MULTILEVEL ANALYSIS

MULTILEVEL ANALYSIS
Tom A. B. Snijders
http://www.stats.ox.ac.uk/~snijders/mlbook.htm
Department of Statistics
University of Oxford
2012
Foreword
This is a set of slides following Snijders & Bosker (2012).

The page headings give the chapter numbers and the page numbers in the book.
Literature:
Tom Snijders & Roel Bosker,
Multilevel Analysis: An Introduction to Basic and Applied Multilevel Analysis,
2nd edition. Sage, 2012.
Chapters 1-2, 4-6, 8, 10, 13, 14, 17.
There is an associated website
http://www.stats.ox.ac.uk/~snijders/mlbook.htm
containing data sets and scripts for various software packages.
These slides are not self-contained, for understanding them it is necessary
also to study the corresponding parts of the book!
2
2. Multilevel data and multilevel analysis 7
2. Multilevel data and multilevel analysis

Multilevel Analysis using the hierarchical linear model :
random coefficient regression analysis for data with several nested levels.
Each level is (potentially) a source of unexplained variability.

3
2. Multilevel data and multilevel analysis 9
Some examples of units at the macro and micro level:
macro-level micro-level
schools teachers
classes pupils
neighborhoods families
districts voters
firms departments
departments employees
families children
litters animals
doctors patients
interviewers respondents
judges suspects
subjects measurements
respondents = egos alters
4
2. Multilevel data and multilevel analysis 11–12
Multilevel analysis is a suitable approach to take into account the social contexts
as well as the individual respondents or subjects.
The hierarchical linear model is a type of regression analysis for multilevel data
where the dependent variable is at the lowest level.
Explanatory variables can be defined at any level
(including aggregates of level-one variables).
Z Z Z
. . . .@
.. ... ...@
... ... . . A. . . . . . .
@ @ A
@
R y @
R y A
@
x @
- x AU - y
Figure 2.5 The structure of macro–micro propositions.
Also longitudinal data can be regarded as a nested structure;

for such data the hierarchical linear model is likewise convenient.
5
2. Multilevel data and multilevel analysis 7–8
Two kinds of argument to choose for a multilevel analysis instead of an OLS

regression of disaggregated data:
1. Dependence as a nuisance
Standard errors and tests base on OLS regression are suspect
because the assumption of independent residuals is invalid.
2. Dependence as an interesting phenomenon
It is interesting in itself to disentangle variability at the various levels;
moreover, this can give insight in the directions
where further explanation may fruitfully be sought.
6
4. The random intercept model 42
4. The random intercept model

Hierarchical Linear Model:
i indicates level-one unit (e.g., individual);
j indicates level-two unit (e.g., group).
Variables for individual i in group j :
Yij dependent variable;
xij explanatory variable at level one;
for group j :
zj explanatory variable at level two; nj group size.
OLS regression model of Y on X ignoring groups :

Yij = β0 + β1 xij + Rij .
Group-dependent regressions:
Yij = β0j + β1j xij + Rij .
7
Distinguish two kinds of fixed effects models:
1. models where group structure is ignored;

2. models with fixed effects for groups: β0j are fixed parameters.
In the random intercept model, the intercepts β0j are random variables
representing random differences between groups:
Yij = β0j + β1 xij + Rij .
where β0j = average intercept γ00 plus group-dependent deviation U0j :
β0j = γ00 + U0j .
In this model, the regression coefficient β1 is common to all the groups.
8
In the random intercept model, the constant regression coefficient β1 is

sometimes denoted γ10:
Substitution yields
Yij = γ00 + γ10 xij + U0j + Rij .
In the hierarchical linear model, the U0j are random variables.

The model assumption is that they are independent,
normally distributed with expected value 0, and variance
τ 2 = var(U0j ).
The statistical parameter in the model is not their individual values,
but their variance τ02 .
9
y12
Y p
regression line group 2
R12


β02

β03

β01
X
Figure 4.1 Different parallel regression lines.
The point y12 is indicated with its residual R12 .
10
4. The random intercept model 46–47
Arguments for choosing between fixed (F ) and random (R) coefficient models for
the group dummies:
1. If groups are unique entities and inference should focus on these groups: F .
This often is the case with a small number of groups.
2. If groups are regarded as sample from a (perhaps hypothetical) population and
inference should focus on this population, then R .
This often is the case with a large number of groups.
3. If level-two effects are to be tested, then R .
4. If group sizes are small and there are many groups, and it is reasonable to
assume exchangeability of group-level residuals, then R makes better use of the
data.
5. If the researcher is interested only in within-group effects, and is suspicious
about the model for between-group differences, then F is more robust.
6. If group effects U0j (etc.) are not nearly normally distributed, R is risky
(or use more complicated multilevel models).
11
4. The random intercept model 49; also see 17–18
The empty model (random effects ANOVA) is a model

without explanatory variables:
Yij = γ00 + U0j + Rij .
Variance decomposition:
var(Yij ) = var(U0j ) + var(Rij ) = τ02 + σ 2 .
Covariance between two individuals (i 6= i0 ) in the same group j :
cov(Yij , Yi0j ) = var(U0j ) = τ02 ,
and their correlation:
τ02
ρ(Yij , Yi0j ) = ρI (Y ) = .
(τ02 + σ 2)
This is the intraclass correlation coefficient.
Often between .05 and .25 in social science research,
where the groups represent some kind of social grouping.
12
Example: 3758 pupils in 211 schools , Y = language test.

Classrooms / schools are level-2 units.
Table 4.1 Estimates for empty model
Fixed Effect Coefficient S.E.

γ00 = Intercept 41.00 0.32
Random Part Variance Component S.E.

Level-two variance:
τ02 = var(U0j ) 18.12 2.16
Level-one variance:
σ 2 = var(Rij ) 62.85 1.49
Deviance 26595.3
13
Intraclass correlation
18.12
ρI = = 0.22
18.12 + 62.85
Total population of individual values Yij has estimated mean 41.00 and standard
√
deviation 18.12 + 62.85 = 9.00 .
Population of class means β0j has estimated mean 41.00 and standard deviation
√
18.12 = 4.3 .
The model becomes more interesting,
when also fixed effects of explanatory variables are included:
Yij = γ00 + γ10 xij + U0j + Rij .
(Note the difference between fixed effects of explanatory variables

and fixed effects of group dummies!)
14
Table 4.2 Estimates for random intercept model with effect for IQ

γ00 = Intercept 41.06 0.24
γ10 = Coefficient of IQ 2.507 0.054

Level-two variance:
τ02 = var(U0j ) 9.85 1.21
Level-one variance:
σ 2 = var(Rij ) 40.47 0.96
Deviance 24912.2
There are two kinds of parameters:
1. fixed effects: regression coefficients γ (just like in OLS regression);

2. random effects: variance components σ 2 and τ02 .
15
Table 4.3 Estimates for ordinary least squares regression

γ00 = Intercept 41.30 0.12

Level-one variance:
σ 2 = var(Rij ) 49.80 1.15
Deviance 25351.0
Multilevel model has more structure (“dependence interesting”);

OLS has misleading standard error for intercept (“dependence nuisance”).
16
Y .........
................
..
................
... ...........................
. .
................
................ ........
..
................ ........
................
55 .
. .. . .... . . . ....
.................
.......................
. . .... . .
........
........ ...............
. . .
. .
........ .....................................
................ ........ ...........................................................
................ ........ . . . . .
................ ........ ........................................................................................
................ ........ . . . . . . .
.............................. ..... . ......... ..................................................................................................................
................. ..... .........................................................................................................
50 ................
..
. ...
................
................
. . .
........
. . . ..
. .. . . . . . . . . . .
........ ......................................................................................................................................................
. . . . . . . . . .
........ ......................................................................................................................................................
................ ........ . . . . . . . . . .
....................... ........ ...................................................................................................................................................... ........
..
. ... . . . . . ... . . . . . . . . . . . . . ........
................ ........ ................ ........ ........................ ................ ........ ........ ........
. ........................... . ..... ........ ...................................................................................................................................................... ... .. ........
. .
..... ..... ..... . ..... ..... ..... .......... ..... ..... ..... ..... ..... ..... .. .......
................ ........ ...................................................................................................................................................... ........ ........
................ ........ . . . . . . . . . . ........ ........
. . ..... ....................... ..... . ......... ...................................................................................................................................................... ............... ....... ........
. . . . . . . . . . . . . . .
................ ........ ...................................................................................................................................................... ........ ........
................ ........ . . . . . . . . . . ........ ........
................ ........ ...................................................................................................................................................... ........ ........
................ ........ . . . . . . . . . . ........ ........
. . ............................ .. ... . ......... ...................................................................................................................................................... ..... .. ........ ..... .. ........
. . . . . . . . . . . . . . .
................ ........ ...................................................................................................................................................... ........ ........
................ ........ . . . . . . . . . . ........ ........
................ ........ ...................................................................................................................................................... ........ ........
................ ........ . . . . . . . . . . ........ ........
...... . ....................... ... .. . ......... ...................................................................................................................................................... ... .... ........ ...............
. . . . . . . . . . . . . . .
................ ........ ...................................................................................................................................................... ........ ........
................ ........ . . . . . . . . . . ........ ........
................ ........ ...................................................................................................................................................... ........ ........
................ ........ . . . . . . . . . . ........ ........
... . . . . ....................... . . . ... . ........ ...................................................................................................................................................... ............... .... ...........
.. ... . . . . . . . . . . . . .
....................... ........ ...................................................................................................................................................... ........ ........
........ . . . . . . . . . . ........ ........
.. ........ ...................................................................................................................................................... ........ ........
........ . . . . . . . . . . ........ ........
. .. .. . ......... ...................................................................................................................................................... ...... . ........ ..... .. ........
....... ............... ....... ..................... .............. ....... ........ ........
.
........
.
........ ....................................................................................................................................................... ........ ........
........... ...... .......... ...... ...... ...... ...... ...... ........ ........
................ ........ ........................ ................ ................ ........ ........
................................................................................................................................
.................................................................................................
.... .... .... .... .... .... .... ........
........
. ...... . ........
........
........
................
30
..................................................................................... ........ ........
........ ........
. ... ... ... ...
....................................................... ... . ... ........ . ..... . ........
. .
. ................. ........ ........
..... ...... ........ ........
........ ........
........
........
..... . . .. ... ........
........
........
. ....
. .. ........
........ 25
........
........
........
. . . . . . .........
.
−4 −3 −2 −1 0 1 2 3 4 X = IQ
Figure 4.2 Fifteen randomly chosen regression lines according to the random intercept model of
Table 4.2.
17
More explanatory variables:
Yij = γ00 + γ10 x1ij + . . . + γp0 xpij + γ01 z1j + . . . + γ0q zqj
+ U0j + Rij .
Especially important:
difference between within-group and between-group regressions.
The within-group regression coefficient is the regression coefficient within each
group, assumed to be the same across the groups.
The between-group regression coefficient is defined as the regression coefficient for
the regression of the group means of Y on the group means of X.
This distinction is essential to avoid ecological fallacies (p. 15–17 in the book).
18
Y
between-group regression line
"
"
"
"
"
"
"
regression line
"
"
within group 2
"
" regression line within group 3
"
"
"

regression line within group 1
"
"
"
"

"
"
"
X
Figure 4.3 Different between-group and within-group regression lines.
This is obtained by having separate fixed effects for the level-1 variable X
and its group mean X̄.
(Alternative:
use the within-group deviation variable X̃ij = (X − X̄) instead of X.)
19
Table 4.4 Estimates for random intercept model

with different within- and between-group regressions

γ00 = Intercept 41.11 0.23
γ01 = Coefficient of IQ (group mean) 1.312 0.262

Level-two variance:
τ02 = var(U0j ) 8.68 1.10
Level-one variance:
σ 2 = var(Rij ) 40.43 0.96
Deviance 24888.0
20
In the model with separate effects for the original variable xij and the group mean
Yij = γ00 + γ10 xij + γ01x.j + U0j + Rij ,
the within-group regression coefficient is γ10 ,
between-group regression coefficient is γ10 + γ01.
This is convenient because the difference between within-group and between-group
coefficients can be tested by considering γ01.
In the model with separate effects for group-centered variable x̃ij
and the group mean
Yij = γ̃00 + γ̃10 x̃ij + γ̃01x.j + U0j + Rij ,
the within-group regression coefficient is γ̃10 ,
the between-group regression coefficient is γ̃01.
This is convenient because these coefficients are given immediately in the results,
with their standard errors.
Both models are equivalent, and have the same fit: γ̃10 = γ10, γ̃01 = γ10 + γ01.
21
Estimation/prediction of random effects

The random effects U0j are not statistical parameters and therefore they are not
estimated as part of the estimation routine.
However, it sometimes is desirable to ‘estimate’ them. This can be done by the
empirical Bayes method; these ‘estimates’ are also called the posterior means.
In statistical terminology, this is not called ‘estimation’ but ‘prediction’, the name
for the construction of likely values for unobserved random variables.
The posterior mean for group j is based on two kinds of information:

⇒ sample information : the data in group j;
⇒ population information : the value U0j was drawn from a normal distribution
with mean 0 and variance τ02 .
The population information comes from the other groups.

If this information is reasonable, prediction is improved on average.
22
Suppose we wish to predict the ‘true group mean’ γ00 + U0j .

The empirical Bayes estimate in the case of the empty model is a weighted average
of the group mean and the overall mean:
EB
β̂0j = λj Ȳ.j + (1 − λj ) γ̂00 ,
where the weight λj is the ‘reliability’ of the mean of group j
τ02 nj ρI
λj = = .
τ02 + σ 2/nj 1 + (nj − 1)ρI
λj
1.0 ρI = 0.4
The reliability coefficient indicates 0.8 ρI = 0.1
how well the true mean γ00 + U0j is
measured by the observed mean Ȳ.j ;
0.4
see Section 3.5.
The picture to the rights gives a plot.
0.0
10 30 50 nj
23
These ‘estimates’ are not unbiased for each specific group, but they are more
precise when the mean squared errors are averaged over all groups.
For models with explanatory variables, the same principle can be applied:
the values that would be obtained as OLS estimates per group are
“shrunk towards the mean”.
The empirical Bayes estimates, also called posterior means,
are also called shrinkage estimators.
24
There are two kinds of standard errors for empirical Bayes estimates:
comparative standard errors

EB EB
S.E.comp Ûhj = S.E. Ûhj − Uhj
for comparing the random effects of different level-2 units
(use with caution – E.B. estimates are not unbiased!);
and diagnostic standard errors

EB EB
S.E.diag Ûhj = S.E. Ûhj
used for model checking (e.g., checking normality of the level-two residuals).
25
U0j
10
---
-- --•••
-
--
- - - ------- •
- -
5 ----------------- ••••••---
- -
--
-- - - - --- ---- -------------------- ••••••••••••••••••••••• --
- -- - --- ----------------- ---- --
- •••••••••••••••• - ------------
- - - •
•
-- - --- --- --- ----- - - - ••••••••••••••• - --- ------ - -
----------------- --------- ---- -- --- •••••••••••••••••••••••••••••••••• - --------------- -- -
0 - -- •••••••••••••••••••• ---------------
-------------------- ••••••••••••••••••••••••••••••••••• ------- - - -
--- ----- - ----- -- -
- ---
- - - ••••••••••• - --- - ---- ---
- -------
•••••••••
••• - --- -------------- -- - ----- -
- ---- ••••• --- - -- - - --
------- ••••••• ------ ----------- --- --- -
•••••• -- - -
−5 - •••••••• - --------- - -
•••••• ---- ----- ---
• --- -- --
-
----- -
-
−10
The ordered added value scores for 211 schools with comparative posterior confidence intervals.
In this figure, the error bars extend 1.39 times the comparative standard errors
to either side, so that schools may be deemed to be significantly different
if the intervals do not overlap (no correction for multiple testing!).
26
5. The hierarchical linear model 74–75
5. The hierarchical linear model

It is possible that not only the group average of Y ,
but also the effect of X on Y is randomly dependent on the group.
In other words, in the equation
Yij = β0j + β1j xij + Rij ,
also the regression coefficient β1j has a random part:
β0j = γ00 + U0j

β1j = γ10 + U1j .
Substitution leads to
Yij = γ00 + γ10 xij + U0j + U1j xij + Rij .
Variable X now has a random slope.
27
It is assumed that the group-dependent coefficients (U0j , U1j )

are independent across j, with a bivariate normal distribution
with expected values (0, 0) and covariance matrix defined by
var(U0j ) = τ00 = τ02 ;
var(U1j ) = τ11 = τ12 ;
cov(U0j , U1j ) = τ01 .
Again, the (U0j , U1j ) are not individual parameters in the statistical sense, but
only their variances, and covariance, are the parameters.
Thus we have a linear model for the mean structure, and a parametrized
covariance matrix within groups with independence between groups.
28
5. The hierarchical linear model 78
5.1 Estimates for random slope model

γ00 = Intercept 41.127 0.234 The equation for this table is
γ10 = Coeff. of IQ 2.480 0.064
γ01 = Coeff. of IQ (group mean) 1.029 0.262
Yij = 41.13 + 2.480 IQij
Random Part Parameters S.E. + 1.029 IQ.j
Level-two random part:
τ02 = var(U0j ) 8.877 1.117 + U0j + U1j IQij + Rij .
τ12 = var(U1j ) 0.195 0.076
τ01 = cov(U0j , U1j ) –0.835 0.217 The slope β1j has
Level-one variance:
σ 2 = var(Rij ) 39.685 0.964
average 2.480
and
Deviance 24864.9
√
s.d. 0.195 = 0.44.
IQ is defined as the group mean.
29
Y
........
............
55 ..... .. .. . . . .. . . ............
............
... . . . . . . . . . . . ............
. .. . . . .
............
. .. ... .. .. .
............ ...........
. .
. . . . . .
. .
. .
..................................................
......................................................................
.
.......................................................
..
............ ............................................. ..................
............ ..................................................... .............................
............ ............................................... ........ .........................................
50 ............
............
............
............
..... . . . . . . . .. . . ............
............
. .... . . .. .. .. .
......................................... ........ .................................. ..........
...................................................................................................................................................................
. . .
............... ........................... ......... ..........................
.............. ............................................................... ..................
...................... ............................................................
........................ ..................... .......................................... ..........
..
.
..........
. .
............ ..........
................
......
.... . .. ............... ...... ......

..
.......................................................................................................................................... .................. . . . . . . . . . . . ..............................
. . . . .... . . . ........ .
...... . ..
..
..
.
...................... .................................................................................................. ............... .
.. . . . . . . . . . ....... ................
.
..... . . .. .. . . . .. .
............ ........................ ....................................................................................................................................................... ............ ..........
............ ....................... . . .. .. . . . . . ............ ..................
............ ....................... .................................................................................................................................... ....................... .
............ ........................ . .. . . . . . ..........
....
.... . . . ............ . . . . . . . . . ................................... .............................................................................................................................................................. .. .. . . . . ...........
..... .. .. ... .. .. .. . .. .. .. .. ..........
............ ....................... ....................... ...................... ......... ........ ................ ............ ..........
............ .................................................................................................................. ............................................................................... ..........
............ ............ ........... . .. .. . . . . .
............ ............ ........... ............. ................................. ............. ........ ..................... ..........
..........
. ............ .....
.... . . ..................................................................................................................... ........................................................................................ . . . . . . . . . ..........
..... . .. .. . . .. . . .
............ ............ ........... ............. ........................... ................. ............................. ..........
............ ............ ........................ ........................................... ................. ................ ....... ..........
............ ............ ........................ . . ................................... ........ ..........
............ ............ ......................... ................................. .................................... ........ ..........
.... . . . . . . . ............ .................................................................... . . ...
....
...
..
.
..
...........................................
. ... . .
. .
. .
..
.........................................................................
. ...... . . ..........
.. . . .. ... .. .
............ ............ ...........
. ... ..
................................. ........... ......................... ........
.. ..........
............ ................. ................................. ............... ...................... ........ ..........
............ ............ ............................... .................... ....... ........ ........ ..........
............ ............... ................................ ..................... ......... ........ ....... ..........
. . . . ............................................................. ......
.............................................................. .. ...... ......................................................................................... . . . . . . .. .........
.
.
....... ........ ....... .
... .. .. ... . .. .. .. .. .. ..
........................................ ........... ......................... ....................... ................................................................................. ..........
........................... ........... .................... . ..........
........... ........... ................... ....................... .................. ........ ........ ........ ..................
........... .................... . .................. ........ ........ ........ .
. . . . . ............................................................................... . . . . . . . ............................................... ............... ..................
....... ....... ...... ....... . . . ..
................................................................................. ............ .............................. .................................
..........
30
. . .
............ .............. .......... ............................... ..................................
............................................... .......... . . . .
........... . .. . ..
..
..
....................... .............................. .................................
.
................. . .
........ ........ .................
. .
.................. ........ ........ ............
................... ........ ........ .......
................... ........ ........ ...........
.................. ..................................................... .
........................... .... . . .
... ... .... ...
........ ..........................
.....
........ .................. ........
........ .................. ........
........ .................. ........
........................ . ...............
... ..........
......
25
........
−4 −3 −2 −1 0 1 2 3 4 X = IQ
Figure 5.2 Fifteen random regression lines according to the model of Table 5.1.
Note the heteroscedasticity: variance is larger for low X than for high X.
The lines fan in towards the right.
Intercept variance and intercept-slope covariance depend on the position of the
X = 0 value, because the intercept is defined by the X = 0 axis.
30
The next step is to explain the random slopes:

β0j = γ00 + γ01 zj + U0j
β1j = γ10 + γ11 zj + U1j .
Substitution then yields
Yij = (γ00 + γ01 zj + U0j )

+ (γ10 + γ11 zj + U1j ) xij + Rij
= γ00 + γ01 zj + γ10 xij + γ11 zj xij
+ U0j + U1j xij + Rij .
The term γ11 zj xij is called the cross-level interaction effect.
31
Table 5.2 Estimates for model with random slope

and cross-level interaction

γ00 = Intercept 41.254 0.235
γ11 = Coefficient of IQ × IQ –0.187 0.064
Random Part Parameters S.E.

τ02 = var(U0j ) 8.601 1.088
τ12 = var(U1j ) 0.163 0.072
τ01 = cov(U0j , U1j ) –0.833 0.210
Level-one variance:
σ 2 = var(Rij ) 39.758 0.965
Deviance 24856.8
32
For two variables (IQ and SES) and two levels (student and school),
the main effects and interactions give rise to a lot of possible combinations:
Table 5.3 Estimates for model with random slopes and many effects

γ00 = Intercept 41.632 0.255
γ20 = Coefficient of SES 0.172 0.012
γ30 = Interaction of IQ and SES –0.019 0.006
γ02 = Coefficient of SES –0.090 0.044
γ11 = Interaction of IQ and IQ –0.081 0.081
γ12 = Interaction of IQ and SES 0.004 0.013
γ21 = Interaction of SES and IQ 0.023 0.018
γ22 = Interaction of SES and SES 0.000 0.002
(continued next page....)
33

τ02 = var(U0j ) 8.344 1.407
τ12 = var(U1j ) 0.165 0.069
τ01 = cov(U0j , U1j ) –0.942 0.204
τ22 = var(U2j ) 0.0 0.0
τ02 = cov(U0j , U2j ) 0.0 0.0
Level-one variance:
σ 2 = var(Rij ) 37.358 0.907
Deviance 24624.0
The non-significant parts of the model may be dropped:
34
Table 5.4 Estimates for a more parsimonious model with a random slope and many effects

γ00 = Intercept 41.612 0.247
γ20 = Coefficient of SES 0.174 0.012
γ02 = Coefficient of SES –0.089 0.042

τ02 = var(U0j ) 8.369 1.050
τ12 = var(U1j ) 0.164 0.069
τ01 = cov(U0j , U1j ) –0.929 0.204
Level-one variance:
σ 2 = var(Rij ) 37.378 0.907
Deviance 24626.8
35
Estimation for the hierarchical linear model
General formulation of the two-level model
As a link to the general statistical literature,

it may be noted that the two-level model can be expressed as follows:
Yj = Xj γ + Zj Uj + Rj
" # " # " #!

Rj ∅ Σj (θ) ∅
with ∼N ,
Uj ∅ ∅ Ω(ξ)
and (Rj , Uj ) ⊥ (R`, U`) for all j 6= ` .
Standard specification Σj (θ) = σ 2Inj ,

but other specifications are possible.
Mostly, Σj (θ) is diagonal, but even this is not necessary (e.g. time series).
36
Estimation for the hierarchical linear model
The model formulation yields

Yj ∼ N Xj γ, Zj Ω(ξ)Zj0 + Σj (θ) .
This is a special case of the mixed linear model

Y = Xγ + ZU + R,
with X[n, r], Z[n, p], and
! ! !!
R ∅ Σ ∅
∼N , .
U ∅ ∅ Ω
For estimation, the ML and REML methods are mostly used.

These can be implemented by various algorithms: Fisher scoring,
EM = Expectation–Maximization, IGLS = Iterative Generalized Least Squares.
See Section 4.7 and 5.4.
37
6. Testing 94–98
6. Testing
To test fixed effects, use the t-test with test statistic
γ̂h
T (γh) = .
S.E.(γ̂h)
(Or the Wald test for testing several parameters simultaneously.)
The standard error should be based on REML estimation.
Degrees of freedom for the t-test, or the denominator of the F -test:
For a level-1 variable: M − r − 1,
where M = total number of level-1 units, r = number of tested level-1 variables.
For a level-2 variable: N − q − 1,
where N = number of level-2 units, q = number of tested level-2 variables.
For a cross-level interaction: again N − q − 1,
where now q = number of other level-2 variables interacting with this level-1
variable.
If d.f. ≥ 40, the t-distribution can be replaced by a standard normal.
38
6. Testing 94–98
For parameters in the random part, do not use t-tests.

Simplest test for any parameters (fixed and random parts)
is the deviance (likelihood ratio) test, which can be used
when comparing two model fits that have used the same set of cases:
subtract deviances, use chi-squared test
(d.f. = number of parameters tested).
Deviance tests can be used to compare any two nested models.
If these two models do not have the same fixed parts,
then ML estimation should be used!
Other tests for parameters in the random part have been developed
which are similar to F -tests in ANOVA.
39
6. Testing 94–98
6.1 Two models with different between- and within-group regressions
Model 1 Model 2
Fixed Effects Coefficient S.E. Coefficient S.E. Test for equality of within- and
γ00 = Intercept 41.15 0.23 41.15 0.23 between-group regressions
γ10 = Coeff. of IQ 2.265 0.065
is t-test for IQ in Model 1:
γ20 = Coeff. of IQ
e 2.265 0.065
γ30 = Coeff. of SES 0.161 0.011 0.161 0.011 t = 0.647/0.264 = 2.45,
γ01 = Coeff. of IQ 0.647 0.264 2.912 0.262 p < 0.02.
Random Part Parameter S.E. Parameter S.E. Model 2 gives

Level-two parameters: within-group coefficient 2.265
τ02 = var(U0j ) 9.08 1.12 9.08 1.12
τ12 = var(U1j ) 0.197 0.074 0.197 0.074
and between-group coefficient
τ01 = cov(U0j , U1j ) –0.815 0.214 –0.815 0.214 2.912 = 2.265 + 0.647.
Level-one variance:
σ 2 = var(Rij ) 37.42 0.91 37.42 0.91
Deviance 24661.3 24661.3
40
6. Testing 98–99
However, one special circumstance: variance parameters are necessarily positive.

Therefore, they may be tested one-sided.
E.g., in the random intercept model
under the null hypothesis that τ02 = 0,
the asymptotic distribution of –2 times the log-likelihood ratio (deviance difference)
is a mixture of a point mass at 0 (with probability 12 )
and a χ2 distribution (also with probability 12 .)
The interpretation is that if the observed between-group variance
is less than expected under the null hypothesis
– which happens with probability 21 –
the estimate is τ̂02 = 0 and the log-likelihood ratio is 0.
The test works as follows:
if deviance difference = 0, then no significance;
if deviance difference > 0, calculate p-value from χ21 and divide by 2.
41
6. Testing 98–99
For testing random slope variances,

if the number of tested parameters (variances & covariances) is p + 1,
the p-values can be obtained as
the average of the p-values for the χ2p and χ2p+1 distributions.
(Apologies for the use of the letter p in two different meanings...)
Critical values for 50–50 mixture of χ2p and χ2p+1 distribution.
α
p 0.10 0.05 0.01 0.001
1 3.81 5.14 8.27 12.81
2 5.53 7.05 10.50 15.36
3 7.09 8.76 12.48 17.61
42
6. Testing 98–99
For example: testing for a random slope in a model that further contains the
random intercept but no other random slopes: p = 1;
testing the second random slope: p = 2;
testing the third random slope: p = 3 – etc.
To test the random slope in the model of Table 5.1,
compare with Table 4.4 which is the same model but without the random slope;
deviance difference 15,227.5 – 15,213.5 = 14.0.
In the table with p = 1 this yields p < 0.001.
Further see p. 99.
43
7. Explained variance 109–110
7. Explained variance
The individual variance parameters may go up when effects are added to the model.
7.1 Estimated residual variance parameters σ̂ 2 and τ̂02 for models

with within-group and between-group predictor variables
σ̂ 2 τ̂02
I. BALANCED DESIGN
A. Yij = β0 + U0j + Eij 8.694 2.271
B. Yij = β0 + β1 X .j + U0j + Eij 8.694 0.819
C. Yij = β0 + β2(Xij − X .j ) + U0j + Eij 6.973 2.443
II. UNBALANCED DESIGN

A. Yij = β0 + U0j + Eij 7.653 2.798
B. Yij = β0 + β1 X .j + U0j + Eij 7.685 2.038
C. Yij = β0 + β2(Xij − X .j ) + U0j + Eij 6.668 2.891
44
7. Explained variance 112–113
The best way to define R2, the proportion of variance explained, is the
proportional reduction in total variance ;
for the random intercept model total variance is (σ 2 + τ02).
Table 7.2 Estimating the level-1 explained variance

(balanced data)
σ̂ 2 τ̂02
A. Yij = β0 + U0j + Eij 8.694 2.271
D. Yij = β0 + β1 (Xij − X .j ) + β2 X .j + U0j + Eij 6.973 0.991
Explained variance at level 1:

2 6.973 + 0.991
R1 = 1 − = 0.27.
8.694 + 2.271
45
8. Heteroscedasticity 119-120
8. Heteroscedasticity
The multilevel model allows to formulate heteroscedastic models where residual
variance depends on observed variables.
E.g., random part at level one = R0ij + R1ij x1ij .
Then the level-1 variance is a quadratic function of X:
var(R0ij + R1ij xij ) = σ02 + 2 σ01 x1ij + σ12 x21ij .
For σ12 = 0, this is a linear function:

var(R0ij + R1ij xij ) = σ02 + 2 σ01 x1ij .
This is possible as a variance function, without random effects interpretation.
Different statistical programs have implemented
various different variance functions.
46
8. Heteroscedasticity 121
8.1 Homoscedastic and heteroscedastic models.
Model 1 Model 2
Fixed Effect Coefficient S.E. Coefficient S.E.
Intercept 40.426 0.265 40.435 0.266
IQ 2.249 0.062 2.245 0.062
SES 0.171 0.011 0.171 0.011
IQ × SES –0.020 0.005 –0.019 0.005
Gender 2.407 0.201 2.404 0.201
IQ 0.769 0.293 0.749 0.292
SES –0.093 0.042 –0.091 0.042
IQ × SES –0.105 0.033 –0.107 0.033
Random Part Parameters S.E. Parameters S.E.

Intercept variance 8.321 1.036 8.264 1.030
IQ slope variance 0.146 0.065 0.146 0.065
Intercept - IQ slope covariance –0.898 0.197 –0.906 0.197
Level-one variance:
σ02 constant term 35.995 0.874 37.851 1.280
σ01 gender effect –1.887 0.871
Deviance 24486.8 24482.2
47
This shows that there is significant evidence for heteroscedasticity:

χ21 = 4.6, p < 0.05.
The estimated residual (level-1) variance is
37.85 for boys and 37.85 – 2×1.89 = 34.07 for girls.
The following models show, however, that the heteroscedasticity as a function of

IQ is more important.
First look only at Model 3.
48
8.2 Heteroscedastic models depending on IQ.

Model 3 Model 4
Fixed Effect Coefficient S.E. Coefficient S.E.
Intercept 40.51 0.26 40.51 0.27
IQ 2.200 0.058 3.046 0.125
SES 0.175 0.011 0.168 0.011
IQ × SES –0.022 0.005 –0.016 0.005
Gender 2.311 0.198 2.252 0.196
IQ 0.685 0.289 0.800 0.284
SES –0.087 0.041 –0.083 0.041
IQ × SES –0.107 0.033 –0.089 0.032
IQ2− 0.193 0.038
IQ2+ –0.260 0.033
Random Part Parameter S.E. Parameter S.E.
Level-two random effects:
Intercept variance 8.208 1.029 7.989 1.002
IQ slope variance 0.108 0.057 0.044 0.048
Intercept - IQ slope covariance –0.733 0.187 –0.678 0.171
Level-one variance parameters:
σ02 constant term 36.382 0.894 36.139 0.887
σ01 IQ effect –1.689 0.200 –1.769 0.191
Deviance 24430.2 24369.0
49
8. Heteroscedasticity 122–123
The level-1 variance function for Model 3 is 36.38 – 3.38 IQ .

Maybe further differentiation is possible between low-IQ pupils?
Model 4 uses(
2
IQ if IQ < 0
IQ2− =
0 if IQ ≥ 0 ,
(
2 0 if IQ < 0
IQ+ =
IQ2 if IQ ≥ 0 .
Y
..............
...............
8 ..
...........
.
...
...
...
..............
.............
............
..........
..........
.........
..
...
...........
.
........
.......
........
4 .
.......
..
...
.
.......
.......
......
......
......
..
.......
.
.
......
......
.....
.....
..
......
−4 −2 .....
......
......
.
......
2 4 IQ
..
..
.......
......
......
......
.......
.
.......
......
.......
..
...
........
−4
.......
.
...
.........
.
..
........
........
........
........
........
.
.........
.........
.........
..
...
..
−8
..........
.
...
...
............
......
...........
Effect of IQ on language test as estimated by Model 4.

50
8. Heteroscedasticity 127–128
Heteroscedasticity can be very important for the researcher

(although mostly she/he doesn’t know it yet).
Bryk & Raudenbush: Correlates of diversity.
Explain not only means, but also variances!
Heteroscedasticity also possible for level-2 random effects:

give a random slope at level 2 to a level-2 variable.
51
10. Assumptions of the hierarchical linear model 152–153
10. Assumptions of the Hierarchical Linear Model
r
X p
X
Yij = γ0 + γh xhij + U0j + Uhj xhij + Rij .
h=1 h=1
Questions:
1. Does the fixed part contain the right variables (now X1 to Xr )?

2. Does the random part contain the right variables (now X1 to Xp)?
3. Are the level-one residuals normally distributed?
4. Do the level-one residuals have constant variance?
5. Are the level-two random coefficients normally distributed with mean 0?
6. Do the level-two random coefficients have a constant covariance matrix?
52
10. Assumptions of the hierarchical linear model 154–156; also 56–59
Follow the logic of the HLM

1. Include contextual effects
For every level-1 variable Xh, check the fixed effect of the group mean X̄h.
Econometricians’ wisdom: “the U0j must not be correlated with the Xhij .
Therefore test this correlation by testing the effect of X̄h (’Hausman test’)
Use a fixed effects model if this effect is significant”.
Different approach to the same assumption:
Include the fixed effect of X̄h if it is significant,
and continue to use a random effects model.
(Also check effects of variables X̄h.j Zj for cross-level interactions involving Xh!)
Also the random slopes Uhj must not be correlated with the Xkij .
This can be checked by testing the fixed effect of X̄k.j Xhij .
This procedure widens the scope of random coefficient models beyond what is
allowed by the conventional rules of econometricians.
53
Within- and between-group regressions 154–156; also 56–59
Assumption that level-2 random effects Uj have zero means.
What kind of bias can occur if this assumption is made but does not hold?
For a misspecified model,
suppose that we are considering a random intercept model:
Zj = 1 j
where the expected value of Uj is not 0 but
E Uj = z2j γ?
for 1 × r vectors z2j and an unknown regression coefficient γ?. Then
Uj = z2j γ? + Ũj
with
E Ũj = 0 .
54
Within- and between-group regressions 154–156; also 56–59
Write Xj = X̄j + X̃j , where X̄j = 1j (10j 1j )−110j Xj are the group means.
Then the data generating mechanism is
Yj = X̄j γ + X̃j γ + 1j z2j γ? + 1j Ũj + Rj ,
where E Ũj = 0 .
There will be a bias in the estimation of γ
if the matrices Xj = X̄j + X̃j and 1j Ũj are not orthogonal.
By construction, X̃j and 1j Ũj are orthogonal, so the difficulty is with X̄j .
The solution is to give X̄j and X̃j separate effects:

Yj = X̄j γ1 + X̃j γ2 + 1j Uj + Rj .
Now γ2 has the role of the old γ:
‘the estimation is done using only within-group information’.
Often, there are substantive interpretations of the difference between the
within-group effects γ2 and the between-group effects γ1.
55
Within- and between-group regressions 155-161
2. Check random effects of level-1 variables.

See Chapter 5.
4. Check heteroscedasticity.
See Chapter 8.
3,4. Level-1 residual analysis

5,6. Level-2 residual analysis
For residuals in multilevel models, more information is in Chapter 3 of
Handbook of Multilevel Analysis (eds. De Leeuw and Meijer, Springer 2008)
(preprint at course website).
56
Residuals 161–165
Level-one residuals
OLS within-group residuals can be written as

R̂j = Inj − Pj Yj
where we define design matrices X̌j comprising Xj as well as Zj

(to the extent that Zj is not already included in Xj ) and
Pj = X̌j (X̌j0 X̌j )−1X̌j0 .
Model definition implies

R̂j = Inj − Pj Rj :
these level-1 residuals are not confounded by Uj .
57
Residuals 161–165
Use of level-1 residuals :

Test the fixed part of the level-1 model using OLS level-1 residuals,
calculated per group separately.
Test the random part of the level-1 model using

squared standardized OLS residuals.
In other words, the level-1 specification can be studied
by disaggregation to the within-group level
(comparable to a “fixed effects analysis”).
The construction of the OLS within-group residuals implies
that this tests only the level-one specification,
independent of the correctness of the level-two specification.
The examples of Chapter 8 are taken up again.
58
Residuals 164
Example: model with effects of IQ, SES, sex.

r̄
- r̄
-
2 --
2 -
••
- - --- - - - -
- - - - • -- - - -
• -
• - - ••• - • - - - • - - • - - - -
-
0 • - - ••• - - • - - -
• • •• - - - • - -•
• •• -
•
•
-
•
- • - -
- - •••
-
0 •••• • • - •• ••
-
- - - - -
• -
- - - -
- -
−2
- - • - - - - - -
- - -
-- −2
- - -
-
−4 −2 0 2 4
IQ
−10 0 10 20 SES
Mean level-one OLS residuals Mean level-one OLS residuals
(bars ∼ twice standard error of the mean) as function of SES.
as function of IQ.
This suggest a curvilinear effect of IQ.
59
Residuals 164
Model with effects also of IQ2− and IQ2+ .
r̄ r̄
2 - -
-- 2
---- - - -
- - - - - - -
- - - - • - •- -
• ••• - - - • • - - - - -
• • ••• -
- --
-• - - •• -•-• - -
0 •- • • •• • - • - •• • •
• • 0
•
• - - - - • -
- -•
•- - • - • - • ••
- - - - -• - - - - -
- - - - - - - -
- - - - -
−2 −2 -
- -
−4 −2 0 2 4 −10 0 10 20
IQ SES
Mean level-one OLS residuals
Mean level-one OLS residuals as function of SES.
as function of IQ.
This looks pretty random.
60
Residuals 165
Are the within-group residuals normally distributed?
observed
.....
. ....
3
......
..
.....
.
2 .
......
. .
1 ......
..
. ....
0 ...
.
...
−1 ...
...
.
...
...
−2 .
......
..
−3 ........
.... ..
. .
.. ... . .. expected
−3−2−1 0 1 2 3
Figure 10.3 Normal probability plot of standardized level-one OLS residuals.
Left tail is a bit heavy, but this is not serious.

61
Residuals 165–167 and 62–67
Level-two residuals
Empirical Bayes (EB) level-two residuals defined as conditional means
Ûj = E {Uj | Y1, . . . , YN }
(using parameter estimates γ̂, θ̂, ξ̂)
= Ω̂ Zj0 V̂j−1 (Yj − Xj γ̂j ) = Ω̂ Zj0 V̂j−1 (Zj Uj + Rj − Xj (γ̂ − γ))
where
Vj = Cov Yj = Zj ΩZj0 + Σj , V̂j = Zj Ω̂Zj0 + Σ̂j ,
with Ω̂ = Ω(ξ̂) and Σ̂j = Σj (θ̂).
You don’t need to worry about the formulae.
62
‘Diagnostic variances’, used for assessing distributional properties of Uj :

Cov Ûj ≈ ΩZj0 Vj−1Zj Ω ,
‘Comparative variances’, used for comparing ‘true values’ Uj of groups:

Cov Ûj − Uj ≈ Ω − ΩZj0 Vj−1Zj Ω .
Note that
Cov (Uj ) = Cov (Uj − Ûj ) + Cov (Ûj ) .
Standardization (by diagnostic variances) :

q
Ûj0 {Cov
d (Ûj )}−1Ûj (with the sign reinstated)
is the standardized EB residual.
63
However,
0
−1
Ûj0 {Cov
d (Ûj )}−1Ûj ≈ Û
j
(OLS)
σ̂ 2(Zj0 Zj )−1 + Ω̂ (OLS)
Ûj
where Ûj(OLS)
= (Zj0 Zj )−1Zj0 (Yj − Xj γ̂j )
is the OLS estimate of Uj , estimated from level-1 residuals Yj − Xj γ̂j .
This shows that standardization by diagnostic variances
takes away the difference between OLS and EB residuals.
Therefore, in checking standardized level-two residuals,
the distinctoin between OLS and EB residuals loses its meaning.
Test the fixed part of the level-2 model using non-standardized EB residuals.
Test the random part of the level-2 model

using squared EB residuals standardized by diagnostic variance.
64
Residuals 166
U0j . .. U0j . . .
. . . .
.. . . . . .
. . . ... . .. . .
4 . . . . .... .. . . 4
. . . .... . . .. . .... . ... . .
..
. . . . . .............. . . . . . .. .. ... .. . .... ..... . .. ... .
. . . .. .. . . . . .
. . . . . . ............ .. ..... .. .. ............ .. . .. .... .. .. . . . ◦◦
0 . . ... ...... ....... . . ◦

◦
◦◦◦◦
◦
◦◦
◦
◦
◦
◦
◦
◦◦◦
◦◦ ◦◦◦
◦
◦◦◦ ◦
◦
◦◦◦
◦
◦
◦◦◦◦
◦
◦
◦◦ ◦◦
◦◦
◦◦
◦◦
◦◦
◦
◦◦◦
◦
◦◦
◦◦
◦◦
◦◦◦◦
◦
◦◦◦
. 0 . ... .. . . . . .. . ... . . . .
◦◦
◦
◦
◦
◦◦
◦ ◦◦
◦
◦◦ ◦◦
◦
◦◦ ◦
◦
◦◦◦
◦◦◦◦
◦◦
◦◦
◦◦◦
◦
◦◦
◦◦◦
◦◦
◦◦◦◦◦
◦
◦◦
◦◦◦◦
◦
◦◦◦
◦◦◦
◦◦
◦ ◦◦
◦
◦◦◦ ◦◦◦◦◦
◦
◦◦◦◦◦
◦◦
◦◦◦
◦◦ ◦◦
◦◦
. .. . . . . .
◦ ◦◦
..... . .. ......... .... ... .. ... . .

◦◦
◦ ◦◦
◦ ◦◦
.. . . ..... .... .......... ....

◦◦◦◦
◦◦
◦ ◦◦◦ ◦◦
◦◦
◦ ◦ ◦◦ ◦ ◦
◦◦◦ ◦
◦◦
◦◦◦◦◦
.. . ◦
◦
◦
◦ ◦◦
◦
◦◦◦
◦◦
◦
◦◦
◦◦◦
◦◦
◦◦
. . . . . . . . . ... . . . .
.. . . . . ... ... ..
◦
. . .. . ... .. . . .
−4 . . . . .... . . −4 .. . . . .. .
. ..
. . . .
. . . .. .
. . mean . . mean
−2 0 2 IQ −10 0 10 SES
Figure 10.4 Posterior intercepts as function of (left) average IQ and (right)

average SES per school. Smooth lowess approximations are indicated by . ◦
The slight deviations do not lead to concerns.
65
Residuals 166
U1j U1j
. .. . ..
. .. . .... .
. . . . .. . . .
0.5
. . 0.5 . ... . . . ...... ... .
. . .. .. ..... . ... . . .......... ..... . . ..
.. ..... . .. ..... .... ...... ........ .
.. . ....... ... .......... .
...... . .. .. . ◦
◦
...... ....... . . . ..... . ..
◦◦◦
0 . ◦ ◦ ◦◦
. ..
◦
...
◦◦◦
.
◦◦◦◦◦◦
◦◦
◦◦◦◦◦
◦◦
◦
◦◦◦ ◦
. ◦◦ ◦◦◦ .
◦
◦ ◦◦◦
◦
................. ..... . .. .. . .
◦◦
◦
◦◦
◦◦
◦◦
◦◦◦
◦◦
. ◦
◦◦◦◦
. . .. .. . ... ................. ...... .. ..

◦
◦◦
◦◦
◦
◦◦◦
◦
◦
◦◦
◦◦
◦◦◦
◦
◦◦◦
◦◦
◦
◦◦
◦
◦
◦◦
◦◦
◦◦
◦
◦ ◦
◦◦◦◦
◦ ◦
◦◦◦◦
◦
◦ ◦
◦
◦◦◦◦
◦◦ ◦◦
◦
◦◦◦◦ ◦
... 0 .. .
.
◦◦
◦◦◦
◦◦◦
◦
◦◦◦
◦
◦◦◦
◦
◦
◦◦
◦ ◦◦
◦
◦◦ ◦◦ ◦
◦
◦◦◦◦
◦
◦◦◦
◦
◦ ◦
◦◦ ◦
◦
◦◦◦◦◦
◦
◦ ◦
◦
◦◦◦
◦ ◦
◦◦
◦
◦◦◦
◦◦◦
◦◦◦
◦◦
◦◦◦
◦
◦◦◦◦◦◦
. ..... .... . ............ .. .... . .

◦◦
◦
◦ ◦◦ ◦◦
◦◦
. . . .. . ........ .. . .. .. .... ....... . . .

. . . ... . . . . .. .
−0.5 . . . .. −0.5 . .
. .
mean . . .
mean
−2 0 2 IQ −10 0 10 SES
Figure 10.5 Posterior IQ slopes as function of (left) average IQ and (right)

average SES per school. Smooth lowess approximations are indicated by . ◦
Again, the slight deviations do not lead to concerns.
66
Residuals 169–170
Multivariate residuals
The multivariate residual is defined, for level-two unit j, as
Yj − Xj γ̂.
The standardized multivariate residual is defined as

Mj2 = (Yj − Xj γ̂j )0 V̂j−1 (Yj − Xj γ̂j ) .
If all variables with fixed effects also have random effects, then
M 2 = (nj − tj ) s2 + Û 0 {Cov
j j j
d (Ûj )}−1 Ûj ,
where
1
s2j = R̂j0 R̂j , tj = rank(Xj ) .
nj − tj
This indicates how well the model fits to group j.
Note the confounding with level-1 residuals.
If an ill-fitting group does not have a strong effect on the parameter estimates,
then it is not so serious.
67
Residuals 169–170
Deletion residuals
The deletion standardized multivariate residual can be used to assess the fit of
group j, but takes out the effect of this group on the parameter estimates:
2
0 −1
M(-j) = Yj − Xj γ̂(-j) V̂(-j) Yj − Xj γ̂(-j)
where
V̂(-j) = Zj Ω̂(-j) Zj0 + Σ̂(-j) ,
(-j) meaning that group j is deleted from the data for estimating this parameter.
Full computation of deletion estimates may be computing-intensive,

which is unattractive for diagnostic checks.
Approximations have been proposed:
Lesaffre & Verbeke: Taylor series; Snijders & Bosker: one-step estimates.
However, with fast computers
full computation of deletion estimates now also is feasible.
The approximate distribution of multivariate residuals, if the model fits well and
sample sizes are large, is χ2, d.f. = nj .
68
Residuals 169–170
Influence diagnostics of higher-level units

The influence of the groups can be assessed by statistics
analogous to Cook’s distance:
how large is the influence of this group on the parameter estimates?
Standardized measures of influence of unit j on fixed parameter estimates :
F 1 0 −1
Cj = γ̂ − γ̂(-j) ŜF (-j) γ̂ − γ̂(-j)
r
where SF is covariance matrix of fixed parameter estimates, and (-j) means
that group j is deleted from the data for estimating this parameter.
on random part parameters :
R 1 0 −1
Cj = η̂ − η̂(-j) ŜR(-j) η̂ − η̂(-j) ,
p
combined :
1
Cj = rCjF + pCjR .
r + p
69
Residuals 169–170
Values of Cj larger than 1 indicate strong outliers.

Values larger than 4/N may merit inspection.
Table 10.1 the 20 largest influence statistics, and p-values for multivariate
residuals,
of the 211 schools; Model 4 of Chapter 8 but without heteroscedasticity.
School nj Cj pj School nj Cj pj
182 9 0.053 0.293 117 27 0.014 0.987
107 17 0.032 0.014 153 22 0.013 0.845
229 9 0.028 0.115 187 26 0.013 0.022
14 21 0.027 0.272 230 21 0.012 0.363
218 24 0.026 0.774 15 8 0.012 0.00018
52 21 0.025 0.024 256 10 0.012 0.299
213 19 0.025 0.194 122 23 0.012 0.005
170 27 0.021 0.194 50 24 0.011 0.313
67 26 0.017 0.139 101 23 0.011 0.082
18 24 0.016 0.003 214 21 0.011 0.546
70
Residuals 169–170
School 15 does not survive Bonferroni correction: 211 × 0.00018 = 0.038.

Therefore now add the heteroscedasticity of Model 4 in Chapter 8.
Table 10.2 the 20 largest influence statistics, and p-values for multivariate
residuals,
of the 211 schools; Model 4 of Chapter 8 with heteroscedasticity.
School nj Cj pj School nj Cj pj
213 19 0.094 0.010 18 24 0.015 0.003
182 9 0.049 0.352 230 21 0.015 0.391
107 17 0.041 0.006 169 30 0.014 0.390
187 26 0.035 0.009 170 27 0.013 0.289
52 21 0.028 0.028 144 16 0.013 0.046
218 24 0.025 0.523 117 27 0.013 0.988
14 21 0.024 0.147 40 25 0.012 0.040
229 9 0.016 0.175 153 22 0.012 0.788
67 26 0.016 0.141 15 8 0.011 0.00049
122 23 0.016 0.004 202 14 0.010 0.511
71
Residuals 169–170
School 15 now does survive the Bonferroni correction: 211 × 0.00049 = 0.103.
Therefore now add the heteroscedasticity of Model 4 in Chapter 8.
Another school (108) does have poor fit p = 0.00008, but small influence
(Cj = 0.008).
Leaving out ill-fitting schools does not lead to appreciable differences in results.
The book gives further details.
72
11. Designing Multilevel Studies 176–179
11. Designing Multilevel Studies
Note: each level corresponds to a sample from a population.

For each level, the total sample size counts.
E.g., in a 3-level design:
15 municipalities;
in each municipality, 200 households;
in each household, 2-4 individuals.
Total sample sizes are 15 (level 3), 3000 (level 2), ≈ 9000 (level 1).
Much information about individuals and households, but little about municipalities.
73
Power and standard errors

Consider testing a parameter β , based on a t-ratio
β̂
.
s.e.(β̂)
For significance level α and power γ
β
≈ (z1−α + zγ ) = (z1−α − z1−γ ) ,
s.e.(β̂)
where z1−α , zγ and z1−γ are values for standard normal distribution.
E.g., for α = .05, γ = .80, effect size β = .20, sample size must be such that
.20
standard error ≤ = 0.081 .
1.64 + 0.84
The following discussion is mainly in terms of standard errors,
always for two-level designs,
with two-stage samples of N clusters each with n units.
74
Design effect for estimation of a mean

Empty model:
Yij = µ + Uj + Rij .
with var(Uj ) = τ 2, var(Rij ) = σ 2.
Parameter to be estimated is µ.
N n
1 XX
µ̂ = Yij ,
N n j=1 i=1
τ2 σ2
var(µ̂) = + .
N Nn
The sample mean of a simple random sample of N n elements from this
population has variance
τ 2 + σ2
.
Nn
75
The relative efficiency of simple random sample w.r.t. two-stage sample

is the design effect (cf. Kish, Cochran)
nτ 2 + σ 2
= 1 + (n − 1)ρI , (1)
τ2 + σ2
where
τ2
ρI = .
τ 2 + σ2
76
11. Designing Multilevel Studies —
Effect of a level-two variable

Two-level regression with random intercept model
Yij = β0 + β1xj + Uj + Eij .
When var(X) = s2X ,
τ 2 + (σ 2/n)
var(β̂1) = .
N s2X
For a simple random sample from the same population,
disaggregated
τ 2 + σ2
var(β̂1 )= .
N ns2X
Relative efficiency again equal to (1).
77
11. Designing Multilevel Studies –
Effect of a level-one variable

Now suppose X is a pure level-one variable,
i.e., X̄ the same in each cluster, ρI = −1/(n − 1).
Assume var(X) = s2X within each cluster.
E.g.: time effect in longitudinal design; within-cluster randomization.
Again
Yij = β0 + β1xij + Uj + Eij .
78
Now
N X
n
1 X
β̂1 = xij Yij (2)
N ns2X j=1 i=1
N X
n
1 X
= β1 + xij Eij
N ns2X j=1 i=1
with variance
σ2
var(β̂1) = .
N ns2X
For a simple random sample from the same population,
disaggregated
σ2 + τ 2
var(β̂1 )= .
N ns2X
79
The design effect now is

σ2
= 1 − ρI .
τ2 + σ2
Note the efficiency due to blocking on clusters!
80
For random intercept models,

design effect < 1 for level-one variables,
and > 1 for level-two variables.
Conclusion: for a comparison of randomly assigned treatments
with costs depending on N n,
randomising within clusters is more efficient than between clusters
(Moerbeek, van Breukelen, and Berger, JEBS 2000.)
But not necessarily for variables with a random slope!
81
Level-one variable with random slope

Assume a random slope for X:
Yij = β0 + β1xij + U0j + U1j xij + Eij .
The variance of (2) now is
nτ12s2X + σ 2
var(β̂1) = .
N ns2X
The marginal residual variance of Y is
σ 2 + τ02 + τ12s2X
so the design effect now is
nτ12s2X + σ 2
.
τ02 + τ12s2X + σ2
82
A two-stage sample (like the ”within-subject design” in psychology)

‘neutralizes’ variability due to random intercept, not due to random slope of X.
In practice:
if there is a random slope for X then the fixed effect does not tell the whole story
and it is relevant to choose a design in which the slope variance can be estimated.
83
Optimal sample size for estimating a regression coefficient

Estimation variance of regression parameters ≈
σ12 σ22
+
N Nn
for suitable σ12, σ22 .
Total cost is usually not a function of total sample size N n, but of the form
c1N + c2N n .
Minimizing the variance under the constraint of a given total budget
leads to the optimum
s
c1σ22
nopt =
c2σ12
(which still must be rounded).
Optimal N depends on available budget.
84
For level-one variables with constant cluster means,

σ12 = 0
so that nopt = ∞: the single-level design.
For level-two variables,
2 2
τ σ
σ12 = 2 σ22 = 2
sX sX
so that
s
c1 σ 2
nopt = 2
.
c2 τ
Cf. Cochran (1977),
Snijders and Bosker (JEBS 1993),
Raudenbush (Psych. Meth. 1997, p. 177),
Chapter Snijders in Leyland & Goldstein (eds., 2001),
Moerbeek, van Breukelen, and Berger (JEBS 2000),
Moerbeek, van Breukelen, and Berger (The Statistician, 2001).
85
How much power is gained by using covariates?

In single-level regression:
covariate reduces unexplained variance by factor 1 − ρ2 .
Assume random intercept models:
Yij = β0 + β1xij + Uj + Rij

Yij = β̃0 + β1xij + β2zij + Ũj + R̃ij
Zij = γ0 + UZj + RZij .
Also assume that Z is uncorrelated with X within and between groups

(regression coefficient β1 not affected by control for Z).
86
Denote the population residual within-group correlation between Y and Z by

ρW = ρ(Rij , RZij ),
and the population residual between-group correlation by
ρB = ρ(Uj , UZj ) .
The reduction in variance parameters is given by
σ̃ 2 = (1 − ρ2W ) σ 2 ,
τ̃ 2 = (1 − ρ2B ) τ 2 .
In formulae for standard errors, control for Z

leads to replacement of σ 2 and τ 2 by σ̃ 2 and τ̃ 2 .
Therefore:
for pure within-group variables, only within-group correlation counts;
for level-two variables, both correlations count,
but between-group correlations play the major role unless n is rather small.
87
Estimating fixed effects in general

The assumptions in the preceding treatment are very stringent.
Approximate expressions for standard errors were given by Snijders and Bosker
(JEBS 1993)
and implemented in computer program PinT (‘Power in Two-level designs’)
available from
http://www.stats.ox.ac.uk/~snijders/multilevel.htm
Raudenbush (Psych. Meth. 1997) gives more detailed formulae
for some special cases.
Main difficulty in the practical application is the necessity
to specify parameter values:
means,
covariance matrices of explanatory variables,
random part parameters.
88
Example:
sample sizes for therapy effectiveness study.
Outcome variable Y , unit variance, depends on
X1 (0–1) course for therapists: study variable,
X2 therapists’ professional training,
X3 pretest.
Suppose pre-thinking leads to the following guesstimates:
Means: µ1 = 0.4, µ2 = µ3 = 0.
Variances between groups:
var(X1) = 0.24 (because µ1 = 0.4)
var(X2) = var(X3) = 1 (standardization)
ρI (X3) = 0.19 (prior knowledge)
ρ(X1, X2) = −0.4 (conditional randomization)
ρ(X1, X̄3 | X2) = 0 (randomization) ⇒ ρ(X1, X̄3) = 0.2
ρ(X2, X̄3) = 0.5 (prior knowledge).
89
2
This yields σX 3 (W )
= 1 − 0.19 = 0.81 and
 
0.24 −0.20 0.04
ΣX(B) =  −0.20 1.0 0.22  .
 
0.04 0.22 0.19
90
Parameters of the random part:
0
var(Yij ) = β12σX(W
2 2
) + β ΣX(B) β + τ0 + σ
2
(cf. Section 7.2 of Snijders and Bosker, 2012).

Therefore total level-one variance of Y is
β12σX(W ) + σ 2
and total level-two variance is
β 0ΣX(B)β + τ02 .
Suppose that ρI (Y ) = 0.2 and that the available explanatory variables together
explain 0.25 of the level-one variance and 0.5 of the level-two variance.
Then σ 2 = 0.6 and τ02 = 0.10.
Budget structure:
assume c1 = 20, c2 = 1, k = 1000.
91
With this information, PinT can run. Results:
S.E.(β̂1)
0.26 ◦◦◦◦◦
◦◦◦◦◦
0.23 ◦◦◦◦◦◦◦
◦◦◦◦◦◦
0.20 ◦◦◦◦◦◦◦
◦◦◦◦◦◦
∗ ∗∗∗∗∗∗∗∗
0.17 ∗∗ ∗ ∗∗∗∗∗∗∗ ∗
∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗ ∗
0.14
5 10 20 40 n 30
Figure 1 Standard errors for estimating β1 ,
for 20N + N n ≤ 1,000;
∗ for σ 2 = 0.6, τ02 = 0.1;
◦ for σ 2 = 0.5, τ02 = 0.2.
92
For these parameters, 7 ≤ n ≤ 17 acceptable.

Sensitivity analysis: also calculate for σ 2 = 0.5, τ02 = 0.2.
(Residual intraclass correlation twice as big.)
Now 5 ≤ n ≤ 12 acceptable.
93
Estimating intraclass correlation

Donner (ISR 1986):
S.E.(ρ̂I) = (1 − ρI )
s
2
× (1 + (n − 1)ρI ) .
n(n − 1)(N − 1)
Budget constraint:
substitute N = k/(c1 + c2n) and plot as a function of n for various ρI .
E.g., suppose 20 N + N n ≤1000
and 0.1 ≤ ρI ≤ 0.2.
94
S.E.(ρ̂I)
0.15 ∗ ◦
0.125
0.10 ◦
∗
◦
∗◦
0.075
∗◦
∗◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦ ρI = 0.20
0.05 ∗∗∗
∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗ ρI = 0.10
0.025
2 10 20 30 40 n
Figure 2 Standard error for estimating
intraclass correlation coefficient for
budget constraint 20N + N n ≤ 1000
with ρI = 0.1 and 0.2.
Clusters sizes between 16 and 27 are acceptable.

95
Variance parameters
Approximate formulae for random intercept model
(Longford, 1993)
s
2 2 2
S.E.(σ̂ ) ≈ σ
N (n − 1)
and
s
2 1 2 τ02 n τ04
S.E.(τ̂02) ≈ σ 2 + + .
Nn n−1 σ2 σ4
Standard errors for estimated standard deviations:
S.E.(σ̂ 2)
S.E.(σ̂) ≈ ,
2σ
and similarly for S.E.(τ̂0).
Same procedure:
substitute N = k/(c1 + c2n) and plot as function of n.
Example in Cohen (J. Off. Stat., 1998).
96
12.2 Sandwich Estimator for Standard Errors 197–201
12.2 Sandwich estimators

Cluster-robust standard errors for fixed effects can be used to account for
clustering (a nested dependence structure) in the data, even when the fixed effects
are estimated by a model without the proper dependence assumptions.
(This estimation method is called GEE = Generalized Estimating Equations.)
These are also affectionately called sandwich estimator because of their sandwich
form
var(γ̂) = (X 0W X)−1X 0V X(X 0W X)−1 .
Here X is the design matrix for the fixed effects, W is the inverse of the assumed
covariance matrix in the clusters, and V is an estimate of the within-cluster
residual variances.
This way of calculating standard errors does not rely on a particular random effects
specification, nor on normality of the residuals.
This works well for 30 or more groups if the groups are similar in size
and in the structure of the covariates.
97
12.2 Sandwich Estimator for Standard Errors 197–201
The advantage is robustness; the disadvantage is non-optimal efficiency.

Another issue is that the model is incompletely specified: the random part of the
model is ignored.
Considerations about whether or not to do this:
1. Can the research question be answered with the incompletely specified model?
(Cf. the nuisance-interest contrast.)
Are only the fixed parameters of interest?
2. Is the model specification of the random part adequate?
3. Are sample sizes etc. sufficient to make the sandwich estimator a reliable
estimator of variance?
4. Is the loss of efficiency acceptable?
For the latter two questions, the degree of imbalance between the clusters will
be an issue.
98
13. Imperfect Hierarchies 205–206
13. Imperfect Hierarchies
So far we had perfect nesting of lower in higher level units

Non-nested structures:
1. Cross-classified
2. Multiple membership
3. Multiple membership & multiple classification
99
13.1 Cross-classified models 206
Cross-classification
Individuals are simultaneously members of several social units
(such as neighborhood and schools)
This leads to crossed factors (random effects)
Individuals uniquely belong to a combination of both factors
Figure 13.1 Example of pupils nested in schools and neighborhoods
100
Adapted notation
Level 1 unit (say, pupils) indicated by i, simultaneously nested in
• Level 2a (schools) indexed by j and

• Level 2b (neighborhoods) indexed by k
Indicate the neighborhood of a pupil i in school j as k(i, j)
HLM for a non-nested model:

Xq p
X
Yij = γ0 + γh xhij + U0j + Uhj xhij + Rij
h=1 h=1
Add neighborhood random effect W0k assuming

2
W0k ∼ N (0, τW )
and
Cov(W0k , Rij ) = 0, Cov(W0k , U0j ) = 0
101
13.1 Cross-classified models 207, 209
HLM with neighborhood random effect W0k :

Xq p
X
Yi(j,k) = γ0 + γh xhij + U0j + Uhj xhij + W0k + Rij
h=1 h=1
Intra-class correlations:
1. Units i within same j, different k
2
τW
2
τW + τ02 + σ 2
2. Units i within same k, different j

τ02
2
τW + τ02 + σ 2
3. Units i within same j, same k

2
τW + τ02
2
τW + τ02 + σ 2
102
Example: Decomposing complex school effects
Secondary school examination scores of cohort study of Dutch pupils

3,658 pupils in 185 secondary schools coming from 1,292 primary schools
Are primary school effects persistent?
Models:
1. secondary school effects only

2. crossed primary and secondary effects
103
Results for models with and without cross-classified random effects
Model 1 Model 2
Fixed effect Coeff. S.E. Coeff. S.E.

γ0 Intercept 6.36 0.03 6.36 0.03
Random part Var. comp. S.E. Var. comp. S.E.

Crossed random effect:
2
τW = var(W0k ) primary school 0.006 0.005
Level-two random effect:
τ02 = var(U0j ) secondary school 0.067 0.014 0.066 0.014
Level-one variance:
σ 2 = var(Rij ) 0.400 0.010 0.395 0.010
104
Three intra-class correlations
1. Correlation between examination grades of pupils who attended the same

primary school but went to a different secondary school
2
τW 0.006
2 2 2
= = 0.013;
τW + τ0 + σ 0.467
2. Correlation between grades of pupils who attended the same secondary school
but came from different primary schools
τ02 0.066
2 2 2
= = 0.141;
τW + τ0 + σ 0.467
3. Correlation between grades of pupils who attended both the same primary and
the same secondary school
2
τW + τ02 0.072
2 2 2
= = 0.154.
τW + τ0 + σ 0.467
105
13.1 Cross-classified models 209–210
Model with added fixed effects
Model 2 Model 3
γ0 Intercept 6.36 0.03

6.39 0.02
γ1 Pretest 0.032 0.002
γ2 SES 0.068 0.011
γ3 Advice 0.054 0.009
γ4 Ethnicity −0.071 0.028
2
τW = var(W0k ) primary school 0.006 0.005 0.003 0.003
σ 2 = var(Rij ) 0.395 0.010 0.330 0.008
106
13.3 Multiple membership models 210–211
Multiple memberships
Individuals are/have been members of several social units
(such as several different secondary schools)
Only one random factor at level two
(as opposed to two or more factors in cross-classified models)
Individuals belong to multiple levels of that factor
Figure 13.2 Example of pupils nested within multiple schools.
107
13.3 Multiple membership models 211–212
Membership weights
We weight membership by importance
(such as duration of stay in particular school)
PN
Weights wih for each pupil i in school h, with h=1 wih = 1
Example:
1. Pupil 1: only in school 1
2. Pupil 2: equal time in schools 1 and 3
3. Pupil 3: equal time in schools 1, 2, 3
School
Pupil 1 2 3
1 1 0 0
2 0.5 0 0.5
3 0.33 0.33 0.33
108
13.3 Multiple membership models 211-212
Multilevel model for multiple memberships
Denote by Yi{j} an individual who might have multiple memberships

Write HLM with level two residuals U0h weighted by wih
N
X
Yi{j} = γ0 + wih U0h + Ri{j}
h=1
For example, pupil 2 will have

1 1
U01 + U03
2 2
To include a fixed effect for membership trajectory define
1
Wi = P 2
−1
h wih
which is 0 for pupil 1 (with one membership) and positive for others with multiple
memberships
109
13.3 Multiple membership models 213
Example: Decomposition of school effects with multiple memberships
Continuation of previous example
94% of pupils never changed schools
Remaining 6%
• 215 attend two schools (wi vary between 0.2 and 0.8)
• 5 attend three (wi vary between 0.2 and 0.25)
Models:
1. Level 2: only first secondary school

2. Level 2: multiple membership model
110
Results from models without and with multiple membership
Model 4 (ML) Model 5 (MM)
γ0 Intercept 6.36 0.03 6.36 0.03
Level-two random effect:

Level-one variance:
σ 2 = var(Rij ) 0.402 0.010 0.401 0.009
111
Plot of school level residuals from ML and MM
Note: Correlation is 0.86
112
13.4 Multiple membership multiple classification models 213-214
Multiple membership multiple classification models
Combining cross-classified random effects with multiple membership weights
Figure 13.4 Example of pupils nested within multiple schools crossed with
neighborhoods
113
13.4 Multiple membership multiple classification models 214
Model 7 Model 8
γ0 Intercept 6.36 0.03 6.40 0.02

γ1 Pretest 0.031 0.002
γ2 SES 0.068 0.011
γ3 Advice 0.053 0.010
γ4 Ethnicity –0.073 0.028
Crossed random effect:
2
τW = var(W0k ) primary school 0.006 0.005 0.003 0.003
Level-two random MM effect:
Level-one variance:
σ 2 = var(Rij ) 0.395 0.010 0.331 0.008
114
14. Survey weights 216
14. Survey weights

Surveys often are organized using a design in which the sample
is not totally random: e.g., a stratified and/or multi-stage sample.
A major issue is the possibility that at some or all levels,
units are represented with non-constant probabilities.
Reasons can be the efficient organization of data collection,
or the wish to over-represent some interesting subgroups.
This is expressed by sampling weights (proportional to sampling probabilities).
Non-constant sampling weights imply that the sample cannot be
regarded as a ‘mini-population’ – it is not directly representative.
What to do with these weights?
115
What is the purpose of the statistical inference:
1. Descriptive
e.g., estimate mean or proportion in a specific population.
2. Analytic
how do variables of interest depend on other variables?
Here often there is a wish to generalize to a larger population.
116
Connected to this, but a distinct issue is the fact that the use of a probability
model can be founded on different bases:
1. Design-based
This uses the probability distribution that is implied by the sampling design.
This usually is for a sample from a finite population.
The probabilities are under control of the researcher,
except for non-response and other sampling errors.
2. Model-based
The researcher assumes that data can be regarded
as outcomes of a probability model with some unknown parameters.
This usually assumes a hypothetical infinite population,
e.g., with normal distributions.
Assumptions are made about independence etc.
117
14. Survey weights 218–219
Most of quantitative social science research is about analytic inference.

If the model is true (or a good approximation),
and a survey design is used for which
the sample is drawn independently of the residuals in the statistical model,
then the design is irrelevant and taking it into account leads to a loss of efficiency.
But one can hardly ever (or never) be certain that a model is true
and design-based estimators can give protection against
distortions caused by incorrect model assumptions.
If the design is potentially relevant,
then the contrast nuisance ⇔ interesting phenomenon occurs again:
the design may tell is something interesting about the social world.
The approach followed here is that we favor a model-based approach,

taking account of the design when this is called for, or when this is interesting.
118
Two kinds of weights
In statistical modeling, two kind of weights are used.
1. Sampling weights, used in complex surveys, which are the inverses of the
probabilities of including an element in the sample;
i.e., population elements that are undersampled
get a higher weight to correct for the undersampling.
Purpose: unbiased estimation, correct for non-uniform sampling.
2. Precision weights, expressing that some data points are associated with more
precision than others, and therefore get a stronger effect on the results.
Example: multiplicity weights (representing number of cases).
Purpose: lower estimation variance, i.e., increased precision.
119
These kinds of weight differ in their implications for the standard errors.
A survey sample with a few cases with very high weights can lead to high standard
errors, because the results are dominated by these high-weight cases.
The effective sample size, for a sample with weights wi, is
P 2
i wi
neff = P 2
i wi
(applicable for estimating a mean, not for regression).
This is equal to the number of sample elements if wi is constant,
and else it is smaller: loss of efficiency.
The design effect is the ratio of effective sample size
to number of sample elements.
120
When can sampling weights be ignored?
A sample design will depend on a number of design variables,

used for stratification, to define primary sampling units, etc.
The survey weights are a function of the design variables.
If the model is specified correctly given all the design variables , i.e.,
the residuals in the model are independent of the design variables,
then the sample design can be ignored in the analysis.
Therefore, the sample design is important for analytic inference
mainly if we are working with a misspecified model.
121
As an example, consider a study of pupils in schools, Y = language achievement,

with SES as an explanatory variable, and ethnic composition as an omitted
variable.
(ethnic comp.)
6 @
@
@
@
@
@
@
@
? @
R
@
SES -
Y
Suppose that the inclusion probability of schools is associated with the ethnic
composition, but no data about this are available.
Then the analysis effectively uses a SES variable that also represents the
unobserved ethnic composition;
if the analysis is unweighted and sampling is not uniform,
the regression coefficient of SES will average over ethn. comp.
in proportions according to the sample, differring from those in the population.
122
We argue for aiming, when possible, for a model-based analysis,

with the analysis being informed by the design variables.
Arguments:
1. The focus of analysis should be on getting a good model.

2. Most social science researchers are more acquainted with model-based
inference, and have less risk of making mistakes there than in design-based
inference.
3. Methods for design-based inference for multilevel data structures
still have strong limitations.
Thus, the design is used to try and make the model more interesting and
appropriate; a design-based analysis is followed only if the model-based analysis
seems unreliable.
123
How to use design variables in a model-based analysis?
1. Check the variability of weights at each level.
If weights are almost constant, then they may be ignored;

if weights are strongly variable, then using them is risky.
2. Add design variables to the model of interest.
If possible, this is the golden way.

Design variables should be considered also for the random part and in interactions.
Design variables are, however, not always available for the researcher.
3. Apply the hierarchical linear model separately to different parts of the survey.
These parts may depend on stratification variables, on design weights, etc.

Grouping according to design weights is important here,
because the assumption for a model-based analysis is that
the model does not depend on the design weights.
124
4. Add weight variables to the model of interest.
This is an option if the design variables themselves are not available.

If the weight variables themselves do not have a clear interpretation,
the purpose of this is mainly the hope that including the weight variables does not
affect the results for the other variables, which would be an argument for pursuing
a model-based analysis without the weight variables.
5. Carry out model-based and design-based elements within each level-2 unit,
and compare the results.
For a two-level design, this leads to single-level within-group analyses.
The differences between the two kinds of analyses can be inspected and tested.
Again, this is done with the hope of being able to choose for a model-based
approach; if this is not justified, then inspection of the differences can give insight
into how the design is associated with the variables of interest.
125
6. Test differences between the model-based and design based results.
Equation (4.17) gives Asparouhov’s (2006) measure for the informativeness of the
sample design, based on such a comparison:
γ̂hHLM − γ̂hW
I2 = ,
S.E. γ̂hHLM
the difference between the estimates,

expressed in units relative to the model-based standard error.
126
Example: Metacognitive strategies in PISA 2009 for USA
PISA is a large international OECD study about school achievement.

We consider the PISA 2009 data for the USA,
with the dependent variable ‘METASUM’, a measure of awareness of appropriate
strategies for summarizing information.
Sampling design: stratified two-stage sample
Stratification by school type {private, public} × region {MW, NE, S, W}.
Schools are sampled within strata with probabilities proportional to enrollment.
Numbers of sampled schools (left) and sampled pupils (right) per stratum,
for PISA data, USA, 2009.
Midwest Northeast South West
Public schools 38 1,264 26 732 55 1,776 35 1,116
Private schools 2 71 2 46 4 108 3 120
127
There are too few private schools to say much specifically about them.
Level-one weights all are between .95 and 1.00: can be disregarded.
Level-two design effects for PISA data (USA 2009).

Midwest Northeast South West
Public schools 0.49 0.74 0.16 0.17
For South and West, this is reason for concern.
128
Example of a descriptive analysis
Suppose we wish to estimate the proportions of schools in the South located in

villages, towns, and cities.
Note that villages tend to have smaller, and cities larger schools.
Design-based and model-based estimates for three proportions,

for public schools in the South.
Parameter Probability-weighted Unweighted
Proportion of schools situated in a β̂ p S.E.p ȳ S.E.(ȳ)
Village 0.18 0.09 0.11 0.04
Town 0.65 0.13 0.56 0.07
City 0.17 0.07 0.33 0.06
The unweighted results are seriously biased for villages and cities,
and unweighted standard errors are unreliable.
129
Analysis of metacognitive competences

Suppose we wish to study how metacognitive competences depend on the
following variables.
• Gender; female = 0, male = 1.

• Grade, originally ranging from 8 to 12; centered at 10, new range –2 to +2.
• Age, between 15 and 17 years, centered at 16 years.
• ESCS, the PISA index of economic, social and cultural status of students; in
this data set this has mean 0.15 and standard deviation 0.92.
• Immigration status, recoded to:
0, at least one parent born in the USA;
1, second generation (born in the USA, both parents born elsewhere);
2: first generation (born outside the USA, parents likewise).
This ordinal variable is treated as having a linear effect.
• Public versus private schools.
130
To explore the consequences of the design,

first we analyze the data set in five separate parts:
private schools;
schools in the four quartiles of the level-two weights.
131
Estimates for model for metacognitive competence for five parts of the data set:
fixed effects.
Private Weight 1 Weight 2 Weight 3 Weight 4

N (schools) 11 38 39 38 39
Fixed effects Par. S.E. Par. S.E. Par. S.E. Par. S.E. Par. S.E.
Intercept –0.72 0.48 –0.12 0.10 –0.15 0.10 –0.17 0.06 –0.18 0.07
Male –0.27 0.12 –0.34 0.06 –0.33 0.06 –0.28 0.06 –0.35 0.06
Age –0.05 0.23 –0.05 0.12 –0.18 0.12 –0.26 0.11 –0.01 0.12
Grade 0.25 0.12 0.20 0.06 0.24 0.07 0.28 0.06 0.18 0.07
Immigrant –0.13 0.09 0.01 0.04 0.05 0.05 –0.03 0.06 0.06 0.09
ESCS 0.04 0.08 0.10 0.04 0.14 0.04 0.08 0.04 0.11 0.04
Sch-imm 0.51 0.37 0.04 0.14 0.14 0.16 0.10 0.16 –0.29 0.36
Sch-ESCS 0.74 0.40 0.16 0.12 0.11 0.12 0.25 0.13 0.10 0.13
132
Estimates for model for metacognitive competence for five parts of the data set:
variance parameters.
N (schools) 11 38 39 38 39
Variances Var. Var. Var. Var. Var.
School lev. 0.14 0.04 0.06 0.02 0.05

Student lev. 0.85 0.93 0.97 0.95 0.95
133
Conclusions:
Private schools differ strongly from public schools;

the four weight groups differ w.r.t. school-level variables.
Which other school variables are available

that might explain these differences?
A major difference between the weight groups is urbanization.
Next page:
Estimates for model for metacognitive competence, including urbanization, for five
parts of the data set: fixed effects.
Reference category for urbanization is ‘city’.
134
N schools 11 38 39 38 39
Fixed effects Par. S.E. Par. S.E. Par. S.E. Par. S.E. Par. S.E.
Intercept –0.90 0.40 –0.16 0.12 0.03 0.13 0.01 0.13 –0.01 0.23
Male –0.27 0.11 –0.33 0.06 –0.33 0.06 –0.28 0.06 –0.35 0.06
Age –0.07 0.23 –0.04 0.12 –0.19 0.12 –0.27 0.11 –0.01 0.12
Grade 0.26 0.12 0.20 0.06 0.25 0.07 0.28 0.06 0.18 0.07
Immigrant –0.13 0.09 0.01 0.04 0.05 0.05 –0.03 0.06 0.06 0.09
ESCS 0.03 0.08 0.10 0.04 0.14 0.04 0.08 0.04 0.11 0.04
Sch-imm 0.85 0.33 0.11 0.15 –0.00 0.18 0.11 0.17 0.12 0.43
Sch-ESCS 1.06 0.33 0.19 0.13 0.14 0.13 0.21 0.14 0.15 0.14
Large city –0.80 0.26 –0.10 0.12 0.01 0.17 –0.32 0.18 –0.51 0.34
Town –0.36 0.24 0.12 0.12 –0.17 0.13 –0.20 0.13 –0.38 0.26
Small town — 0.02 0.21 –0.42 0.17 –0.22 0.13 –0.24 0.25
Village — — — –0.23 0.17 –0.12 0.24
135
Estimates for model for metacognitive competence, including urbanization,

for five parts of the data set: variance parameters.
N schools 11 38 39 38 39
Variances Var. Var. Var. Var. Var.
School lev. 0.05 0.04 0.05 0.02 0.05

Student lev. 0.85 0.93 0.97 0.95 0.95
136
Conclusion
Differences between weights groups are very small now.
Let us at this stage compare the model-based and design-based estimates.
Next page:
Design-based and model-based estimates for model for metacognitive competence,
entire data set: fixed effects.
I2 is Asparouhov’s (2006) informativeness measure.
137
Design-based Model-based
Fixed effects Par. S.E. Par. S.E. I2
Intercept –0.077 0.089 –0.075 0.058 0.03

Male –0.350 0.040 –0.321 0.029 1.00
Age –0.134 0.102 –0.121 0.055 0.24
Grade 0.192 0.043 0.224 0.031 1.03
Immigrant –0.014 0.031 0.008 0.027 0.81
ESCS 0.075 0.035 0.107 0.019 1.68
Sch-imm 0.240 0.104 0.094 0.068 2.41
Sch-ESCS 0.426 0.119 0.189 0.049 4.84
Private –0.211 0.179 –0.008 0.118 1.72
Large city –0.406 0.121 –0.152 0.081 3.14
Town –0.390 0.111 0.093 0.057 8.47
Small town –0.382 0.108 0.172 0.072 7.69
Village –0.203 0.110 –0.087 0.073 1.59
138

entire data set: variance parameters.
Variances Var. Var.
School lev. 0.031 0.007 0.040 0.009

Student lev. 0.893 0.083 0.941 0.018
139
Conclusion:
Differences remain, mainly with respect to school-level variables.
A residual analysis showed that there is one outlier:
a school with 6,694 pupils enrolled,
while other schools range between 100 and 3,592.
This school and all private schools were excluded,
and square root of school size included as control variable.
Next page:
public schools without outlier: fixed effects.
140
Intercept –0.050 0.068 –0.122 0.061 1.18

Male –0.359 0.046 –0.318 0.032 1.28
Age 0.011 0.080 –0.117 0.059 2.17
Grade 0.162 0.040 0.215 0.032 1.66
Immigrant 0.019 0.034 0.017 0.030 0.07
ESCS 0.090 0.032 0.111 0.019 1.11
Sch-imm 0.114 0.092 0.068 0.078 0.59
Sch-ESCS 0.204 0.081 0.143 0.054 1.13
√
school size − 35 0.004 0.005 0.004 0.003 0.00
Large city –0.193 0.091 –0.094 0.078 1.27
Town –0.094 0.059 –0.032 0.060 1.03
Small town –0.128 0.088 –0.100 0.080 0.35
Village 0.034 0.102 0.023 0.088 0.13
141

public schools without outlier: variance parameters.
Variances Var. Var.
School lev. 0.035 0.010 0.038 0.009

Student lev. 0.939 0.085 0.940 0.018
Conclusion:
Only the coefficient of age still differs between the two types of analysis.
This suggests that age effects depend on the design variables.
This could be, e.g., school enrollment or urbanization.
142
Next page:
public schools without outlier, with more extensive controls: fixed effects.
143
Intercept –0.067 0.065 –0.120 0.061 0.87

Male –0.357 0.045 –0.319 0.032 1.19
Age –0.083 0.065 –0.113 0.058 0.52
Grade 0.167 0.040 0.216 0.032 1.53
Immigrant 0.019 0.034 0.017 0.030 0.07
ESCS 0.090 0.032 0.111 0.019 1.11
Sch-imm 0.108 0.092 0.065 0.060 0.72
Sch-ESCS 0.203 0.081 0.143 0.054 1.11
√
school size − 35 0.002 0.005 0.003 0.003 0.33
Large city –0.194 0.091 –0.093 0.078 1.29
Town –0.096 0.060 –0.033 0.060 1.05
Small town –0.134 0.089 –0.101 0.080 0.41
Village 0.036 0.104 0.023 0.088 0.15
√
Age × ( school size − 35) –0.011
144
0.005 –0.004 0.004 1.75

public schools without outlier, with more extensive controls: variance parameters.
Variances Var. Var.
School lev. 0.036 0.010 0.038 0.009

Student lev. 0.937 0.084 0.940 0.018
Conclusion:
It seems that now we arrived at a good model.
145
Overview: how was the analysis modified by the design variables
1. Public schools differ from private schools,

but the data contains too few private schools to say much about them.
2. One very large school seemed to differ from the rest.
3. Urbanization has an effect on metacognitive competences.
4. Age (controlling for grade!) has an effect,
and interacts with urbanization and/or school size;
it is hard to say which of these, because of correlated effects.
5. For the student-level variables, the final model-based results differ hardly from
the model-based results for the entire data set, controlling for urbanization.
There is a difference w.r.t. the school-level variable ‘proportion of immigrants’,
which may be due to the difference between private and public schools.
Here again there is an issue of correlated effects.
6. The refinement of excluding the very large school and using school size as a
control variable did not importantly modify the other results.
146
15. Longitudinal data 247-248
15. Longitudinal data

Level two: ‘subjects’;
level one: ‘measurements’ made at certain time points.
Multilevel (‘random coefficients’) approach: flexible model specification,
easy treatment of unbalanced data sets (variable n per subject).
1. Fixed occasions :
All subjects measured at the same time points.
2. Variable occasions :
time points can be different; also number of measurements may differ.
(Then there are few alternative approaches.)
An advantage in both cases is the interpretability of the parameters

in terms of rate of increase, etc.
Notation: Yti is measurement at time t for individual i.
t indicates level 1, i indicates level 2.
147
15.1 Fixed Occasions

Simplest case: random intercept model.
This is also called the ‘compound symmetry’ model.
Yti = µt (+ further fixed part) + U0i + Rti.
The subjects differ only with respect to their constant mean deviations U0i,
the time points could have arbitrary means µt .
The residuals are independent.
The intraclass correlation here compares individual differences to
over-time fluctuations, and will be much larger than
the .05–.25 range that is usual for data sets of individuals nested in groups.
Arbitrary differences between time points can be represented
by defining dummy variables for all measurement occasions:
X
µhdhti + U0i + Rti where dhti = 1 for h = t, 0 otherwise
h
(or all minus one, if a constant term — intercept — is included).
148
Example: Life Satisfaction of German 55-60 year olds
1236 respondents, panel data with yearly waves;

Fixed occasions: t = 1, 2, 3, 4, 5, 6 representing ages 55, 56, . . . , 60.
The first model has a constant mean over time,
the second an age-dependent mean.
149
Table 12.1 Estimates for random intercept models

Model 1 Model 2
Fixed effect Coefficient S.E. Coefficient S.E.
µ1 Mean at age 55 6.937 0.044 6.882 0.053

µ2 Mean at age 56 6.937 0.044 6.956 0.054
µ3 Mean at age 57 6.937 0.044 7.021 0.056
µ4 Mean at age 58 6.937 0.044 6.907 0.057
µ5 Mean at age 59 6.937 0.044 6.894 0.059
µ6 Mean at age 60 6.937 0.044 6.985 0.060
Random effect Parameter S.E. Parameter S.E.

Level-two (i.e., individual) variance:
τ02 = var(U0i) 1.994 0.095 1.991 0.094
Level-one (i.e., occasion) variance:
σ 2 = var(Rti) 1.455 0.030 1.452 0.030
Deviance 21,791.34 21,780.70
150
Model 2 is borderline significantly better than Model 2

(χ2 = 10.64, df = 5, p = 0.06),
but the mean differences are not important.
Estimated variances of observations are τ̂02 + σ̂ 2 = 1.994 + 1.455 = 3.449,
estimated within-subjects correlation is
1.994
ρ̂I = = 0.58 .
1.994 + 1.455
Note that these correlations are the same between all measurements of the same
subject, whether they are for adjacent waves or for waves spaced far apart.
If we wish to estimate a person’s long-term life satisfaction µ + U0i,
then the reliability of measurement when using n repeated measurements is
nρI
λi =
1 + (n − 1)ρI
(see Sections 3.5 and 4.8); e.g., for n = 3 this is 0.81.
For this purpose again we can use the posterior means.
151
Here again the assumptions of the random intercept model are quite restrictive.
It is likely that individual respondents differ not only in their mean value over time,
but also in rate of change and other aspects of time dependence.
This is modeled by including random slopes of time,
and of non-linear transformations of time.
Random slope of time:
Yti = µt + U0i + U1i (t − t0) + Rti ;
with a covariate, e.g.,
Yti = µt + αzi + γzi(t − t0) + U0i + U1i (t − t0) + Rti .
152
15. Longitudinal data 254
Here we use as covariates:

birth year (representing cohort & time) and interaction birth year × age.
Birth year date ranges 1929-1951, coded here as
birth year date − 1940

10
so that it ranges between –1.1 and +1.1;
division by 10 to avoid very small coefficients.
Note that age is included non-linearly as a main effect (dummy variables)
and linearly in interaction with birth year.
Linear age (t − t0) ranges from 0 to 5.
The choice of variables implies that the intercept refers to individuals
born in 1940 and now aged 55 years.
153
Model 3 Model 4
µ1 Effect of age 55 6.883 0.055 6.842 0.055

µ2 Effect of age 56 6.955 0.055 6.914 0.055
µ3 Effect of age 57 7.022 0.055 6.988 0.055
µ4 Effect of age 58 6.912 0.056 6.886 0.057
µ5 Effect of age 59 6.906 0.058 6.892 0.060
µ6 Effect of age 60 7.004 0.060 7.005 0.064
α Main effect birth year –0.394 0.078
γ Interaction birth year × age 0.049 0.019
Random effect Parameter S.E. Parameter S.E.
Level-two variation:
τ02 Intercept variance 2.311 0.126 2.253 0.124
τ12 Slope variance age 0.025 0.005 0.025 0.005
τ01 Intercept–slope covariance –0.103 0.021 –0.097 0.021
σ 2 Residual variance 1.372 0.031 1.371 0.031
Deviance 21,748.03 21,722.45
154
Interpretation :
Random slope is significant: χ2 = 32.7, df = 2, p < 0.0001.

Birth year has effect –0.394: older cohorts tend to be happier;
its contribution for −1.1 ≤ B.Y. ≤ 1.1 ranges from 0.43 to –0.43.
This is of medium effect size given the inter-individual s.d.(U0i) = τ̂ 0 = 1.5.
The interaction birth year × age has a somewhat smaller contribution,
ranging between ±(5 × 1.1 × 0.049) = ±0.27.
Deviations from average pattern in rate of change
√
have standard error 0.025 = 0.16.
Since age ranges from 0 to 5, this represents larger variations in rate of change
than those accounted for by birth year.
Negative intercept-slope covariance:
those who start higher (at 55 years) tend to go down
compared to those who start lower.
155
The covariance matrix implied by this model is
 
3.67 2.16 2.06 1.96 1.87 1.77
 2.16 3.50 2.01 1.94 1.87 1.80 
 
 
c
 2.06 2.01 3.38 1.92 1.87 1.82 
Σ(Y ) = 
b
 1.96

 1.94 1.92 3.31 1.87 1.85 

 1.87 1.87 1.87 1.87 3.29 1.88 
 
1.77 1.80 1.82 1.85 1.88 3.33
and the correlation matrix is
 
1.00 0.60 0.58 0.56 0.54 0.51
 0.60 1.00 0.58 0.57 0.55 0.53 
 
 
 0.58 0.58 1.00 0.57 0.56 0.54 
R̂(Y c) = 
 0.56
.
 0.57 0.57 1.00 0.57 0.56 

 0.54 0.55 0.56 0.57 1.00 0.57 
 
0.51 0.53 0.54 0.56 0.57 1.00
156
The variance does not change a lot with age,

and correlations attenuate slightly as age differences increase
(in the matrix, this means going away from the main diagonal);
they are close to the intra-subject correlation
estimated for the compound symmetry model as ρ̂I = 0.58.
157
Multivariate model: unstructured general covariance matrix.

A further possibility available for fixed occasion data
(not in this way for variable occasion data) is the fully multivariate model, which
makes no restrictions on the covariance matrix.
The advantage is that there is less concern about failure of the assumptions.
In some software implementations (e.g., MLwiN), this can be achieved
by giving all dummy variables for measurement occasions random slopes at level 2,
and dropping the random part at level 1.
158
Simplest case: incomplete bivariate data.
This is like the combination of a paired-samples t-test

and an independent-samples t-test.
For the REML estimation method, this method reproduces
the paired-samples t-test if data are complete
and the independent-samples t-test if no respondent has both data points;
but still allowing for different variances at both time points.
The model fitted is
Yti = γ0 + γ1 d1ti + Uti,
where the dummy variable d1 indicates whether or not t = 1
159
Example: comparison of life satisfaction between ages 50 and 60,

for individuals born in 1930-1935 and alive in 1984.
Available respondents: 591; of these,
136 with measurements for both ages,
112 with a measurement only for age 50 (t = 0),
343 with a measurement only for age 60 (t = 1)..
Null hypothesis of no age effects: ‘γ1 = 0’.
160
Estimates for incomplete paired data.
Fixed effect Coefficient S.E.
γ0 Constant term 7.132 0.125

γ1 Effect time 1 –0.053 0.141
Deviance 2,953
Estimated covariance matrix
!
4.039 1.062
Σ(Y
b )=
c
.
1.062 3.210
Age has no significant effect: γ̂1 = −0.053 (t = −0.053/ .141 = 0.38).
161
15. Longitudinal data (not in the book)
Now we estimate Model 5,

which is Model 4 modified to have an unstructured covariance matrix.
Model 4 Model 5
µ1 Effect of age 55 6.842 0.055 6.842 0.055

µ2 Effect of age 56 6.914 0.055 6.913 0.056
µ3 Effect of age 57 6.988 0.055 6.987 0.055
µ4 Effect of age 58 6.886 0.057 6.882 0.059
µ5 Effect of age 59 6.892 0.060 6.886 0.060
µ6 Effect of age 60 7.005 0.064 7.001 0.060
α Main effect birth year –0.394 0.078 –0.395 0.078
γ Interaction birth year × age 0.049 0.019 0.043 0.019
Deviance 21,722.45 21,676.84
162
15. Longitudinal data (not in the book)
The covariance matrices implied by Models 4 and 5 are

Model 4 Model 5
   
3.67 2.16 2.06 1.96 1.87 1.77 3.61 2.11 1.92 1.92 1.84 1.60
 2.16 3.50 2.01 1.94 1.87 1.80   2.11 3.60 2.15 1.95 1.90 1.80 
   

 2.06 2.01 3.38 1.92 1.87

1.82  and 
 1.92 2.15 3.26 1.96 1.73

1.73 
   
 1.96 1.94 1.92 3.31 1.87 1.85   1.92 1.95 1.96 3.49 1.85 1.85 
   
 1.87 1.87 1.87 1.87 3.29 1.88   1.84 1.90 1.97 2.07 3.29 1.87 
   
1.77 1.80 1.82 1.85 1.88 3.33 1.60 1.80 1.73 1.85 1.87 2.88
Although Model 5 is significantly better than Model 4

(χ2 = 45.6, df = 15 − 4 = 11, p < 0.001),
the differences are not important.
163
15.2 Variable occasions

This is a model of populations of curves.
Again, there are random slopes for time and non-linear functions of time.
If there are no variables representing individual × time interactions,
the fixed effects represent the population averages
and the random effects at level two represent
the between-individual differences in the curves.
Example : height of children with retarded growth, 5–10 years old.

1,886 measurements of a total of 336 children.
Height in cm, age in years, intercept at t0 = 5 years.
The time variable is age; children were measured at quite varying ages.
164
First a model of linear growth.
Table 15.7 Linear growth model for 5–10-year-old children with retarded growth.

γ10 Age 5.53 0.08
Random Effect Parameter S.E.
Level-two (i.e., individual) random effects:
τ02 Intercept variance 19.79 1.91
τ12 Slope variance for age 1.65 0.16
τ01 Intercept-slope covariance –3.26 0.46
σ 2 Residual variance 0.82 0.03
Deviance 7099.87
165
Note the small residual standard deviation, σ̂ = 0.9,

which includes measurement error and deviations from linearity.
It follows that growth in this age range is quite nearly linear,
with an average growth rate of 5.53 cm/y,
and a between-individual standard deviation in growth rate of
√
τ̂0 = 1.65 = 1.3 cm/y.
Negative slope-intercept variance:
children who are taller at age 5 grow more rapidly (correlation = –0.63).
Correlation between heights at ages 5 and 10 years is
cov(U0i, U0i + 5U1i)

p
var(U0i) × var(U0i + 5U1i)
19.79 − 5 × 3.26 3.49
= p = = 0.65.
19.79 × (19.79 − 10 × 3.26 + 25 × 1.65) 5.33
166
The ‘predicted’ (postdicted?) curves for the children,

calculated by using empirical Bayes estimates of the parameters:
Next to linear models, polynomial models and spline models can be considered.
Here a simple approach to splines: fixed nodes determined by trial and error.
167
Table 15.8 Cubic growth model for 5–10-year-old children with retarded growth

γ10 t − 7.5 5.23 0.12
γ20 (t − 7.5)2 −0.007 0.038
γ30 (t − 7.5)3 0.009 0.020
Random Effect Variance S.E.

τ12 Slope variance t − t0 2.97 0.32
τ22 Slope variance (t − t0)2 0.255 0.032
τ32 Slope variance (t − t0)3 0.066 0.009
Deviance 6603.75
168
Estimated correlation matrix for level-two random slopes (U0i, U1i, U2i, U3i) :
 
1.0 0.17 −0.27 0.04
0.17 1.0 0.11 −0.84
 
R
bU =   .
 
 −0.27 0.11 1.0 −0.38 
0.04 −0.84 −0.38 1.0
Again ‘predicted’ curves from empirical Bayes estimates – looks almost linear:
169
Another option is to fit piecewise linear function :

continuous functions of age that can have different slopes
for each interval between birthdays years t and t + 1.
170
Table 15.9 Piecewise linear growth model for 5–10-year-old children with retarded growth

τ12 Slope variance f1 3.97 0.82
γ10 f1 (5–6 years) 5.79 0.24
γ20 f2 (6–7 years) 5.59 0.18
γ30 f3 (7–8 years) 5.25 0.16
γ40 f4 (8–9 years) 5.16 0.15
γ50 f5 (9–10 years) 5.50 0.16
Deviance 6481.87
171
Estimated correlation matrix of the level-two random effects (U0i, ..., U5i)
and ‘predicted’ curves :
 
1.0 0.22 0.31 0.14 −0.05 0.09
0.22 1.0 0.23 0.01 0.18 0.33 
 

 
 0.31 0.01 1.0 0.12 −0.16 0.48 
R
bU =   .

 0.14 0.01 0.12 1.0 0.47 −0.23 

−0.05 0.18 −0.16 0.47 1.0 0.03 
 

0.09 0.33 0.48 −0.23 0.03 1.0
172
The advantage of piecewise linear functions is that they are

less globally sensitive to local changes in the function values.
The disadvantage is their angular appearance, with kinks in the “knots”.
A further step is to use splines,
which are a flexible and smooth family of functions that are like polynomials,
but without the global sensitivity to local changes.
Spline functions are polynomials on intervals bounded by points called knots,
and smooth when passing the knots.
We mostly use quadratic or cubic splines.
When the knots have been found with a bit of trial and error,
using the spline function amounts to transforming the time variable
and then further on using linear models – see the book.
173
For example, a cubic spline with one node, positioned at t0,

is defined by the four functions
f1(t) = t − t0 (linear function)

f2(t) = (t − t0)2 (quadratic function)
(
(t − t0)3 (t ≤ t0)
f3(t) = (cubic to the left of t0)
0 (t > t0)
(
0 (t ≤ t0)
f4(t) = (cubic to the right of t0).
(t − t0)3 (t > t0)
Example: growth in the same population of children with retarded growth,

now 12-17 year olds. A total of 321 children with 1,941 measurements.
A good model was obtained for cubic splines with one knot at t0 = 15 years.
174
Table 15.10 Cubic spline growth model for 12–17-year-old children with retarded growth

γ10 f1 (linear) 6.43 0.19
γ20 f2 (quadratic) 0.25 0.13
γ30 f3 (cubic left of 15) −0.038 0.030
γ40 f4 (cubic right of 15) −0.529 0.096
2
Deviance 6999.06
175
Estimated correlation matrix of the level-two random effects (U0i, ..., U4i) :
 
1.0 0.26 −0.31 0.32 0.01
 0.26 1.0
 0.45 −0.08 −0.82 

b U =  −0.31 0.45 1.0 −0.89 −0.71 
R

 .
 0.32 −0.08 −0.89 1.0
 
0.40 
0.01 −0.82 −0.71 0.40 1.0
176
length
170 . . .. . ..................................
. . . ...........................................
. ... .. . . . . . . . . .. . .
. . .. .................... ............................
160 . . . .. . ..... ... . . . ..
. .
... . .
.........................∗ ......∗
....∗
. .....∗.....∗
. .....∗ ....
........ .... . . . . .. . .
...........................∗......∗ . ....∗
. .
....∗ ..∗
. . ..
................ .....
. .. . . .
..
.. . . .... . . ..
.
. . . . . . . ... .. ..
... . . . .. .. .
. .
..
. .... .
.
......................∗ ....∗ ....∗ . ..∗
.
. ...............................................
150 ... ...... ......................∗ ......∗...∗
.....∗ ∗ .............................
..........................................................................∗ .
.......∗.....∗
....
. . .. .
...
..
....... ..........
. .. .
....
. .
............................................∗ ..
.....∗ .....∗
..∗ ..
....∗
..
..∗
.. ∗ . .
..... ......... . . . . . .
. . . . .. . .. .
140 . . .. . .. . .. . . . ..∗. . .. . .
..
..
..
. . . . . . . . . .. ... ..
. .. ...............................
. .
.......∗
............∗ .
.
...∗. .....∗
. . ∗
. . . .. .. .
..................................................................
....∗ . .
.
130 ∗.....∗ .....................................................
............
120
12 13 14 15 16 17
t (age in years)
Figure 15.1 Average growth curve (∗) and 15 random growth curves for
12–17-year-olds for cubic spline model.
Next: explain growth variability by gender and parents’ height (Table 15.11 ).
177

γ10 f1 (linear) 5.85 0.18
γ20 f2 (quadratic) 0.053 0.124
γ30 f3 (cubic left of 15) –0.029 0.030
γ40 f4 (cubic right of 15) –0.553 0.094
γ01 Gender –0.385 0.426
γ11 f1 × gender −1.266 0.116
γ21 f2 × gender –0.362 0.037
γ02 Parents’ length 0.263 0.071
γ12 f1 × parents’ length 0.0307 0.0152
2
Deviance 6885.18
178
Estimated correlation matrix of the level-two random effects (U0i, ..., U4i) :
 
1.0 0.22 −0.38 0.35 0.07
 0.22 1.0
 0.38 −0.05 −0.81 

b U =  −0.38 0.38 1.0 −0.91 −0.75 
R

 .
 0.35 −0.05 −0.91 1.0
 
0.48 
0.07 −0.81 −0.75 0.48 1.0
The correlation matrix is given only for completeness,

not because you can see a lot from it.
179
17. Discrete dependent variables 292–293
17. Discrete dependent variables

Heterogeneous proportions.
Religious attendance at least once a week, 136,611 individuals in 60 countries.
0.0 0.2 0.4 0.6 0.8 1.0

Figure 17.1: Proportion of religious attendance.
Average proportion 0.238.

Difference between proportions: X 2 = 29,733, d.f. = 59, p < 0.0001.
√
τ̂ = 0.0404 = 0.201.
180
17. Discrete dependent variables 294-295
The logarithm transforms a multiplicative to an logit(p)

additive scale and transforms the set of .
...
positive real numbers to the whole real line. 3 ...

...
.....
.....
...
.....
....
.......
One of the most widely used transformations of ....

..
..
...
..
..
2 ...
...
probabilities is the log odds, defined by ..

.
..
..
.
...
...
..
...
...
..
...

p 1 ...
...
.
..
...
.
logit(p) = ln , ..
...
...
...
.
....
1−p ...
...
...
.
...
....
.
0 ...
...
...
.
...
.
.
p
where ln(x) denotes the natural logarithm .....
..
...
...
...
.
....
1
...
...
of the number x. The logit function, of which −1

..
.
...
...
.
...
...
...
.
...
.
graphs are shown here, is an increasing function .

...
...
..
...
...
...
defined for numbers between 0 and 1, and its −2 ...

..
.
...
..
..
...
......
....
......
range is from minus infinity to plus infinity. −3 ..

..
...
......
....
.
...
For example, p = 0.269 is transformed to logit(p) = −1

and p = 0.982 to logit(p) = 4. The logit of p = 0.5 is exactly 0.
181
17. Discrete dependent variables 294-295
p
0.05 0.12 0.27 0.50 0.73 0.88 0.95
.. ... ... ... ... ... ... ... ...
... ... ...
... ... ... ... ... ... ... ...
...
...
... ... ... ... .. ... ... ... ...
.... ... ... ... ... ... ..
. ..
. ...
.. ... ...
...
... ... ... .
... ... ...
... ... ...
...
... ... ... ... ... ...
... ...
... ... ... ... ... ... ... ...
.... ... ... ... ... ... ... ..
. ...
.. ... ...
... ... ... ... ... ... ...
... ... ... ... ... ... ... ... ...
...
... ... ... ...
... ... ... ... ... ...
.... ... ... ... ... ... ... ..
. ...
.. ... ... ... ... ... ... ... ...
... ...
...
...
... ... ... ... ... ... ...
... ... ... ... ... ... ... ... ...
.... ... ... ...
... ... ... .... ..
. ...
.. ... ... ... ... ... ... ... ...
... ... ...
... ... ... ... ... ... ...
... ... ... ... ... ... ... ...
...
...
.... ...
... ... ... ... ... ... ..
. ...
.. ... ... ... ... ... ... ... ...
... ... ...
...
...
... ... ... ... ... ...
... ... ... ... ... ... ... ... ...
.... ... ... ... ... ... ..
. ..
. ...
.. ... ... ... ... ... .
... ... ...
... ... ... ... ... ... ... ... ...
... ...
...
...
...
...
... ... ... ... ... ...
...
.... ... ... ... ... ... ... ..
. ...
.. ... ... ... ... ... ... ... ...
... ... ... ... ... ... ... ... ...
... ... ...
...
... ... ... ... ... ...
.... ... ...
...
... ... ... ... ..
. ...
.. ...
... ... ... ... ... ... ... ...
... ... ... ... ... ... ... ... ...
... ... ... ... ... ... ... ... ...
.... ...
...
... ... ... ... .... ..
. ...
...
.. ... ... ...
... ... ... ... ... ...
... ... ... ... ... ... ... ... ...
... ... ...
... ... ... ... ... ... ...
.... ...
... ... ... ... ... ... ..
. ...
.. ... ... ... ... ... ... ... ...
... ... ... ...
... ... ... ... ... ...
... . .. .. ... ... ... ..
−4 −3 −2 −1 0 1 2 3 4
logit(p)
Correspondence between p and logit(p).
182
Empty multilevel logistic regression model: logit(Pj ) = γ0 + U0j .
Table 17.1 Estimates for empty multilevel logistic model

γ0 = Intercept -1.447 0.180
Random effect Var. Comp. S.E.

Level two variance:
τ0 = S.D.(U0j ) 1.377
.........................................................
.............. ...........
.......... .........
......... .........
.
...
..........
. ........
...
...
..... ........
........
...
.......
. .......
...
.......
. .......
.......
.
...
....... .......
.
...
....... .......
........
...
...
..... ........
...
...
...... ........
........
.
...
........
. .........
...
...
.......
. .........
..........
...
...
...
........ ...........
.............
...
...
...
...
......... ................
.
...
...
...
...
...
...
............. .....................
...............
...........
−5 −2 −1 0 −41 2 −3
logit(p)
Figure 17.5: Observed log-odds and estimated normal distribution
of population log-odds of religious attendance.
183
The random intercept model for binary data:

Xr
logit(Pij ) = γ0 + γh xhij + U0j .
h=1
There is no level-1 residual because Pij is not a random variable,

but itself a group-level random coefficient.
184
Explanatory variables:
• Educational level, measured as the age at which people left school (14–21 years), minus 14.
This variable was centered within countries. The within-country deviation variable has mean 0,
standard deviation 2.48.
• Income, standardized within country; mean −0.03, standard deviation 0.99.
• Employment status, 1 for unemployed, 0 for employed; mean 0.19, standard deviation 0.39.
• Sex, 1 for female, 0 for male; mean 0.52, standard deviation 0.50.
• Marital status, 1 for single/divorced/widowed, 0 for married/cohabiting; mean 0.23, standard
deviation 0.42.
• Divorce status, 1 for divorced, 0 for other; mean 0.06, standard deviation 0.25.
• Widowed, 1 for widowed, 0 for other; mean 0.08, standard deviation 0.27.
• Urbanization, the logarithm of the number of inhabitants in the community or town of
residence, truncated between 1,000 and 1,000,000, minus 10; mean 0.09, standard deviation
2.18.
185
Constructed level-2 variables:
• Average educational level: mean 0.82, standard deviation 0.94.

• Average unemployment: mean 0.19, standard deviation 0.08.
• Average divorce status: mean 0.06, standard deviation 0.03.
• Gini coefficient measuring income inequality, minus 35.
Mean 0.10, standard deviation 9.58.
Next page:
Table 17.2.
Logistic random intercept model for religious attendance in 59 countries.
186
γ0 Intercept –2.069 0.646

γ1 Education (within-country deviation) –0.0290 0.0032
γ2 Income –0.0638 0.0082
γ3 Unemployed 0.017 0.020
γ4 Female 0.508 0.016
γ5 Single –0.269 0.019
γ6 Divorced –0.489 0.036
γ7 Widowed 0.518 0.027
γ8 Urbanization –0.0665 0.0039
γ9 Gini coefficient 0.035 0.017
γ10 Country average education –0.330 0.135
γ11 Country average unemployment 6.033 2.116
γ12 Country average divorce –7.120 4.975
Random effect S.D. S.E.
Random intercept:
τ0 = S.D.(U0j ) intercept standard deviation 1.08
Deviance 115,969.9
187
Random slopes can be added in the usual way

to the multilevel logistic model.
The following is a model for the logit with one random slope:
X r
logit(Pij ) = γ0 + γhxhij + U0j + U1j x1ij .
h=1
188
Table 17.3.
Logistic random slope model for religious attendance in 59 countries.

γ1 Education (within-country deviation) –0.0388 0.0092
γ2 Income –0.0738 0.0161
γ3 Unemployed 0.019 0.020
γ4 Female 0.511 0.016
γ5 Single –0.271 0.019
γ6 Divorced –0.493 0.036
γ7 Widowed 0.482 0.027
γ8 Urbanization –0.0650 0.0040
γ9 Gini coefficient 0.028 0.017
γ10 Country average education –0.333 0.132
γ11 Country average unemployment 5.44 2.06
γ12 Country average divorce –6.43 4.84
(continued...)
189
(... continuation)
Random part parameters S.D. / Corr.
τ0 = S.D.(U0j ) Intercept standard deviation 1.09

τ1 = S.D.(U1j ) Income-slope standard deviation 0.096
τ2 = S.D.(U2j ) Education-slope standard deviation 0.063
ρ01 = ρ(U0j , U1j ) Intercept–income slope correlation 0.29
ρ02 = ρ(U0j , U2j ) Intercept–education slope correlation –0.07
ρ12 = ρ(U1j , U2j ) Income–education slopes correlation 0.27
For a data set with large groups like these country data,
a two-step approach (Section 3.7) might be preferable.
190
Estimated level-two intercept variance may go up when level-1 variables are added
and always does when these have no between-group variance.
This can be understood by threshold model
which is equivalent to logistic regression:
(
0 if Y̆ ≤ 0
Y =
1 if Y̆ > 0 ,
where Y̆ is a latent continuous variable

Xr
Y̆ij = γ0 + γh xhij + U0j + Rij
h=1
and Rij has a logistic distribution, with variance π 2/3.

The fact that the latent level-1 variance is fixed implies that
explanation of level-1 variation by a new variable Xr+1
will be reflected by increase of γh (0 ≤ h ≤ r) and of var(U0j ).
191
Measure of explained variance (‘R2’) for multilevel logistic regression can be based
on this threshold representation, as the
proportion of explained variance in the latent variable.
2
Because of the arbitrary fixation of σR to π 2/3,
these calculations must be based on one single model fit.
192
Let
r
X
Ŷij = γ0 + γh xhij
h=1
be the latent linear predictor; then
Y̆ij = Ŷij + U0j + Rij .
Calculate Ŷij (using estimated coefficients) and then

2
σF = var Ŷij
in the standard way from the data; then

var Y̆ij = σF2 + τ02 + σR 2
2
where σR = π 2/3 = 3.29.
193
The proportion of explained variance now is

2 σF2
Rdicho = .
σF2 + τ02 + 2
σR
Of the unexplained variance, the fraction
τ02
σF2 + τ02 + σR
2
is at level 2, and the fraction

2
σR
σF2 + τ02 + σR
2
is at level one.
194
Table 17.4 Estimates for probability to take a science subject
Model 1
γ1 Gender –1.515 0.102
γ2 Minority status –0.727 0.195
Random Effect Var. Comp. S.E.

Level-two variance:
τ02 = var(U0j ) 0.481 0.082
Deviance 3238.27
Linear predictor
Ŷij = 2.487 − 1.515 genderij − 0.727 minorityij
0.582
has variance σ̂F2 = 0.582 . Therefore Rdicho
2
= 0.582 + 0.481+3.29
= 0.13 .
195
Multicategory ordinal logistic regression

‘Measurement model’:


 0 if Y̆ ≤ θ0

1 if θ0 < Y̆ ≤ θ1 ,
Y =

 k if θk−1 < Y̆ ≤ θk (k = 2, ..., c − 2),

c−1 if θc−2 < Y̆ .

‘Structural model’:
X r
Y̆ij = γ0 + γh xhij + U0j + Rij .
h=1
196
17. Discrete dependent variables 313
Table 17.6 Multilevel 4-category logistic regression model number of science subjects
Model 1 Model 2
Threshold parameters Threshold S.E. Threshold S.E.
θ1 Threshold 1 - 2 1.541 0.041 1.763 0.045
θ2 Threshold 2 - 3 2.784 0.046 3.211 0.054
Fixed Effects Coefficient S.E. Coefficient S.E.

γ0 Intercept 1.370 0.057 2.591 0.079
γ1 Gender girls -1.680 0.066
γ2 SES 0.117 0.037
γ3 Minority status -0.514 0.156
Level two random effect Parameter S.E. Parameter S.E.

τ02 Intercept variance 0.243 0.034 0.293 0.040
deviance 9308.8 8658.2
197
Multilevel Poisson regression
r
X

ln E (Lij ) = γ0 + γh xhij + U0j .
h=1
Lij is a count variable.

For overdispersed counts, negative binomial models
may be more appropriate than Poisson models;
or a Poisson-type model with overdispersion parameter may be used.
Example next page:

Number of memberships of voluntary organizations of individuals in 40 regions in
The Netherlands (ESS data).
198
Table 17.7. Two Poisson models for number of memberships.
Model 1 Model 2
γ0 Intercept 0.846 0.026 0.860 0.032

γ1 Female –0.118 0.044
γ2 (Age − 40)/10 0.198 0.028
γ3 (Age − 40)2 /100 –0.061 0.014
γ4 Protestant 0.296 0.041
γ5 Female × (Age − 40)/10 –0.097 0.030
Level-two random part S.D. S.D. / Corr.
τ0 Intercept standard deviation 0.112 0.070

τ1 Slope S.D. Female 0.146
ρ01 (τ ) Int.–slope correlation 0.168
Deviance 6,658.3 6,522.6
The total contribution of being female is −0.118 − 0.097(Age − 40)/10 + U1j ,

which is 0 for age 28 if U1j = 0.
199
Table 17.8. Two negative binomial models for number of memberships.
Model 3 Model 4
γ0 Intercept 0.848 0.027 0.861 0.034

γ1 Female –0.116 0.046
γ2 (Age − 40)/10 0.197 0.031
γ3 (Age − 40)2 /100 –0.061 0.016
γ4 Protestant 0.300 0.047
γ5 Female × (Age − 40)/10 –0.095 0.033
Level-two random part Parameter Parameter
τ0 Intercept standard deviation 0.111 0.051

τ1 Slope S.D. Female 0.127
ρ01 (τ ) Int.–slope correlation 0.497
α Negative binomial parameter 7.40 10.32
Deviance 6,588.6 6,483.3
200

MULTILEVEL ANALYSIS - Snijders 2012

Uploaded by

Copyright:

Available Formats

MULTILEVEL ANALYSIS - Snijders 2012

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MULTILEVEL ANALYSIS - Snijders 2012

Uploaded by

Copyright:

Available Formats

This is a set of slides following Snijders & Bosker (2012).

2. Multilevel data and multilevel analysis

Each level is (potentially) a source of unexplained variability.

Some examples of units at the macro and micro level:

Figure 2.5 The structure of macro–micro propositions.

Also longitudinal data can be regarded as a nested structure;

Two kinds of argument to choose for a multilevel analysis instead of an OLS

4. The random intercept model

OLS regression model of Y on X ignoring groups :

Distinguish two kinds of fixed effects models:

1. models where group structure is ignored;

In this model, the regression coefficient β1 is common to all the groups.

In the random intercept model, the constant regression coefficient β1 is

In the hierarchical linear model, the U0j are random variables.

Figure 4.1 Different parallel regression lines.

The point y12 is indicated with its residual R12 .

The empty model (random effects ANOVA) is a model

Example: 3758 pupils in 211 schools , Y = language test.

Fixed Effect Coefficient S.E.

Random Part Variance Component S.E.

(Note the difference between fixed effects of explanatory variables

Fixed Effect Coefficient S.E.

Random Part Variance Component S.E.

There are two kinds of parameters:

1. fixed effects: regression coefficients γ (just like in OLS regression);

Table 4.3 Estimates for ordinary least squares regression

Fixed Effect Coefficient S.E.

Random Part Variance Component S.E.

Multilevel model has more structure (“dependence interesting”);

More explanatory variables:

Figure 4.3 Different between-group and within-group regression lines.

Table 4.4 Estimates for random intercept model

Fixed Effect Coefficient S.E.

Random Part Variance Component S.E.

Estimation/prediction of random effects

The posterior mean for group j is based on two kinds of information:

The population information comes from the other groups.

Suppose we wish to predict the ‘true group mean’ γ00 + U0j .

5. The hierarchical linear model

β0j = γ00 + U0j

It is assumed that the group-dependent coefficients (U0j , U1j )

5.1 Estimates for random slope model

Fixed Effect Coefficient S.E.

IQ is defined as the group mean.

.... . .. ............... ...... ......

The next step is to explain the random slopes:

Substitution then yields

Yij = (γ00 + γ01 zj + U0j )

The term γ11 zj xij is called the cross-level interaction effect.

Table 5.2 Estimates for model with random slope

Fixed Effect Coefficient S.E.

Random Part Parameters S.E.

Fixed Effect Coefficient S.E.

Random Part Parameters S.E.

The non-significant parts of the model may be dropped:

Fixed Effect Coefficient S.E.

Random Part Parameters S.E.