The Importance of Generative Models for Assessing Network Structure
Matthew J. Denny | Penn State | Contact: [email protected], www.mjdenny.com | Paper: ssrn.com/abstract=2798493
Application: International Lending
Project Overview
• Network structure can only be assessed using generative models.
• Important to consider explanations for observed structure at three levels:
2005
2000Settlements (BIS) aggregate (private
Bank for International
and public) international lending volumes for 16-24 large economies from Oatley et al. (2013).
node, group, and system. Explanations confound eachother.
• Multicollinearity in (G)ERGM statistics creates specification challenges.
• Bookending problem: multicollinearity and omitted variable bias. Need
for strong theory to guide specification.
• Application of framework to extend Oatley et al. (2013). Generative
modelling results do not agree with descriptive analysis.
2000
2005
●
TAW
PAN
●
BEL
●
AUT
●
●
ITA
●
FIN
SWE
●
Specification and Testing of Relational Theories
ITA
●
TUR
●
GRE
●
BRA
●
POR
●
●
AUT
●
AUS
CHL
●
●
ESP
ESP
●
SWE
●
CAN
●
FRA
●
●
SWIUK
DEU
●
●
DEU
●
DEN
●
●
CAN
US
2008
●
UK
●
JAP
●
SWI
US
●
NET
●
●
FRA
JAP
2010
●
TAW
●
GRE
●
MEX
BRA
●
●
POR
●
●
AUS
●
DEN
2010
●
BRA
●
IRE
●
BEL
●
POR
●
NET
2008
●
DEN
●
MEX
●
●
IRE
●
CHL
BEL
●
●
TUR
●
GRE
●
BEL
●
SWE
●
PAN
●
AUT
●
DEN
●
MEX
JAP
●
TAW
●
●
CAN
●
●
ITA
1.00
0.95
0.90
0.85
0.80
●
●
●
●
0.98
0
1
●
●
G8 Sender, G8 Receiver
●
●
G8 Sender, Non−G8 Receiver
●
●
log(GDP) Receiver
●
●
log(GDP) Sender
●
●
Non−G8 Sender, G8 Receiver
●
●
Nodes in Network
Oatley, T., Winecoff, W. K., Pennock, A., Danzman, S. B. (2013).
“The Political Economy of Global Finance: A Network Model."
Perspectives on Politics, 11(01), 133–153.
3
●
Transitive Triads
Normalized Net Exports
2
●
Out 2−Stars
●
Normalized Net Exports
−1
Mutual Dyads
log(GDP) Sender
Non−G8 Sender, G8 Receiver
4
●
●
Variable
G8 Sender, Non−G8 Receiver
2010
Coefficient
Coefficient
2
●
0.94
●
●
●
Variable
0.90
●
●
●
●
Normalized Net Exports
log(GDP) Receiver
50
●
●
Non−G8 Sender, G8 Receiver
Transitive Triads
G8 Sender, G8 Receiver
4
●
●
●
log(GDP) Sender
0
40
Mutual Dyads
2008
Out 2−Stars
30
2
log(GDP) Receiver
Mutual Dyads
20
0
G8 Sender, Non−G8 Receiver
log(GDP) Receiver
Nodes in Network
10
−2
G8 Sender, G8 Receiver
50
out2star
in2star
ctriads
recip
edgeweight
6
Transitive Triads
Variable
●
G8 Sender, G8 Receiver
G8 Sender, Non−G8 Receiver
2005
Coefficient
●
Variable
40
●
POR
●
PAN
Out 2−Stars
Transitive Triads
Non−G8 Sender, G8 Receiver
30
●
FRA
●
ITA
●
ESP
Coefficient
2
4
log(GDP) Sender
20
US
●
AUS
●
●
TUR
0
Out 2−Stars
10
●
NET
●
●
DEU
UK
●
AUT
Mutual Dyads
out2star
in2star
ctriads
recip
edgeweight
●
CAN
2000
Normalized Net Exports
Correlations
●
SWE
GERGM Estimation Results
Figure: Plots of correlation coefficients between five common network statistics and the transitive
triads statistic for samples of 2,000 random n-node networks with 10 to 50 nodes for dichotomous
networks (top) and continuous networks (bottom). Correlations between common network statistics
are higher in continuous networks, and the correlations are generally increasing in network size.
Correlations
●
CHL
ESP
US ●
UK
●
SWI
●
DEU
●
FRA
●
Multicollinearity in (G)ERGM Statistics
●
SWI
●
JAP
●
NET
This work was supported under NSF
grants: DGE-1144860, SES-1357606,
and CISE-1320219.
4