rare, so estimates of
balance sheet information across thinner tier buckets will be noisy. No matter how long
the actual production chain is, the basic tier structure of primary producers, intermediate
producers, and final producers remains the same.
Measurement error in tier classifications will attenuate differences across tiers. Supply
chain networks consisting of only two tiers are defined to include only a primary and final
producer. Empirically, 60% of the observed firm production chains consist of only two tiers,
26
Defining tiers based only on observed customers or suppliers in a particular year attenuates the differences in average leverage across tiers. The attenuation arises naturally if each year’s number of customers
or suppliers are measured with noise.
83
while the remaining clusters consist of at least three production tiers. The data truncation causes intermediate producers to be under-represented relative to primary and final
producers.
The firm-level production tier classification is related to industry definitions. Letting tier
numbers to represent the distance from final consumers, final producers are Tier 1, intermediate producers are Tier 2, and primary producers are Tier 3. Although the tier classification
is subject to some measurement error, it shows intuitive patterns across industries: Retail
is most like a final producer and Fabricated Products is closer to a primary producer. In
the model, final producers face prices that are more flexible because the realized final goods
markets must clear. Meanwhile, there are no spot markets for inputs. Binding supply chain
contracts impose fixed prices in the short term.
The production tiers in the directed network framework are also related to measures of
network centrality used in existing network research. Many existing studies of equilibrium
network formation and relationships generate first-order conditions that specify key relationships between outcome variables and measures of centrality.27 For example, Gao (2014) finds
that firms with higher closeness use less leverage, where closeness measures the number of
steps required to access every other vertex in the network from a given vertex. The production tier structure used in this paper has strong implications for closeness. In a symmetric
tiered production network, primary producers will have higher closeness, followed by intermediate producers then final producers. Empirical results from Gao (2014) are consistent
with both empirical and theoretical results in this paper.
27
See Bala and Goyal (2000), Ballester et al. (2006), and Acemoglu et al. (2013). For a thorough
simulation-based study of different network centrality measures, see Valente et al. (2008).
84
B.6
B.6.1
Empirical Results
Leverage and Supply Diversification: Robustness
The relationship remains stable when focusing on state law changes where the corporate
profits tax increase by at least 0.5, 1, or 2%, with the first-stage regression getting stronger
when using indicators for tax changes. The relationship between capital structure and sourcing also does not appear to be driven only by large firms, as shown in Table 8, which allows
for heterogeneous relationships of sourcing and debt based on lagged total assets.
When the instrument violates the relevance condition for leverage, I do not find a relationship between the instrumented debt and supply chain variables. The first stage results
are consistent with results found in Heider and Ljungqvist (2015), where although firms
increase debt when tax rates rise, they do not decrease debt when tax rates fall.
85
Table 7: Long-Term Leverage & Sourcing using Continuous Changes in Tax Rates
The table below shows the relationships between long term leverage and number of suppliers. Long term leverage is instrumented
by the decimal increase in corporate profits tax. All regressions include year by industry fixed effects, where industries are
defined based on 3-digit SIC. Standard errors are clustered at the state-level.
Dependent Variable: Yt =
(1)
T + > 0%
30
1,212
Num Events
Num Affected Firms
\ Lev
LT Book
ROAt
6.619⇤⇤⇤
(1.660)
0.583⇤⇤⇤
(0.206)
0.001
(0.004)
0.385⇤
(0.205)
0.046
(0.140)
0.012
(0.077)
1.573
(2.160)
0.007⇤⇤
(0.003)
13,832
0.575
0.839
1
(M/B)t
1
ln Salet
1
ln Assetst
1
U nemp Ratet
Incomet
1
1
First Stage.
= 1 if tax increase in t =
Num Obs
R2 first stage
R2 second stage
1
⇤
p<0.1;
⇤⇤
Number of Supplierst
(2)
(3)
T + > 0.5%
T + > 1%
21
18
492
465
(4)
T + > 2%
9
383
6.940⇤⇤⇤
(1.972)
0.602⇤⇤⇤
(0.196)
0.001
(0.004)
0.391⇤
(0.219)
0.034
(0.149)
0.017
(0.082)
1.376
(2.333)
0.007⇤⇤
(0.003)
13,832
0.616
0.831
7.392⇤⇤⇤
(2.056)
0.629⇤⇤⇤
(0.200)
0.001
(0.004)
0.399⇤
(0.231)
0.017
(0.156)
0.024
(0.086)
1.099
(2.336)
0.007⇤⇤
(0.003)
13,832
0.616
0.820
p<0.05;
86
⇤⇤⇤
p<0.01
6.745⇤⇤⇤
(1.881)
0.590⇤⇤⇤
(0.196)
0.001
(0.004)
0.387⇤
(0.212)
0.041
(0.144)
0.014
(0.079)
1.496
(2.295)
0.007⇤⇤
(0.003)
13,832
0.616
0.836
Table 8: Long-Term Leverage & Sourcing based on Firm Size
The table below shows the relationships between long term leverage and number of suppliers or customers, allowing for heterogeneous relationships for small versus large firms. Long term leverage is instrumented by an indicator of whether its headquarter
state had a corporate profits tax increase. All regressions include year by industry fixed effects, where industries are defined
based on 3-digit SIC. Standard errors are clustered at the state-level.
For
T + > 0.5
Num Events
Num Affected Firms
Num Affected Firms, size > Median
Dependent Variable: Yt =
21
492
193
(1)
Number of Supplierst
(2)
Number of Customerst
4.922⇤⇤⇤
(1.020)
12.781
(11.400)
0.134
(0.317)
0.0003
(0.002)
0.621⇤⇤
(0.302)
0.448
(0.559)
0.093
(0.119)
1.946
(3.713)
0.025⇤⇤⇤
(0.007)
19,163
0.584
0.875
2.042
(1.429)
1.684
(2.341)
0.128
(0.080)
0.001⇤⇤⇤
(0.0004)
0.112
(0.130)
0.160
(0.140)
0.032
(0.055)
2.932⇤⇤
(1.156)
0.025⇤⇤⇤
(0.007)
19,163
0.584
0.579
\ Lev
LT Book
\ Lev ⇥ 1Sizet
LT Book
ROAt
1 >M edian
1
(M arket/Book)t
ln Salet
1
ln Assetst
1
U nemp Ratet
Incomet
1
1
1
First Stage.
= 1 if tax increase in t =
Num Obs
R2 first stage
R2 second stage
1
⇤
87
p<0.1;
⇤⇤
p<0.05;
⇤⇤⇤
p<0.01
Table 9: Long-Term Leverage & Sourcing with Tax Decreases
The table below shows the relationships between long term leverage and number of suppliers or customers. Long term leverage
is instrumented by an indicator of whether its headquarter state had a corporate profits tax decrease. All regressions include
year by industry fixed effects, where industries are defined based on 3-digit SIC. Standard errors are clustered at the state-level.
For
Num Events
Num Affected Firms
Dependent Variable: Yt =
Tax Change:
\ Lev
LT Book
ROAt
1
(M arket/Book)t
ln Salet
1
ln Assetst
1
U nemp Ratet
Incomet
1
1
1
First Stage.
= 1 if tax decrease in t =
Num Obs
R2 first stage
R2 second stage
1
T <
46
634
0.5
(1)
(2)
Number of Supplierst
Indicator
Continuous
(3)
(4)
Number of Customerst
Indicator
Continuous
9.187
(11.083)
0.342
(0.574)
0.009
(0.012)
0.002
(0.401)
0.674
(0.481)
0.019
(0.084)
1.688
(3.374)
0.010
(0.007)
37,258
0.702
0.815
7.461
(8.818)
0.347
(0.473)
0.010
(0.012)
0.059
(0.352)
0.266
(0.357)
0.034
(0.060)
1.287
(2.561)
0.010⇤
(0.005)
37,258
0.119
0.660
⇤
p<0.1;
32.094
(121.750)
1.243
(5.389)
0.016
(0.061)
0.683
(3.853)
1.464
(4.436)
0.050
(0.242)
8.059
(29.008)
0.0004
(0.001)
33,318
0.720
-0.281
⇤⇤
p<0.05;
⇤⇤⇤
28.291
(96.813)
1.194
(4.313)
0.017
(0.052)
0.688
(3.045)
0.985
(3.514)
0.013
(0.191)
4.318
(23.083)
0.0004
(0.001)
33,318
0.720
0.152
p<0.01
Moreover, it is worth noting alternative economic forces likely biases the effect on the
number of customers upwards. If firms diversify customers rather than suppliers, then we
expect to find a positive coefficient on the number of customers. However, if customers
are also diversifying supply, then suppliers are more likely to have less customers following
leverage increases due to higher supply riskiness.
The truncation bias in the data are non-trivial. The data are truncated either because
some customer public companies may not be reported, and private companies will not be
88
matched in the final dataset due to missing balance sheet information. Regarding the missing
private companies, I look at the representativeness of the sample in Section B.7 by aggregating up to the industry level to get a sense of where the truncation bias may be biggest. This
allows me to see the fraction of overall production flows based on industry-level input-output
represented in the final sample. Additional data sources such as CapitalIQ may help provide
connection and capital structure even on some private firms, but the data are not standardized and are not necessarily audited for accuracy. Apart from that, there are no obvious
solutions for incorporating private network data, since they do not need to report customers
and typically do not make their capital structure information consistently available.
Finally, the financial statement coverage of public firms in Compustat may be less reliable
in the beginning of the period and there may be other regulatory concerns such as the passage
of BAPCA in 2005. To this end, I confirm that all the results in the paper are robust to
starting the analysis after 1992, 1995, or 2005.
B.6.2
Heterogeneity
Table 10 shows that the relationship between capital structure and supply chain variables
does not depend on the production tier. In the model, production tiers differ due to price
stickiness. However, the incentives of primary and intermediate producers to diversify their
supply are the same.
89
Table 10: Long-Term Leverage & Sourcing depending on Producer Type
The table below shows the relationships between long term leverage and number of suppliers or customers, allowing for heterogeneous relationships across producer types. Long term leverage is instrumented by an indicator of whether its headquarter state
had a corporate profits tax increase. There are 662 Final producers, 543 intermediate producers, and 2,119 primary producers
in the regression below. Of these, 15 final producers, 17 intermediate producers, and 56 primary producers were affected by the
tax changes. All regressions include year by industry fixed effects, where industries are defined based on 3-digit SIC. Standard
errors are clustered at the state-level.
For
T + > 0.5
Num Events
Num Affected Firms
Num Affected Firms, Intermediate
Dependent Variable: Yt =
\ Lev
LT Book
\ Lev ⇥ Intermediate
LT Book
ROAt
1
(M arket/Book)t
ln Salet
1
ln Assetst
1
U nemp Ratet
Incomet
1
1
1
First Stage.
= 1 if tax increase in t =
Num Obs
R2 first stage
R2 second stage
1
21
492
145
(1)
Number of Supplierst
(2)
Number of Customerst
7.523⇤⇤⇤
(2.417)
1.300
(1.387)
0.525⇤⇤⇤
(0.158)
0.002
(0.002)
0.428⇤⇤
(0.193)
0.024
(0.182)
0.025
(0.095)
0.888
(2.595)
0.024⇤⇤⇤
(0.007)
19,163
0.596
0.824
4.013⇤
(2.240)
1.219
(0.910)
0.251⇤⇤
(0.113)
0.002⇤⇤⇤
(0.001)
0.128
(0.161)
0.167
(0.128)
0.054
(0.063)
3.871⇤⇤
(1.631)
0.024⇤⇤⇤
(0.007)
19,163
0.596
0.456
⇤
p<0.1;
⇤⇤
p<0.05;
⇤⇤⇤
p<0.01
Focusing on supply chain effects of capital structure, Table 11 shows that firms whose
supplier or customer receive treatment do not significantly adjust their capital structure
unless they are also simultaneously treated. The results vary depending on whether the
treated neighbor is a supplier or customer. When both a firm and at least one of its suppliers
receive tax increases, it does not change its leverage. This is compared to when baseline case
90
where firms increase leverage by 2.4 percentage points following tax increases. This provides
evidence that firms seem to adjust capital structure in response to supplier capital structure
as well. On the other hand, when a firm and at least one of its customers are treated, it still
increases leverage.
Table 11: Treated Neighbors
The table below shows the first stage results in detail, including a dummy of whether at least one of the firm’s neighbor is
treated. The neighbor is a customer in Regression (1) and supplier in Regression (2). The results are for T + > 0.5.
Standard errors are clustered at the state level. All regressions include year by industry fixed effects, where industries are
defined based on 3-digit SIC. Standard errors are clustered at the state-level.
Dependent Variable: Yt =
Num Events
Num Affected Firms
Num Affected Neighbor
Num Affected Both Firm & Neighbor
Neighbor Type:
0.009
(0.011)
0.025⇤⇤⇤
(0.009)
0.036⇤⇤⇤
(0.011)
0.048⇤
(0.026)
0.0003
(0.0003)
0.026
(0.029)
0.037
(0.027)
0.003
(0.008)
0.161
(0.180)
15,381
0.596
N eighbor T reated = 1
T reated = 1
N eighbor T reated = 1 &
T reated = 1
ROAt 1
(M arket/Book)t
ln (Salet
1
ln (Assetst
+ 1)
U nemp Ratet
Incomet
1
+ 1)
1
LT Book Lev
21
492
153
372
66
121
(1)
(2)
Suppliers
Customer
1
1
Num Obs
R2
⇤
p<0.1;
⇤⇤
0.001
(0.017)
0.024⇤⇤⇤
(0.008)
0.019
(0.012)
0.048⇤
(0.026)
0.0003
(0.0003)
0.026
(0.029)
0.037
(0.027)
0.003
(0.008)
0.161
(0.180)
15,381
0.596
p<0.05;
⇤⇤⇤
p<0.01
While a firm indirectly responds to the capital structure of firms in its supply chain, the
first order impact remains whether the firm itself has an increase in the marginal benefit of
91
debt. Given the increase in marginal benefit of debt, firms increase debt and also diversify
their supply. At the same time, customers are also more likely to drop suppliers whose debt
increases.
B.7
Sample Representativeness
To help judge the representativeness of the data, I compare the representativeness of
the final sample to the overall Compustat sample and the aggregated industry-level network
from industry input-output tables provided by the Bureau of Economic Analysis.
B.7.1
Comparison to Cohen and Frazzini (2008)
My sample is very very similar that from Cohen and Frazzini (2008). I compare the
fraction of links going to the same industry using the same 48 industry definitions provided
on Ken French’s website. The numbers using data from 1987 through 2012 are very similar to
the numbers reported in Cohen and Frazzini (2008) which is based on 24 annual observations
from 1981 - 2004.
Table 12: Fraction of Links in the Same Industry
Industry Definition Min Max Mean SD Median
4-Digit SIC
0.08 0.11 0.09 0.01
0.09
Fama-French 48
0.22 0.29 0.26 0.02
0.26
2-Digit SIC Groups 0.36 0.51 0.43 0.05
0.42
Cohen & Frazzini 0.21 0.27 0.23 0.02
0.23
92
Figure 19: Links in the Same Industries
The plot below shows the fraction of firms whose links are to another firm in the same industry, based
on whether industries are defined by 2-digit SIC codes, Fama-French industries, or 4-digit SIC codes. The
ranking for these industry classification from coarsest to finest is 2-digit SIC, Fama-French industries, then
4-digit SIC codes.
Links in the Same Industries
2−digit SIC
FF Ind
4−digit SIC
0.5
Fraction of Links
0.4
0.3
0.2
0.1
1990
B.7.2
1995
2000
2005
2010
Comparison to Overall Compustat
Comparing the representativeness of the final data sample used in the paper to the overall
Compustat firms helps to evaluate the model applicability and amount of debt attributable
to this model. Figure 21 shows the representativeness of the data based on sales, while
Figure 20 shows the representativeness of the data based on overall debt usage. Both these
figures suggest that the sample used in the paper capture almost half of all debt use, the
results in the paper concern an economically relevant sample of the economy.
93
Figure 20: Representativeness based on Debt Usage
The plot below shows the sample representativeness compared to the overall Compustat data. The data are
shown both in levels as well as fraction representativeness. On average across the years available, from 1980
to 2012, 35% of overall short-term debt and 28% of long term debt usage are represented relative to all the
data available in Compustat.
Overall Dollar Debt Representation
In Sample
Not in Sample
20
Long Term Debt
15
10
Trillion $
5
0
40
Short Term Debt
30
20
10
0
1980
1990
2000
2010
2000
2010
Debt Represented
0.4
Long Term Debt
0.3
0.2
Fraction
0.1
0.0
0.4
Short Term Debt
0.3
0.2
0.1
0.0
1980
1990
94
Figure 21: Representativeness based on Sales
The plot below shows the sample representativeness compared to the overall Compustat data. The data are
shown both in levels as well as fraction representativeness. On average across the years available, from 1980
to 2012, 35% of overall short-term debt and 28% of long term debt usage are represented relative to all the
data available in Compustat.
Overall Gross Sales Represented
Final Sample
Not in Final Sample
30
Trillion $
20
10
0
2000
2005
2010
Gross Sales Represented
1.00
Fraction
0.75
0.50
0.25
0.00
2000
2005
95
2010
B.7.3
Comparison to Bureau of Economic Analysis Data
Although the sample does not seem to capture most of industry flows in the entire economy, the aggregated industry-network based on firm-level connections show similar network
features. This is based on two comparisons. First, I compare the Compustat aggregated
industry-level network to the BEA input-output network for 2007 based on NAICS codes
to more directly check the sample representativeness. Second, I note that the Compustat industry-level network aggregated based on 2-digit SIC codes contain similar network
features as those from the BEA.
To see how representative the network information from the Compustat firm-level data is
relative to the BEA industry input-output data, I aggregate both the firm-level and industrylevel data to 4-digit NAICS codes. The industry input-output data are downloaded directly
through the Bureau of Economic Analysis website, through the Industry tab, then through
the Input-Output Accounts data section.28 To get to the finest level of data possible, I
download the 2007 version with 389 industries.29 In the spectrum of firm-level data to aggregated industry-level data, having more finely defined industries will capture more features
similar to the firm-level network. I use this data to show that even using 389 industries, the
production network looks very different from firm-level networks.
Aggregating the firm-level network to industry-level and directly evaluating the representativeness gives room for pause: across 15 NAICs industries, only 2.5% of industry-toindustry flows are correctly represented in 2007. While most flows are not represented by the
Compustat network sample, some flows are very well captured. Table 13 shows that 99.3%
of the total BEA recorded Wholesale Trade to Retail Trade industry flows are captured in
the Compustat sample. 77 out of 225 industry-to-industry links are represented, of which
12 capture at least 10% of the flow, and 24 capture at least 1% of total flows. The largest
truncation factors are due to private firms and small public customer firms. Industries with
28
Link is: http://www.bea.gov/industry/io_annual.htm. Last accessed on May 26, 2016.
This is the most recent version of this data set that was available as of the download date on March 3,
2016. Last accessed and confirmed on May 26, 2016.
29
96
a large fraction of private firm output will not be included since the final sample only relies
on publicly traded firms. Even within the sample of publicly traded firms, links are more
likely to be observed going from small suppliers to large customers.
Figure 22: NAICs Input-Output Representativeness
The plot below shows the sample representativeness compared to the BEA input-output data based on
NAICS industry codes, aggregated to the 15 NAICS industries defined by the BEA. Numbers in the tiles
are in percentages.
Representativeness of NAICs Input − Output for 2007
(0,5]
(5,10]
(10,20]
(20,50]
(50,80]
(80,100]
Wholesale Trade
Utilities
Transportation and Warehousing
Retail Trade
Professional Services
Other Services, Except Government
Mining
Manufacturing
Information
Government
Finance, Insurance, Real Estate
Educational Services, Healthcare, and Social Assistance
Construction
Arts, Entertainment, Recreation, Accommodation, Food Services
n,
tio
ea
cr
Ed
uc
at
Re
A
rts
,E
nt
er
ta
in
m
en
t,
A
gr
ic
ul
tu
re
,F
or
es
try
io
A
na
c
co , Fi
lS
s
m
er
m hin
vi
g,
od
ce
a
s,
tio and
H
n,
H
ea
Fo un
lth
od tin
ca
Se g
Fi re, a
na
C rv
n
nc d S on ices
s
e,
o
In cia truc
su
l A tio
ra
nc ssis n
e,
t
Re anc
e
al
Es
G
t
ov ate
er
n
In me
O
nt
fo
th
er
M rma
Se
a
t
n
i
rv
o
uf
ic
ac n
es
tu
,E
rin
xc
g
e
M
Pr pt G
in
i
of
es ove ng
Tr
sio rn
an
na me
sp
l S nt
or
ta
e
tio
Re rvic
n
es
ta
an
il
d
W Tra
ar
eh de
ou
sin
g
W
ho Util
iti
le
sa
es
le
Tr
ad
e
Agriculture, Forestry, Fishing, and Hunting
97
Table 13: Most Represented Industry to Industry Links
This table shows the top 5 most represented industry-to-industry captured by the final sample. The comparison is to the 2007 BEA input-output data based on NAICS industry codes, aggregated to the 15 NAICS
industries defined by the BEA.
% Represented
99.3
70.6
68.2
49.3
41.8
BEA Flows ($bn)
55.1
12.7
1.2
11.5
30.9
Supplier Industry
Wholesale Trade
Retail Trade
Agriculture, Forestry, Fishing, and Hunting
Retail Trade
Utilities
Customer Industry
Retail Trade
Manufacturing
Retail Trade
Finance, Insurance, Real Estate
Utilities
Despite the low representativeness of the final sample relative to economy-level inputoutput information gathered by the BEA, industry-level features of the network are preserved. The aggregated industry network structure looks relatively stable, and this aggregated input-output matrix shows that there seems to be high levels of cycles such as flows
from Manufacturing to Manufacturing, as shown in Figure 23. On the other hand, firm-level
networks show very little such cycles. Looking at industry-level networks rather than firmlevel networks masks important network features such as little reciprocity and transitivity,
key features used in the main model of this paper.
98
Figure 23: Compustat Network Aggregated Industry Input-Output
Industries are based on 2-digit SIC codes, defined as follows: 01 - 09 is Agriculture, Forestry, and Fishing;
10 - 14 is Mining; 15 - 17 is Construction; 20 - 39 is Manufacturing; 40 - 49 is Transportation & Public
Utilities; 50 - 51 is Wholesale Trade; 52 - 59 is Retail Trade; 60 - 67 is Finance, Insurance & Real Estate; 70
- 89 is Services; 91 - 98 is Public Administration. 99 is Unclassified, and is not included in the plot below.
2−Digit SIC Input−Output
in millions $
1
100
10,000
1995
2000
2005
2010
Wholesale Trade
Transportation & Public Utilities
Services
Mining
Manufacturing
Finance, Insurance, & Real Estate
Construction
Wholesale Trade
Transportation & Public Utilities
Services
Mining
Manufacturing
Finance, Insurance, & Real Estate
de
es
til
Tr
a
iti
vi
c
le
sa
ho
le
bl
Pu
&
tio
n
W
ic
U
Se
r
in
M
nc
Tr
a
ns
po
rta
su
ra
In
e,
nc
na
Fi
es
g
in
g
M
an
uf
ac
t
ur
sta
in
te
n
Re
e,
&
Co
es
,F
or
re
tu
ul
ic
gr
A
al
E
uc
ns
tr
Fi
try
,&
le
sa
ho
le
W
tio
n
rta
po
ns
tio
in
sh
Tr
a
til
U
ic
bl
Pu
&
g
de
es
iti
vi
c
Se
r
in
M
Tr
a
In
e,
nc
na
Fi
es
g
in
g
ur
M
in
te
sta
an
uf
ac
t
al
E
Re
su
ra
re
tu
ul
ic
gr
A
B.8
&
nc
,F
or
e,
es
Co
try
,&
ns
tr
Fi
uc
sh
in
tio
g
n
Construction
More Detailed Descriptive Statistics
Figure 24 shows measures of network structure through time, and corroborates the conclusion from the paper that the firm-level production network can be characterized by a
one-directional supply chain network. Of all years, less than 1% of all connections observed
contain loops. Most of the loops come from firms that report themselves as a major customer.
99
Figure 25 shows that from 1976 to 2012, a maximum of 10 firms a year report themselves as
a major customer.
Figure 24: Network Structure
The plot below shows some measures of the Compustat firm-level network structures through time. The
gray bars are NBER recession periods.
Network Structure
0.9
Reciprocity
0.3
0.0
0.9
Transitivity
Percentage (%)
0.6
0.6
0.3
0.0
1980
1990
2000
2010
Figure 25: Firm Sales to Itself
Number of Firms with Large Sales to Itself
10
8
6
4
2
0
1980
1990
2000
2010
Figure 26 shows that the Compustat firm-level production network is relatively stable
through time. On average, over 60% of the directed firm-to-firm connections are preserved,
with the higher number of connections preserved going from years 2010 to 2011, and the
lowest number of connections preserved going from year 1999 to 2000. Of the links that
are preserved, the dollar measures of sales from a supplier to customer also is very stable,
100
as measured by the cosine similarity measure. The cosine measure provides a method to
compare multiple cross sections through time, and is a cross-sectional based measure of
aggregate network similarity across years.
Figure 26: Similarity Measures
The plot below shows the similarity measures of the Compustat firm-level network across years. “Links”
shows the fraction of network links that remain from year t 1 to year t. The other three measures, “sales”,
“cust weight”, and “supp weight” focus only on the links that are preserved. Of the links that are preserved,
“sales” is the cosine similarity between total dollar sales between a supplier and customer, “cust weight” is
the cosine similarity between the fraction of a supplier firm’s sales that goes to the particular customer firm,
and “supp weight” is the cosine similarity between the fraction of a customer firm’s input coming from a
particular supplier. All three cosine similarity measures are compared between from year t 1 to year t. The
t ·yt 1
cosine similarity measure is defined as s = ||yyt ||||y
where s is a similarity measure and yt is the vector of
t 1 ||
interest in year t. Similarity measures are bounded between 0 and 1 (scaled to 100%), and a value of 100%
means that all values in year t 1 are the same as those in year t. A value of 0% means that the vectors
from t 1 and t are orthogonal. The gray bars are NBER recession periods.
Network Similarity Measures
Links
Dollar Sales
Customer Weights
Supplier Weights
100
75
50
25
%
0
100
75
50
25
0
1985 1990 1995 2000 2005 2010 2015
1985 1990 1995 2000 2005 2010 2015
Figure 27 shows the degree distributions for in and out-degrees. Over 25% of the firms
do not have an out-degree. Through the lens of the model, they are viewed as the final
consumption good producers. However, in reality they may still be intermediate goods
producers whose subsequent production chains are unobserved.
101
Figure 27: Degree Distributions
The plot below shows some measures of the Compustat firm-level sample degree distributions based on all
years of data from 1980 to 2012.
In−Degree Distribution
% of Observations
60
40
20
0
0
1
2
3
Number of Suppliers
4
5
>5
5
>5
Out−Degree Distribution
% of Observations
40
30
20
10
0
0
1
2
3
Number of Customers
4
Figure 28 shows the histogram and time series of the average short and long-term debt
usage in the sample. Figure 29 shows the distribution of log dollar changes in long term
debt, based on whether firms issue or reduce their debt. Long term debt may be reduced
either due to scheduled payments or due to active extinguishment of debt such as by calling
some outstanding corporate bonds.
102
Figure 28: Book Leverage Distribution
The plot below shows a histogram of the firm book-leverage distribution in the final sample. Long term
leverage is defined to be the ratio of total outstanding long-term debt (DLTT) divided by the total assets of
the firm (AT).
Leverage Distribution
20
% of Observations
15
10
5
0
0.05
0.15
0.25
0.35
0.45
0.55
0.65
0.75
0.85
0.95
DLTT AT
Average Book Leverage
Long Term
Short Term
0.6
0.4
0.2
0.0
1985
1990
1995
2000
103
2005
2010
2015
Figure 29: Changes in Long Term Debt
The plot below shows a histogram of the firm book value of long term debt in the final sample, across all
years from 1980 to 2012. Changes in long term debt is defined separately for increases in long term debt
(DLTIS) and decreases (DLTR).
Changes in Long Term Debt
0.3
Issuances
0.2
0.1
0.0
0.3
Density
Reductions
0.2
0.1
0.0
0.3
Net Changes
0.2
0.1
0.0
0.1
B.8.1
($) millions
100.0
100,000.0
Industries
SIC codes were originally implemented in 1937, but were revised dramatically in 1987 by
the Office of Management and Budget (OMB). Therefore, I use the industry classifications
only from 1987 onward.
Figure 30 shows the median leverage by industries defined according to 2-digit SIC groups,
while Figure 31 shows the median leverage by industries defined according to Fama-French
industries. The correlations between book leverage and market leverage is between 0.88 and
0.89 depending on how the industries are defined.
104
Figure 30: Industry Leverage: 2-digit SIC
The plot below shows the mean leverage ratio across industries, based on Compustat data from 1980 to 2012.
Industries are based on 2-digit SIC codes, defined as follows: 01 - 09 is Agriculture, Forestry, and Fishing; 10
- 14 is Mining; 15 - 17 is Construction; 20 - 39 is Manufacturing; 40 - 49 is Transportation & Public Utilities;
50 - 51 is Wholesale Trade; 52 - 59 is Retail Trade; 60 - 67 is Finance, Insurance & Real Estate; 70 - 89 is
Services; 91 - 98 is Public Administration. 99 is Unclassified, and is not included in the plot below. Industry
correlation is the correlation by industry of book and market leverage measures. Firm-level correlation is
the correlation across firms of book and market leverage measures.
Median Leverage by Industry
Industry Correlation = 0.81
Firm−Level Correlation = 0.14
Book
Market
Wholesale Trade
●
Transportation & Public Utilities
●
Construction
●
●
Mining
●
●
Agriculture, Forestry, & Fishing
●
●
Manufacturing
●
Nonclassifiable
●
Services
●
●
●
●
●
0.2
●
0.3
0.4
105
0.5 0.2
0.3
0.4
0.5
Figure 31: Industry Leverage: Fama-French Definitions
The plot below shows the mean leverage ratio across industries, based on Compustat data from 1980 to
2012. Industries are based 4-digit SIC codes. The exactly definitions are available on Ken French’s data
library site. The industry definition information was accessed online on May 23, 2016. Industry correlation
is the correlation by industry of book and market leverage measures. Firm-level correlation is the correlation
across firms of book and market leverage measures.
Median Leverage by Industry
Industry Correlation = 0.85
Firm−Level Correlation = 0.14
Book
Shipping Containers
Wholesale
Fabricated Products
Textiles
Candy & Soda
Automobiles and Trucks
Personal Services
Entertainment
Utilities
Business Supplies
Rubber and Plastic Products
Transportation
Tobacco Products
Consumer Goods
Food Products
Coal
Restaurants, Hotels, Motels
Healthcare
Communication
Construction Materials
Retail
Construction
Beer & Liquor
Apparel
Steel Works Etc
Petroleum and Natural Gas
Chemicals
Electrical Equipment
Shipbuilding, Railroad Equipment
Recreation
Aircraft
Defense
Printing and Publishing
Machinery
Agriculture
Business Services
NonMetallicand
Electronic Equipment
Computers
Medical Equipment
Precious Metals
Pharmaceutical Products
Measuring and Control Equipment
Computer Software
Market
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.1
●
0.2
0.3
0.4
106
0.5
0.60.1
0.2
0.3
0.4
0.5
0.6
B.9
Institutional Details
Ever since 1976, with the introduction of IAS No. 14, firms with customers that represent
at least 10% of total reported sales must provide the name of the customer on their financial
statements. The business segment and major customer disclosures are typically reported in
the footnote of annual 10-K filings. Later in June 1997 the FASB issued SFAS No. 131
which superseded the existing customer reporting requirements and mandated additional
operating segment disclosures. Rule No. 131 went into effect for fiscal years beginning
after December 15, 1997. These major customers can be other firms or various government
entities. However, in addition to this, firms may also voluntarily report their customers and
suppliers using other means, such as through press releases and Form 8-K’s. Therefore, the
network data I use is an underestimate of the set of actual firm links.
To better describe the institutional setting, I provide excerpts from the relevant disclosure laws. I focus only on the relevant events for the disclosure of customers and business
segments:
1. SFAS No. 14 required disclosure of major customers accounting for 10 percent or more
of the revenue of an enterprise. Paragraph 39 of FASB Statement No. 14 (“Financial
Reporting for Segments of a Business Enterprise” reads:
If 10 percent or more of the revenue of an enterprise is derived from sales to any
single customer, that fact and the amount of revenue from each such customer
shall be disclosed. (For this purpose, a group of customers under common control
shall be regarded as a single customer.) Similarly, if 10 percent or more of the
revenue of an enterprise is derived from sales to domestic government agencies
in the aggregate or to foreign governments in the aggregate, that fact and the
amount of revenue shall be disclosed. The identity of the industry segment or
segments making the sales shall be disclosed. The disclosures required by this
paragraph shall be made even if the enterprise is not required by this Statement to
107
report information about operations in different industries or foreign operations.
1. Although the accounting standard does not require the disclosure of names of those
customers, Regulation S-K Item 101 in CFR requires the disclosure of those customer
names.
(a) Specifically, this is in Section 229.101 (Item 101) Description of business, part (c)
(vii) reads:
The dependence of the segment upon a single customer, or a few customers,
the loss of any one or more of which would have a material adverse effect on
the segment. The name of any customer and its relationship, if any, with the
registrant or its subsidiaries shall be disclosed if sales to the customer by one
or more segments are made in an aggregate amount equal to 10 percent or
more of the registrant’s consolidated revenues and the loss of such customer
would have a material adverse effect on the registrant and its subsidiaries
taken as a whole. The names of other customers may be included, unless in
the particular case the effect of including the names would be misleading.
For purposes of this paragraph, a group of customers under common control
or customers that are affiliates of each other shall be regarded as a single
customer.
(a) This included sales to domestic governmental agencies in the aggregate or to
foreign governments in the aggregate. Effective for fiscal years beginning after
December 15, 1979, SFAS No. 30 amended SFAS No. 14 to break down the
individual domestic government or foreign government whose revenues are 10
percent or more of the enterprise’s revenues. The amended paragraph reads:
An enterprise shall disclose information about the extent of the enterprise’s
reliance on its major customers. If 10 percent or more of the revenue of an
108
enterprise is derived from sales to any single customer, that fact and the
amount of revenue from each such customer shall be disclosed. For this purpose, a group of entities under common control shall be regarded as a single
customer, and the federal government, a state government, a local government (for example, a county or municipality), or a foreign government shall
each be considered as a single customer.* The identity of the customer need
not be disclosed, but the identity of the industry segment or segments making the sales shall be disclosed. The disclosures required by this paragraph
shall be made by an enterprise subject to this Statement [Statement No. 14]
even if the enterprise operates only in one industry or has no foreign operations. [*Footnote reads: If sales are concentrated in a particular department
or agency of government, disclosure of that fact and the amount of revenue
derived from each such source is encouraged.]
2. SFAS No. 131 amended SFAS No. 14 and superseded it in some ways but not others.
The paragraph concerning major customers reads:
Since the adoption of SFAS No. 14, GAAP has required disclosure of revenues
from major customers. 32 SFAS No. 131 now requires issuers to disclose the
amount of revenues from each external customer that amounts to 10 percent or
more of its revenue as well as the identity of the segment(s) reporting the revenues.
The accounting standards, however, have never required issuers to identify major
customers. On the other hand, Regulation S-K Item 101 historically requires
naming a major customer if sales to that customer equal 10 percent or more of
the issuer’s consolidated revenues and if the loss of the customer would have a
material adverse effect on the issuer and its subsidiaries. 33 Since we continue to
believe that the identity of major customers is material information to investors,
we propose to retain this Regulation S-K requirement.
109
Comment: The point of concern is whether the required reporting is only for
external customers, or whether that also includes sales from one segment to
another but within the same firm. The wording in SFAS No. 14 shows only
major customers, with no specification of whether they need to be internal or
external. There are still firms reporting themselves as major customers, but that
represents a very small fraction of the population of firm links observed.
110