Structural equation
modeling
Advanced topics
Rex B Kline
Concordia
QICSS Set D
D2
Moderation + mediation
Topics
o Moderation (MR, PA)
o Conditional process modeling
o Causal mediation analysis
D3
MMR
o X, W, Y are continuous
o Y = BX X + BWW + BXW XW + A
o Interpretation
D4
Myths
o
Edwards, J. R. (2009). Seven deadly myths of
testing moderation in organizational
research. In C. E. Lance & R. J.
Vandenberg (Eds), Statistical and
methodological myths and urban
legends: Doctrine, verity and fable in the
organizational and social sciences (pp.
143164). New York: Taylor & Francis.
D5
Myth
You must center, to reduce
extreme collinearity
D6
Truth
Centering changes nothing
Optional, if 0 is not on scale
Center some, others not
D7
Myth
You must use hierarchical
entry
D8
Truth
Not required
Possibly misleading
D9
Myth
You can ignore score
reliability
D10
Truth
Score reliability is critical
rXX > .90
D11
Myth
Y = BX X + BWW + BXW XW + A
X, W are main effects
D12
Truth
X, W are linear only
D13
Myth
You can ignore curvilinear
effects
D14
Truth
Estimate X2 and W2, too
D15
2
6
8
11
4
7
8
11
7.125
10
12
13
10
24
19
18
25
16.375
5
9
11
11
11
10
7
5
D16
Y = .112 X .064 W + 8.873
2
R = .033
D17
2
10
6
12
8
13
11
10
4
24
7
19
8
18
11
25
7.125 16.375
5.125
1.125
.875
3.875
3.125
.125
.875
3.875
0
6.375
4.375
3.375
6.375
7.625
2.625
1.625
8.625
0
5
9
11
11
11
10
7
5
D18
Y = .112 X .064 W + 8.873
Y = .112 x .064 w + 8.625
2
R = .033
D19
11
10
Y
7
4
1
10
11
D20
11
10
W < MW
8
W > MW
Y
7
4
1
10
11
D21
Products
o XW with X, W
o xw with x, w
o XWres with X, W
D22
Products
o XWres:
1. Regress XW on X, W
2. Create XW
3. Create XWres = XW XW
D23
X
XW
XWres
.747 .706
0
xw .284 .113
D24
Products
o No difference:
BXW = Bxw = BXWres
R2
D25
Y = .112 X .064 W + 8.873
2
R = .033
Y = 1.768 X + .734 W .108 XW 3.118
2
R = .829
D26
Y = 1.768 X + .734 W .108 XW 3.118
20
15
10
5
0
30
25
20
15
14
12
10
10
2
D27
Y = .112 x .064 w + 8.625
2
R = .033
Y = .000 x .035 w .108 xw + 8.903
2
R = .829
D28
Y = .112 X .064 W + 8.873
2
R = .033
Y = .112 X .064 W .108 XWres + 8.873
2
R = .829
D29
Simple regressions
o Simple slopes
o Simple intercepts
o Generate equations
D30
Y on X as a function of W
Y = 1.768 X + .734 W .108 XW 3.118
Y = 1.768 X .108 XW + .734 W 3.118
Y = (1.768 .108 W) X + .(734 W 3.118)
D31
Y = (1.768 .108 W) X + (.734 W 3.118)
MW = 16.38
4.34 10.36 16.38 22.40 28.42
D32
Y = (1.768 .108 W) X + .(734 W 3.118)
4.34, 10.36, 16.38, 22.40, and 28.42
YW =22.40 = (1.768 .108 * 22.40) X + (.734 * 22.40 3.118)
YW =22.40 = .651X + 13.324
D33
W
Level
Score
Regression equation
+2 SD
28.42
Y = 1.301 X + 17.712
+1 SD
22.40
Y = .651 X + 13.324
Mean
16.38
Y = .001 X + 8.905
1 SD
10.36
Y = .649 X + 4.486
2 SD
4.34
Y = 1.299 X + .068
D34
11
2 SDW
SDW
10
MW
9
8
Y
+SDW
7
6
+2 SDW
5
4
1
10
11
D35
[Link]
2 SDW
1 SDW
MW
+1 SDW
+2 SDW
D36
(a) Regression perspective
(b) Compact symbolism
XW
XW
XW
W
W
(c) X as focal variable,
W as moderator
(d) W as focal variable,
X as moderator
W
XW
X
X
D
XW
D37
Mediation
o Time precedence:
Experimental
Longitudinal
o No? Indirect effect
D38
Mediation
o
Kline, R. B. (in press). The mediation
myth. Basic and Applied Social
Psychology.
Little, T. D. (2013). Longitudinal structural
equation modeling. New York:
Guilford.
D39
CPM
o
Hayes, A. F. (2013a). Conditional process modeling:
Using structural equation modeling to examine
contingent causal processes. In G. R. Hancock &
R. O. Mueller (Eds.), Structural equation modeling:
A second course (2nd ed.) (pp. 219266).
Greenwich, CT: IAP.
Hayes, A. F. (2013b). Introduction to mediation,
moderation, and process control analysis: A
regression-based approach. New York: Guilford.
D40
CPM
o Mediated moderation
o Moderated mediation
o Cause mediator
D41
Mediated moderation
X
1
1
DY
DM
Y
M
Y
W
D42
Example
o
Lance, C. E. (1988). Residual centering,
exploratory and confirmatory moderator
analysis, and decomposition of effects in
path models containing interaction
effects. Applied Psychological
Measurement, 12, 163175.
D43
D44
Moderated mediation (1)
1st-stage moderation
XMY
X M depends on W
X
1
1
DY
DM
Y
M
Y
W
D45
Moderated mediation (2)
1st-stage moderation
WMY
W M depends on X
X
1
1
DY
DM
Y
M
Y
W
D46
Moderated mediation
2nd-stage moderation
XMY
M Y depends on W
X
DM 1
M
Y
DY
D47
Moderated mediation
1st-and-2nd-stage moderation
X M, M Y depend on W
W M, M Y depend on X
1
DM
M
Y
1
DY
D48
Example
o
Curran, T., Hill, A. P., & Niemiec, C. P.
(2013). A conditional process model
of children's behavioral engagement
and behavioral disaffection in sport
based on self-determination theory.
Journal of Sport & Exercise
Psychology, 35, 3043.
D49
D50
Example
o
Desrosiers, A., Vine, V., Curtiss, J., &
Klemanski, D. H. (2014). Observing
nonreactively: A conditional process
model linking mindfulness facets,
cognitive emotion regulation
strategies, and depression and
anxiety symptoms. Journal of
Affective Disorders, 165, 3137.
D51
D52
BaronKenny
o Continuous variables
o Linear model
o No interaction
D53
Product estimator = ab
DM
1
DY
1
D54
Cause mediator
X
DY
Y
DM 1
M
Y
=BX+A
M
1
1
Y = B2 X + B3 M + B4 XM + A2
D55
X M in CPM
o No single direct effect
o Effect decomposition?
o Nonlinear models?
D56
Causal mediation
o Judea Pearl
o Graph theory
o Structural causal model
D57
Causal mediation
o SCM:
Parametric SEM
Nonparametric models
Potential outcomes
D58
Causal mediation
o SCM:
Assumes X M
Linear or nonlinear
Counterfactuals
D59
Counterfactuals
o Experimental design, Tx, Cn
o What if Tx were not treated?
o What if Cn were treated?
D60
Counterfactuals
o Rubin Causal Model
o Inference with missing data
o Latent variables
D61
Counterfactuals
o
Rubin, D. B. (2005). Causal inference
using potential outcomes: Design,
modeling, decisions. Journal of the
American Statistical Association, 100,
322331.
D62
Mediation in SCM
o Design:
Experimental design
X = 0, 1
M, Y are continuous
D63
Mediation in SCM
o Direct effects:
Controlled (CDE)
Natural (NDE)
D64
Mediation in SCM
o No interaction:
CDE = NDE
D65
Mediation in SCM
o CDE:
How much Y changes
Given X = 0 to X = 1
If M = m for all cases
D66
Mediation in SCM
o CDE:
Estimated for M
Policy goal: Lift all to m
D67
Mediation in SCM
o NDE:
How much Y changes
Given X = 0 to X = 1
If M varies as under X = 0
D68
Mediation in SCM
o Natural indirect effect (NIE):
How much Y changes in X = 1
As M changes from under
X = 0 to X = 1
D69
Mediation in SCM
o Total effect:
TE = NDE + NIE
D70
Counterfactuals
CDE = E [ Y (X = 1, M = m) ] E [ Y (X = 0, M = m) ]
NDE = E [ Y (X = 1, M = m0) ] E [ Y (X = 0, M = m0) ]
NIE = E [ Y (X = 1, M = m1) ] E [ Y (X = 1, M = m0) ]
TE = E [ Y (X = 1) ] E [ Y (X = 0) ]
D71
Example
o
Petersen, M. L., Sinisi, S. E., & van der
Laan, M. J. (2006). Estimation of direct
causal effects. Epidemiology, 17,
276284.
D72
X = 0, control; X = 1, anti-viral therapy
M = viral load
Y = CD4 T-cells
D73
CDE
Mean T-cells if viral load were
the same for all cases
D74
NDE
Mean T-cells if viral load were as
among untreated cases
D75
NIE
Mean T-cells among treated if
viral load changed from untreated
to treated levels
D76
= + X
M
0
1
Y = 0 + 1X + 2 M + 3 XM
CDE = 1 + 3 m
NDE = 1 + 3 0
NIE = (2 + 3)1
D77
= + X
M
0
1
Y = 0 + 1X + 2 M + 3 XM
If 3 = 0:
CDE = 1
NDE = 1
NIE = 2 1
D78
= 1.70 .20 X
M
Y = 450.00 + 50.00 X 20.00 M 10.00 XM
0 = 1.70 and 1 = .20
0 = 450.00, 1 = 50.00, 2 = 20.00,
and 3 = 10.00
D79
CDE = 50.00 10.00 m
NDE = 50.00 10.00 (1.70) = 33.00
NIE = (20.00 10.00) .20 = 6.00
TE = 33.00 + 6.00 = 39.00
D80
Resources
o
Valeri, L., & VanderWeele, T. J. (2013).
Mediation analysis allowing for
exposuremediator interactions and
causal interpretation: Theoretical
assumptions and implementation
with SAS and SPSS macros.
Psychological Methods, 2, 137150.
D81
Resources
o
Imai, K., Keele, L., & Tingley, D. (2010). A
general approach to causal
mediation analysis. Psychological
Methods, 15, 309334.
D82
Resources
o
Hicks, R., & Tingley, D. (2011). Causal
mediation analysis. Stata Journal, 11,
605619.
[Link]
D83
Resources
o
Muthn, B., & Asparouhov, T. (2015).
Causal effects in mediation
modeling: An introduction with
applications to latent variables.
Structural Equation Modeling, 22, 12
23.
D84
D85
D86
