To Main Proceedings Document

Analyzing Dynamic Systems: A

Comparison of System Dynamics and
Structural Equation Modeling
Peter Hovmand
School of Social Work
Michigan State University

Ralph Levine
Department of Psychology and
Department of Resource Development
Michigan State University
 Overview
• Background
• Approach
• Comparisons
• Conclusions
 Background
• System dynamics modeling (SDM)
• Structural equation modeling (SEM)
• Similarity of features that are modeled
• Tendency to see SDM and SEM as the same
• Demonstrating differences between SDM
and SEM
Do SDM and SEM generate
“equivalent” models?
θ1 θ2
System ? Structural
Dynamics Equation
Model (SDM) c Model (SEM)

a b

S
Observed
data
 Approach
‡ 'HYHORSDV\VWHPG\QDPLFVPRGHO
‡ 8VHWKHV\VWHPG\QDPLFVPRGHOWRJHQHUDWH
VLPXODWHGREVHUYHGGDWDZKLFKLPSOLHVE\GHILQLWLRQ
WKDWWKHV\VWHPG\QDPLFVPRGHOILWVWKHREVHUYHG
GDWD
‡ *HQHUDWHVWUXFWXUDOHTXDWLRQPRGHOVWKDWILWWKH
REVHUYHGGDWD
‡ 7HVWVWUXFWXUDOHTXDWLRQPRGHOILWZLWKREVHUYHGGDWD
ZKLFKLVWKHQDWHVWRIILWQHVVZLWKWKHV\VWHP
G\QDPLFVPRGHO
 “Fixes That Fail”
Causal Loop Diagram

L1
❈✒
Problem
❈✓
Fix

L2
D
EL
A
Y

❈✑
Side effects
 Path Diagram
ε4 ε5 ε6
ε7 ε8 ε9
y4 y5 y6
y7 y8 y9

ζ2
❈✒ β23
❈✓ ζ3
β32

β21 β13
❈✑
ζ1

y1 y2 y3

ε1 ε2 ε3
 Latent Model
System Dynamics Model Structural Equation Model
∂η1 η1 = β13η3 + ζ 1
= ξ 2η3 + η1,t = 0
∂t η 2 = β 21η1 + β 23η3 + ζ 2
∂η 2
= −η 2η3 + η1 + ξ1 η3 = β 32η2 + ζ 3
∂t
∂η3
= −1.1 ⋅ η3 + η 2 + η3,t =0
∂t
 Measurement Model
System Dynamics Model Structural Equation Model
y1 (t s ) = η1 (t s ) + ε1 y1 = η1 + ε1
y2 (t s ) = η1 (t s ) + ε 2 y2 = η1 + ε 2
y3 (t s ) = η1 (t s ) + ε 3 y3 = η1 + ε 3
y4 (t s ) = η 2 (t s ) + ε 4 y4 = η2 + ε 4
y5 (t s ) = η2 (t s ) + ε 5 y5 = η 2 + ε 5
y6 (t s ) = η 2 (t s ) + ε 6 y6 = η 2 + ε 6
y7 (t s ) = η3 (t s ) + ε 7 y 7 = η3 + ε 7
y8 (t s ) = η3 (t s ) + ε 8 y8 = η3 + ε 8
y9 (t s ) = η3 (t s ) + ε 9 y9 = η3 + ε 9
 SDM
4

2
Eta_1
Eta_2
Eta_3
1

0
0 1 2 3 4 5 6 7 8 9 10
Time
1% of simulated values

Eta 1

Eta 2
Eta 3
LISREL Model 1
y1

y2

ETA 1
y3

y4
ETA 2
y5

y6
ETA 3
y7

y8

y9
 Model 1 Chi-Square

100.000
90.000
80.000
70.000
Chi-Square

60.000
50.000
40.000
30.000
20.000
10.000
0.000
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
Time
 Model 1 RMSEA

1.000
0.900
0.800
0.700
0.600
RMSEA

0.500
0.400
0.300
0.200
0.100
0.000
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
Time
LISREL Model 2
y1

y2

ETA 1
y3

y4

ETA 2
y5

y6
ETA 3
y7

y8

y9
 Model 2 Chi-Square

59.000

49.000

39.000
Chi-Square

29.000

19.000

9.000

-1.000
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
Time
 Model 2 RMSEA

0.900

0.700
RMSEA

0.500

0.300

0.100

-0.100 0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0

Time
 Summary of results
3KDVH RI ORRS 7LPH LQWHUYDO 'RPLQDQW ORRSV 0RGHO  0RGHO 
GRPLQDQFH ILW ILW
P1 0.0 to 0.6 L1 and L2 No Yes
Transition 0.6 - No Yes
P2 0.6 to 2.4 L1 No Yes
Transition 2.4 - Yes No2,3
P3 2.4 to 3.7 L2 Yes Yes
Transition 3.7 - Yes No2,3
P4 3.7 to 6.3 L1 Yes Yes
Transition 6.3 - Yes Yes
P5 6.3 to 10.0 L1 and L2 Yes1 Yes
1
RMSEA increases up to about 0.045 and then decreases again.
2
Fails to converge.
3
Not admissible after 50 iterations.
 Conclusions
• The structural equation model that
corresponded to the system dynamics model
did not fit the data during the initial shifts in
loop dominance.
• The structural equation model that did fit
the data did not correspond to the system
dynamics model.