You are on page 1of 27

Four Innovative Applications

of Hierarchical Linear
Modeling (HLM)

Shenyang Guo, Ph.D.


University of North Carolina at Chapel Hill
sguo@email.unc.edu
Acknowledgment
Support for this research is provided by the
Discretionary Grants Program of Children’s
Bureau to Shenyang Guo (PI). The project
aims to develop innovative quantitative
methods for child welfare research.
Overview of HLM
 Why HLM? The need to study multilevel influences on an
outcome variable, and to run growth curve analysis.
 Central problem: intraclass correlation.
 Conceptually one may view HLM as running regression
model several times, or at 2 or 3 levels.
 Other names:multi-level analysis, mixed-effects model,
random-effects model, growth curve analysis, random-
coefficient regression model, covariance components model.
 The key idea is to estimate random effects. In addition to
traditional regression coefficients, HLM estimates a set of
random effects associated with each high-level unit, which
can be used to control for autocorrelation.
Four innovative models
1. Latent-variable analysis of HLM
2. Omnibus score of CBCL & TRF using
latent-variable HLM
3. Meta analysis using HLM
4. Modeling multivariate change
Latent-variable analysis (1)
 Latent variables: variables that are not directly
observed. Under this framework, any observed
variable is an indicator, and can be viewed as a
latent true-score plus measurement error.
 Statistical models for analyzing latent variables:
structural equation modeling: (1) measurement
model – relations between indicator and latent
variable; (2) structural model – relations among
latent variables.
 In HLM, a latent-variable analysis consists of two
parts: measurement model, and structural model
involving explanatory variables.
Latent-variable analysis (2)
Example
 Sampson, Raudenbush, & Earls (1997, Science 277(15):
918-924) applied this approach to analyzing multilevel
influences of collective efficacy, in which they view
collective efficacy as a latent variable. Their three-level
HLM treats ten items collected from all survey
respondents as level 1, and conceptualizes that these
items are commonly determined by a latent true score
“collective efficacy” plus measurement errors. Their
model then explores how informants within neighborhoods
(i.e., level 2) vary randomly around the neighborhood
mean of “collective efficacy”, and how neighborhoods
across whole study area (i.e., level 3) vary randomly about
the grand mean of “collective efficacy”.
Omnibus score of CBCL & TRF (1)
 Disentangle multiple raters’ measurement error from
clients true change (Guo & Hussey, 1999, Social Work
Research 23(4): 258-269).
 Ratings are likely to be collected by multiple raters (e.g.,
Achenbach instruments: CBCL, TRF, & YSR).
 Attritions can also occur in raters.
 None of the prior studies (before 1999) ever controlled
for raters’ impact on ratings, though many used multiple
raters to collect ratings.
 A theoretical framework to investigate multiple sources
of measurement error: Cronbach’s Generalizability
Theory.
Omnibus score of CBCL & TRF (2)

The Need: Hypothetical Data


Two raters’ ratings on a single subject
1a 1b 1c 1d
Y Y Y Y
70 70 70 70
60 60 60 60
50 50 50 50
40 40 40 40
30 30 30 30
20 20 20 20
10 10 10 10
0 0 0 0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Time Time Rater A Time Time
Rater B
Omnibus score of CBCL & TRF (3)
Problems & solutions
Problem Type of Causes of the Problem Solution Major Results

Data

Divergent ratings Both cross- Three versions of rating A three-level  Both models estimate a “true
made by caregiver, sectional & forms: CBCL, TRF, & HLM with latent- score” (an omnibus score based
teacher, and youth longitudinal YSR variable analysis; on multiple ratings) for each
self or study child.
Guo & Hussey’s  Both facilitate a multivariate
three-level HLM analysis identifying significant
predictors of true-score
differences in the study sample.
Attrition of raters or Longitudinal A teacher may change job Guo & Hussey’s The model estimates a “true
missing ratings in a longitudinal study and three-level HLM change trajectory” (an omnibus
make herself no longer a trajectory based on all available
member of the TRF ratings) for each study child;
collection team. The model facilitates an inter-
A caregiver may miss one individual analysis that identifies
or more CBCL collections. significant predictors of the overall
A youth may miss one or trajectory of the study sample.
more YSR collections.
Omnibus score of CBCL & TRF (4)

L e v e l 1 :
R ijk   jk   D ij k   ijk

L e v e l 2 :
 jk   0 k   01 k ( RATER _ X 1 ) jk   02 k ( RATER _ X 2 ) jk  r 0 jk

L e v e l 3 :
 0 k   00   01 ( CHILD _ W 1 ) k   02 ( CHILD _ W 2 ) k  u 0 k

 0 0 k is th e o m n ib u s s c o re
Omnibus score of CBCL & TRF (5)
Model 1

L evel 1: R ijk   jk  D ijk   ijk

L evel 2:  jk   00 k  r 0 jk
L evel 3:  00 k   000  u 00 k

 00k is th e o m n ib u s s c o re
Omnibus score of CBCL & TRF (6)
Model 2
L ev el 1 :
R i jk   jk   D ij k   i jk

L ev el 2 :
 jk   0 k  r0 jk

L ev el 3 :
 0 k   00   01 ( CHILD _ W 1)k.   0 q ( CHILD _ Wq )k  u 0 k

 0 0 k is th e o m n ib u s s c o re
Omnibus score of CBCL & TRF (6)
Illustrating example
 Acknowledgment to Dr. Richard Barth and Ms. Ariana
Wall at UNC for their help.
 Data: National Survey of Child and Adolescent Well-
being (NSCAW).
 We focus on externalizing and internalizing scores.
Each child has four such scores: two from caregiver
(CBCL), and two from teacher (TRF). The task: how
to create one score?
 Variables employed in level 3 of Model 2: age gender,
race, social behavior, MBA reading score, MBA math
score, count of risky behaviors of delinquency, count
of risky behaviors of substance abuse, and count of
risky behavior of suicidal attempt.
Omnibus score of CBCL & TRF (7)

Correlation coefficients and descriptive


statistics on disagreement between
caregiver and teacher’s scores (N=448)
____________________________________________________________________________
Ec Et Ic It
_______ _______ _______ _______
Externalizing rated by caregiver (Ec) 1.000
Externalizing rated by teacher (Et) .405** 1.000
Internalizing rated by caregiver (I c) .639** .142** 1.000
Internalizing rated by teacher (I t) .174** .476** .196** 1.000

Mean (S.D.) 60.8 (11.52) 58.7 (9.60) 57.3 (11.91) 56.0 (9.69)
____________________________________________________________________________
** p < .01
Omnibus score of CBCL & TRF (8)
Evaluation Schemes:
C1 Caregiver's scores only
.5Ec + .5Ic
C2 Teacher's scores only
.5Et + .5It
C3 All 4 scores from both versions with equal weights
.25Ec + .25Ic + .25Et + .25It
C4 Similar to C3 but heavier weights giving to caregiver's sco
.35Ec + .35Ic + .15Et + .15It
C5 Similar to C3 but heavier weights giving to teacher's score
.15Ec + .15Ic + .35Et + .35It
C6 Similar to C3, a 50/50 split between caregiver and teacher'
scores but heavier weights giving to externalizing scores
.35Ec + .15Ic + .35Et + .15It
Omnibus score of CBCL & TRF (9)
Evaluation Schemes (continued):
C7 Similar to C3, a 50/50 split between caregiver and
teacher's scores but heavier weights giving to internalizing
scores
.15Ec + .35Ic + .15Et + .35It
C8 Extreme value, low end
.5 [Min (Ec,Et)] + .5 [Min (Ic,It)]
C9 Extreme value, high end
.5 [Max (Ec,Et)] + .5 [Max (Ic,It)]
C10 Arbitrary: half high externalizing and half low
internalizing
1.5 [Max (Ec,Et)] + .5 [Min (Ic,It)]
C11 Arbitrary: half low externalizing and half high
internalizing
.5 [Min (Ec,Et)] + .5 [Max (Ic,It)]
Omnibus score of CBCL & TRF (10)
Evaluations
_____________________________________________________________________________

Scheme Mean S.D. Minimum Maximum Correlation Coefficient


_____________________________
with Omnibus 1 w ith Omnibus 2
_____________________________________________________________ ______________

Omnibus 1 60.56 3.29 52.30 69.21


Omnibus 2 60.56 4.78 49.34 72.81 .707
C1 59.03 10.61 32.00 82.50 .855 .700
C2 57.32 8.29 39.50 80.00 .747 .406
C3 58.18 7.63 39.00 78.25 1.000 .707
C4 58.52 8.49 37.00 79.35 .966 .731
C5 57.84 7.39 39.40 77.15 .955 .620
C6 58.79 7.90 39.30 77.60 .979 .702
C7 57.57 7.69 37.20 79.35 .978 .681
C8 53.00 8.22 32.00 74.00 .929 .656
C9 63.36 8.21 41.50 82.50 .929 .657
C10 57.74 8.09 37.50 76.50 .961 .698
C11 58.62 7.81 39.00 81.00 .958 .658

_____________________________________________________________________________
All correlation coefficients are statistically significant (p<.01)
Omnibus score of CBCL & TRF (11)
Use the score as a dependent variable
______________________________________________________________________

Scheme Employed the score created


by the scheme
as an outcome variable

R2
_______________________________________

Omnibus 1 .388
Omnibus 2 .987
C1 .431
C2 .147
C3 .388
C4 .434
C5 .293
C6 .388
C7 .361
C8 .337
C9 .337
C10 .383
C11 .332

______________________________________________________________________
Omnibus score of CBCL & TRF (12)
Use the score as independent variable:
___________________________________________________________________________________________

Scheme Employed the score as an independent variable


_____________________________________________________________________
DV=Substance Abuse Risk DV=Delinquency Risk DV=Suicidal Risk
________________________ ________________________ ________________________
P-value for R2 P-value for R2 P-value for R2
B Incremental B Incremental B Incremental
________________________________ _____________________ _____________________

Omnibus 1 .005 0.012 .894 .000 .000 0.080


Omnibus 2 .000 0.043 .797 .000 .000 0.244
C1 .061 0.005 .695 .000 .000 0.076
C2 .008 0.010 .793 .000 .000 0.026
C3 .005 0.012 .894 .000 .000 0.080
C4 .014 0.009 .787 .000 .000 0.085
C5 .003 0.013 .969 .000 .000 0.062
C6 .001 0.017 .635 .000 .000 0.067
C7 .033 0.007 .825 .000 .000 0.087
C8 .015 0.009 .505 .000 .000 0.067
C9 .007 0.011 .678 .000 .000 0.070
C10 .004 0.013 .883 .000 .000 0.078
C11 .015 0.009 .916 .000 .000 0.069

___________________________________________________________________________________________
Meta analysis using HLM (1)
 R & B (2002): Chapter 7
 Meta analysis: research synthesis, or a “study
of the studies”. Objective: summarize results
from a series of related studies.
 Collect the following data from literature review:
the mean outcome for the experimental group;
Y Ej the mean outcome for the control group;
Y Cj the pooled, within-group standard deviation;
Sj the sample size of the experimental group;
nEj
the sample size of the control group;
where j indicates the jth study.
nCj
Meta analysis using HLM (2)
Based on these data, calculate effect size:
d j  (Y Ej  Y Cj ) / S j

And variance of the effect size:


V j  (nEj  nCj ) /( nEj nCj )  d 2 j /[ 2(nEj  nCj )]
Square root of Vj is called “standard error of dj”
Meta analysis using HLM (3)
 General model
Level 1: dj = j + ej
Level 2: j = 0 + uj
or combined model: dj = 0 + uj + ej
where dj ~N(0, j) with j =  + Vj
 In this model, we only have one subscript j to
indicate study. This is a special case of two-level
model, in which subscript i is omitted, because we
don’t have original data at the study subject level.
 V-known model: unlike previous HLM, this model
has known variance Vj,or S.E.(dj)=  Vj.
Meta analysis using HLM (4)
 Use HLM DOS version to run the v-
known model.
 Data look like this:
1 .030 .016 2.000
2 .120 .022 3.000
3 -.140 .028 3.000
……
19 -.070 .030 3.000

 Format of the raw data file: (a11,3f11.3)


 See HLM 5 manual pp. 221-226.
Meta analysis using HLM (5)
Experimental Studies of Teacher
Expectancy Effects on Pupil IQ
Effect
Size Standard Weeks of
Estimate Error of Prior
Study dj dj Contact

1. Rosenthal et al. (1974) 0.030 0.126 2.000


2. Conn et al. (1968) 0.120 0.148 3.000
3. Jose & Cody (1971) -0.140 0.167 3.000
4. Pellegrini & Hicks (1972) 1.180 0.373 0.000
5. Pellegrini & Hicks (1972) 0.260 0.369 0.000
6. Evans & Rosenthal (1969) -0.060 0.105 3.000
7. Fielder et al. (1971) -0.020 0.105 3.000
8. Claiborn (1969) -0.320 0.219 3.000
9. Kester & Letchworth (1972) 0.270 0.164 0.000
10. Maxwell (1970) 0.800 0.251 1.000
11. Carter (1970) 0.540 0.302 0.000
12. Flowers (1966) 0.180 0.224 0.000
13. Keshock (1970) -0.020 0.290 1.000
14. Henrickson (1970) 0.230 0.290 2.000
15. Fine (1972) -0.180 0.158 3.000
16 Greiger (1970) -0.060 0.167 3.000
17. Rosenthal & Jacobson (1968) 0.300 0.138 1.000
18. Fleming & Anttonen (1971) 0.070 0.095 2.000
19. Ginsburg (1970) -0.070 0.173 3.000
Meta analysis using HLM (6)
Running HLM, we obtain the following findings:
The estimated grand-mean effect size is
0.084, implying that, on average, experimental
students scored about .084 standard deviation
units above the controls.
However, the estimated variance of the
effect parameter is =.019. This corresponds
to a standard deviation of .138 (i.e., .019 = .
138), which implies that important variability
exists in the true-effect sizes. For example,
an effect one standard deviation above the
average would be .084+.138=.222, which is of
nontrivial magnitude.
Multivariate change (1)
 In a cross-sectional study, we use
correlation coefficients to see the level of
association of an outcome variable with other
variables.
 In a longitudinal study, we have a similar

task, that is, we need to model multivariate


change: whether two change trajectories
(outcome measures) correlate over time?
For details of this method, see MacCallum,

R.C., & Kim, C. (2000). “Modeling


multivariate change”, in Little, Schnabel, &
Baumert edited, Modeling Longitudinal and
Multilevel Data. Lawrence Erlbaum
Associates, pp.51-68.
Multivariate change (2)
What kind of questions can be
answered?
Whether benefits clients gained from an
intervention over time negatively correlate
with the intervention’s side effects?
Whether clients’ change in physical health

correlates with their change in mental health?

Whether a program’s designed change in


outcome (e.g., abstinence from alcohol or
substance abuse) correlates with clients’ level
of depression?
 Use software MLn/MLwiN to estimate
the model. It’s possible to use SAS Proc

You might also like