Multilevel Mixed (hierarchical) models
Christopher F Baum
EC 823: Applied Econometrics
Boston College, Spring 2013
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
1 / 19
Introduction to mixed models
Introduction to mixed models
Stata supports the estimation of several types of multilevel mixed
models, also known as hierarchical models, random-coefficient
models, and in the context of panel data, repeated-measures or
growth-curve models. These models share the notion that individual
observations are grouped in some way by the design of the data.
Mixed models are characterized as containing both fixed and random
effects. The fixed effects are analogous to standard regression
coefficients and are estimated directly. The random effects are not
directly estimated but are summarized in terms of their estimated
variances and covariances. Random effects may take the form of
random intercepts or random coefficients.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
2 / 19
Introduction to mixed models
For instance, in hierarchical models, individual students may be
associated with schools, and schools with school districts. There may
be coefficients or random effects at each level of the hierarchy. Unlike
traditional panel data, these data may not have a time dimension.
In repeated-measures or growth-curve models, we consider multiple
observations associated with the same subject: for instance, repeated
blood-pressure readings for the same patient. This may also involve a
hierarchical component, as patients may be grouped by their primary
care physician (PCP), and physicians may be grouped by hospitals or
practices.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
3 / 19
Introduction to mixed models
Linear mixed models
Linear mixed models
The simplest sort of model of this type is the linear mixed model, a
regression model with one or more random effects. A special case of
this model is the one-way random effects panel data model
implemented by xtreg, re. If the only random coefficient is a
random intercept, that command should be used to estimate the
model.
For more complex models, the command xtmixed may be used to
estimate a multilevel mixed-effects regression. Consider a dataset in
which students are grouped within schools (from Rabe-Hesketh and
Skrondal, Multilevel and Longitudinal Modeling Using Stata, 3rd
Edition, 2012). We are interested in evaluating the relationship
between a students age-16 score on the GCSE exam and their age-11
score on the LRT instrument.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
4 / 19
Introduction to mixed models
Linear mixed models
As the authors illustrate, we could estimate a separate regression
equation for each school in which there are at least five students in the
dataset:
.
.
.
>
use gcse, clear
egen nstu = count(gcse + lrt), by(school)
statsby alpha = _b[_cons] beta = _b[lrt], by(school) ///
saving(indivols, replace) nodots: regress gcse lrt if nstu > 4
command: regress gcse lrt if nstu > 4
alpha: _b[_cons]
beta: _b[lrt]
by: school
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
5 / 19
Introduction to mixed models
Linear mixed models
That approach gives us a set of 64 s (intercepts) and s (slopes) for
the relationship between gcse and lrt. We can consider these
estimates as data and compute their covariance matrix:
. use indivols, clear
(statsby: regress)
. summarize alpha beta
Obs
Variable
Mean
alpha
64
-.1805974
beta
64
.5390514
. correlate alpha beta, covariance
(obs=64)
alpha
beta
alpha
beta
10.833
.208622
Christopher F Baum (BC / DIW)
Std. Dev.
3.291357
.1766135
Min
Max
-8.519253
.0380965
6.838716
1.076979
.031192
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
6 / 19
Introduction to mixed models
Linear mixed models
To estimate a single model, we could consider a fixed-effects approach
(xtreg, fe), but the introduction of random intercepts and slopes for
each school would lead to a regression with 130 coefficients.
Furthermore, if we consider the schools as a random sample of
schools, we are not interested in the individual coefficients for each
schools regression line, but rather in the mean intercept, mean slope,
and the covariation in the intercepts and slopes in the population of
schools.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
7 / 19
Introduction to mixed models
Linear mixed models
A more sensible approach is to specify a model with a school-specific
random intercept and school-specific random slope for the i th student
in the j th school:
yi,j = (1 + 1,j ) + (2 + 2,j )xi,j + i,j
We assume that the covariate x and the idiosyncratic error are both
independent of 1 , 2 .
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
8 / 19
Introduction to mixed models
Linear mixed models
The random intercept and random slope are assumed to follow a
bivariate Normal distribution with covariance matrix:
11 21
=
21 22
Implying that the correlation between the random intercept and slope is
12
Christopher F Baum (BC / DIW)
21
11 22
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
9 / 19
Introduction to mixed models
Linear mixed models
We could estimate a special case of this model, in which only the
intercept contains a random component, with either xtreg, re or
xtmixed, mle.
The syntax of Statas xtmixed command is
xtmixed depvar fe_eqn [ || re_eqn]
[ || re_eqn] [, options]
The fe_eqn specifies the fixed-effects part of the model, while the
re_eqn components optionally specify the random-effects part(s),
separated by the double vertical bars (||). If a re_eqn includes only
the level of a variable, it is listed followed by a colon (:). It may also
specify a linear model including an additional varlist.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
10 / 19
Introduction to mixed models
Linear mixed models
. xtmixed gcse lrt || school: if nstu > 4, mle nolog
Mixed-effects ML regression
Number of obs
Group variable: school
Number of groups
Obs per group: min
avg
max
Wald chi2(1)
Prob > chi2
Log likelihood = -14018.571
gcse
Coef.
lrt
_cons
.5633325
.0315991
Std. Err.
.0124681
.4018891
Random-effects Parameters
P>|z|
45.18
0.08
0.000
0.937
=
=
=
=
=
4057
64
8
63.4
198
=
=
2041.42
0.0000
[95% Conf. Interval]
.5388955
-.7560891
.5877695
.8192873
Estimate
Std. Err.
[95% Conf. Interval]
sd(_cons)
3.042017
.3068659
2.496296
3.70704
sd(Residual)
7.52272
.0842097
7.35947
7.689592
school: Identity
LR test vs. linear regression: chibar2(01) =
Christopher F Baum (BC / DIW)
403.32 Prob >= chibar2 = 0.0000
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
11 / 19
Introduction to mixed models
Linear mixed models
By specifying gcse lrt || school:, we indicate that the
fixed-effects part of the model should include a constant term and
slope coefficient for lrt. The only random effect is that for school,
which is specified as a random intercept term which varies the schools
intercept around the estimated (mean) constant term.
These results display the sd(_cons) as the standard deviation of the
random intercept term. The likelihood ratio (LR) test shown at the foot
of the output indicates that the linear regression model in which a
single intercept is estimated is strongly rejected by the data.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
12 / 19
Introduction to mixed models
Linear mixed models
This specification restricts the school-specific regression lines to be
parallel in lrt-gcse space. To relax that assumption, and allow each
schools regression line to have its own slope, we add lrt to the
random-effects specification. We also add the cov(unstructured)
option, as the default is to set the covariance (21 ) and the
corresponding correlation to zero.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
13 / 19
Introduction to mixed models
Linear mixed models
. xtmixed gcse lrt || school: lrt if nstu > 4, mle nolog ///
> covariance(unstructured)
Mixed-effects ML regression
Number of obs
Group variable: school
Number of groups
Obs per group: min
avg
max
Wald chi2(1)
Prob > chi2
Log likelihood = -13998.423
gcse
Coef.
lrt
_cons
.5567955
-.1078456
Std. Err.
.0199374
.3993155
Random-effects Parameters
P>|z|
27.93
-0.27
0.000
0.787
=
=
=
=
=
4057
64
8
63.4
198
=
=
779.93
0.0000
[95% Conf. Interval]
.5177189
-.8904895
.5958721
.6747984
Estimate
Std. Err.
[95% Conf. Interval]
sd(lrt)
sd(_cons)
corr(lrt,_cons)
.1205424
3.013474
.497302
.0189867
.305867
.1490473
.0885252
2.469851
.1563124
.1641394
3.676752
.7323728
sd(Residual)
7.442053
.0839829
7.279257
7.608491
school: Unstructured
LR test vs. linear regression:
chi2(3) =
443.62
Prob > chi2 = 0.0000
Note: LR test is conservative and provided only for reference.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
14 / 19
Introduction to mixed models
Linear mixed models
These estimates present the standard deviation of the random slope
parameters (sd(lrt))) as well as the estimated correlation between
the two random parameters (corr(lrt,_cons)). We can obtain the
corresponding covariance matrix with estat:
. estat recovariance
Random-effects covariance matrix for level school
lrt
_cons
lrt
_cons
.0145305
.1806457
9.081027
These estimates may be compared with those generated by
school-specific regressions. As before, the likelihood ratio (LR) test of
the model against the linear regression in which these three
parameters are set to zero soundly rejects the linear regression model.
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
15 / 19
Introduction to mixed models
Linear mixed models
The dataset also contains a school-level variable, schgend, which is
equal to 1 for schools of mixed gender, 2 for boys-only schools, and 3
for girls-only schools. We interact this qualitative factor with the
continuous lrt model to allow both intercept and slope to differ by the
type of school:
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
16 / 19
Introduction to mixed models
Linear mixed models
. xtmixed gcse c.lrt##i.schgend || school: lrt if nstu > 4, mle nolog ///
> covariance(unstructured)
Mixed-effects ML regression
Number of obs
=
4057
Group variable: school
Number of groups
=
64
Obs per group: min =
8
avg =
63.4
max =
198
Wald chi2(5)
Prob > chi2
Log likelihood = -13992.533
gcse
Coef.
lrt
.5712245
.0271235
schgend
2
3
.8546836
2.47453
schgend#
c.lrt
2
3
_cons
804.34
0.0000
P>|z|
[95% Conf. Interval]
21.06
0.000
.5180634
.6243855
1.08629
.8473229
0.79
2.92
0.431
0.003
-1.274405
.8138071
2.983772
4.135252
-.0230016
-.0289542
.057385
.0447088
-0.40
-0.65
0.689
0.517
-.1354742
-.1165818
.0894709
.0586734
-.9975795
.5074132
-1.97
0.049
-1.992091
-.0030679
Christopher F Baum (BC / DIW)
Std. Err.
=
=
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
17 / 19
Introduction to mixed models
Random-effects Parameters
Linear mixed models
Estimate
Std. Err.
[95% Conf. Interval]
sd(lrt)
sd(_cons)
corr(lrt,_cons)
.1198846
2.801682
.5966466
.0189169
.2895906
.1383159
.0879934
2.287895
.2608112
.163334
3.43085
.8036622
sd(Residual)
7.442949
.0839984
7.280122
7.609417
school: Unstructured
LR test vs. linear regression:
chi2(3) =
381.44
Prob > chi2 = 0.0000
Note: LR test is conservative and provided only for reference.
The coefficients on schgend levels 2 and 3 indicate that girls-only
schools have a significantly higher intercept than the other school
types. However, the slopes for all three school types are statistically
indistinguishable. Allowing for this variation in the intercept term has
reduced the estimated variability of the random intercept
(sd(_cons)).
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
18 / 19
Introduction to mixed models
Logit and Poisson mixed models
Just as xtmixed can estimate multilevel mixed-effects linear
regression models, xtmelogit can be used to estimate logistic
regression models incorporating mixed effects, and xtmepoisson can
be used for Poisson regression (count data) models with mixed effects.
More complex models, such as ordinal logit models with mixed effects,
can be estimated with the user-written software gllamm by
Rabe-Hesketh and Skrondal (see their earlier-cited book, or
ssc describe gllamm for details).
Christopher F Baum (BC / DIW)
Multilevel Mixed (hierarchical) models
Boston College, Spring 2013
19 / 19
