(Multi-level Modeling)

!
(hierarchical linear model),
(random coefficient model)
(slope-as-outcome model)

(, ), (individual growth
curve model) (individual change model)

:
HLM,
HLM MLwiN,
ML iN AMOS,
AMOS LISREL MPlus
MPl

 (OLS)

',' '' ''

. ,
.

) = + ( )

(OLS / )
(error term) ,

(error terms) .
OLS: R-square, R = R() + R( )

(iterative convergence)
.
OLS / ,
.

90

80

70

60

50

40

30

100

95

90

85

80

75

70

89

88

78

68

58

48

38

65

60

55

50

45

40

35

rAB=.336
= 336 (p=n.s.,
=n s n=14)
.
A rA=1.00
1 00 (p<.001,
< 001 n=7),
7)
B rB=.991 (p<.001, n=7)
.
.



,
1.
2
2.
3.

<>
1. :

2. :
(, cluster)

(Growth Curve Model)

Purpose is to model change over time

Li
Linear
or nonlinear
li
models
d l possible
ibl

Variability in change over time by modeling individual growth

curves

Variability in initial or average levels

Predictors can be used to account for variability

Two general approaches

Hierarchical linear models (HLM)
Structural equation models (SEM)

-- (serially)
(auto-regression analysis) (individual
growth curve model) .
--
.
(B ll & Curran,
(Bollen
Cu
2004)
2004).
-- (time-lag) (:
1st lag,
1st.
lag 2nd.
2nd lag ) .

--
.

-- .
(t-1)
(t
1)
,
.

(Growth Curve Model)

?
availability of longitudinal data

emphasis
h i on iindividual
di id l diff
differences
accessibility of SEM software

HLM Approach to Growth Curves

Conceptualization
Two levels: within individual and between
i di id l
individual
Regression equation for each level

Example Growth Curves

High Variability in Intercepts and Slopes

Low Variability in Intercepts and High

Variability in Slopes

t
Low Variability in Intercepts and Slopes

Y
t

Y
t

HLM Approach to Growth Curves

Level 1: Time Level

yti = 0i + 1i xti + rti

Yti = 0i + 1ixti + rti

Examines change in the dependent variable as a function

of time for each individual
Intercepts and slopes obtained for each individual
Intercept is initial or average value of the dependent
variable for a given individual (depending on coding of
time variable)
Slope describes linear increase or decrease in the
dependent variable over time of a given individual
With predictors, intercepts and slopes represent adjusted
means and
d slopes
l

HLM Approach to Growth Curves

Level 2: Individual Level
0j = 00 + 01z1i + U0j
1j = 10 + 11z1i + U1j
Intercepts and slopes obtained from Level 1 serve as dependent
variables
With no predictors, Level 2 intercept represents average of
intercepts or slopes from Level 1
With no predictors, Level 2 residual provides information about
variance of intercepts or slopes across individuals
Can incorporate predictors measured at the individual level
(gender, income, etc.)
Predictors explain variation in intercepts or slopes across
individuals

SEM Approach to Growth Curves

Example of a latent growth curve model with four time points

yt1

yt2

1 1

0
(Intercept)

yt3

1
0

yt4

1
(Slope)

:
:
-- (KYPS)
4 1, 2, 3 (2004, 2005, 2006)

-- 2004 1
(stratified multistage cluster sampling) 2,949
2004 1 .

:
-------

ID/
,
/
/ /
/
.

1 :
1.

1)
-- //
-- (E-mail)
(E
il)
-- /
--
2)
-- ,
-- ,
2. : ,

RMSEA=.08

GFI= .86

(Embedded Model)

30

, ,

60

50

( )

40

( )
( )

30

( )
20

( )
( )

10

0
3 16

4 2

4 10

4 15

45

40

( )
35

( )
30

( )
( )

25

20
3 16

4 2

4 10

4 15

2030

55
50
45

20 ( )

40
35

20 ( )

30

30 ( )
25

30 ( )

20
15
10
3 16

4 2

4 10

4 15

40 50
4050

45

40

40 ( )

35

40 ( )
30

50 ( )
50 ( )

25

20
3 16

4 2

4 10

4 15

HLM ( 3)

HLM ( 2)

3 16

19.3

19.8

44.9

47.1

4 2

24 1
24.1

22 6
22.6

42 3
42.3

43 1
43.1

4 10

26.7

25.7

39.7

41.4

4 15

29 0
29.0

29 2
29.2

37 2
37.2

38 7
38.7

4 23
()

31.3

34.7

35.7

36.0

4 23 27.9%,
35.3%.

52

47

42

(HLM)
37

(Logistic)

32

(HLM)

27

22

(Logistic)

17
3 16

4 2

4 10

4 15

423
( )

, ?
.
.
, (error)
(
)

.

.
.

.
(unbalanced design)
,
.
().
, .
.
OLS
.