Title: 3. Models with Random Effects
13. Models with Random Effects
- 3.1 Error-Components/Random-Intercepts model
- Model, Design issues, GLS estimation
- 3.2 Example Income tax payments
- 3.3 Mixed-Effects models
- Linear mixed effects model, mixed linear model
- 3.4 Inference for regression coefficients
- 3.5 Variance components estimation
- Maximum likelihood estimation, Newton-Raphson and
Fisher scoring, restricted maximum likelihood
(REML) estimation - Appendix 3A REML calculations
23.1 Error components model
- Sampling - Subjects may consist of a random
subset from a population, not fixed subjects - Inference - In the fixed effects models, our
inference deals, in part, with subject-specific
parameters ai . - These parameters are based on the subjects in our
sample. - We may wish to make statements about the entire
population. - In the fixed effects model, because n is
typically large, there are many nuisance
parameters ai .
3Basic model
- The error components model is yit ?i xit ?
?it . - This portion of the notation is the same as the
basic fixed model. However, now the quantities ?i
are assumed to be random variables, not fixed
unknown parameters. - We assume that ?i are independently and
identically distributed (i.i.d) with mean
zero?and variance ???. - We assume that ?i are independent of the error
random variables, ?it . - We still assume that xit is a vector of
covariates, or explanatory variables, and that
??is a vector of fixed, yet unknown, population
parameters. - In the error components model, we assume no
serial correlation, that is, Var ?i ? ? Ii . - Thus, the variance of the ith subject is
- Var yi ?a? Ji ? ? Ii Vi
4Traditional ANOVA set-up
- Without the covariates, this is the traditional
random effects (one way) ANOVA set-up. - This model can be interpreted as arising from a
stratified sampling scheme. - We draw a sample from a population of subjects.
- We observe each subject over time.
- Is there heterogeneity among subjects? One
response is to test the null hypothesis H0 sa2
0. - Estimates of sa2 are of interest but require
scaling to interpret. A more useful quantity to
report is sa2 /(sa2 s 2 ),
the intra-class correlation.
5Sampling
- The experimental design specifies how the
subjects are selected and may dictate the model
choice. - Selecting subjects based on a (stratified) random
sample implies use of the random effects model. - This sampling scheme also suggests that the
covariates are random variables. - Selecting subjects based on characteristics
suggests using a fixed effects model. - In the extreme, each i represents a
characteristic. - Another example is where we sample the entire
population. For example, the 50 states in the US.
6- Figure 3.1. Two-stage random effects sampling.
7Error Components Model Assumptions
- E (yit ?i ) ?i xit? ß.
- xit,1, ... , xit,K are nonstochastic variables.
- Var (yit ?i ) s 2.
- yit are independent random variables,
conditional on a1, , an. - yit is normally distributed, conditional on a1,
, an. - E ?i 0, Var ?i sa 2 and a1, , an are
mutually independent. - ai is normally distributed.
8Observables Representation of the Error
Components Model
- E yit xit? ß.
- xit,1, ... , xit,K are nonstochastic variables.
- Var yit s 2 sa 2 and
- Cov (yir, yis) sa 2, for r ? s.
- yi are independent random vectors.
- yi are normally distributed.
9Structural Models
- What is the population?
- A standard defense for a probabilistic approach
to economics is that although there may be a
finite number of economic entities, there is an
infinite range of economic decisions. - According to Haavelmo (1944)
- the class of populations we are dealing with
does not consist of an infinity of different
individuals, it consists of an infinity of
possible decisions which might be taken with
respect to the value of y. - See Nerlove and Balestras chapter in a monograph
edited by Mátyás and Sevestre (1996, Chapter 1)
in the context of panel data modeling.
10Inference
- If you would like to make statements about a
population larger than the sample, design the
sample to use the random effects model. - If you are simply interested in controlling for
subject-specific effects (treating them as
nuisance parameters), then use the fixed model. - In addition to sampling and inference, the model
design may also be influenced by a desire to
increase the degrees of freedom available for
parameter estimation. - Degrees of freedom
- There are nK linear regression parameters plus 1
variance parameter in the fixed effects model,
compared to only 1K regression plus 2 variance
parameters in the random effects model. - Choose the random effects models to increase the
degrees of freedom available for parameter
estimation.
11Time-constant variables
- If the primary interest is in testing for the
effects of time-constant variables, then, other
things being equal, design the sample to use a
random effects model. - An important example of a time-constant variable
is a variable that classifies subjects by groups - Often, we wish to compare the performance of
different groups, for example, a treatment
group and a control group. - In the fixed effects model, time-constant
variables are perfectly collinear with
subject-specific intercepts and hence are
inestimable.
12GLS estimation
- Expressing the model in matrix form, we have
- E yi Xi b and Var yi Vi sa2 Ji s 2
Ii. - Ji is a Ti Ti matrix of ones, Ii is a Ti Ti
identity matrix. - Here, Xi is a Ti K matrix of explanatory
variables, - Xi (xi1, xi2, ... , xiTi ) .
- The generalized least squares (GLS) equations
are - This yields the error-components estimator of b
- The variance of the error components estimator is
13Feasible generalized least squares
- This assumes that the variance parameters sa2 and
s 2 are known . One way to get a feasible
generalized least squares estimate is - First run a regression assuming Vi Ii,
ordinary least squares. - Use the residuals to determine estimates of sa2
and s 2 . - This estimation procedure yields estimates of sa2
that can be negative, although unbiased. - Determine bEC using the estimates of sa2 and s 2
. - This procedure could be iterated. However,
studies have shown that iterated versions do not
improve the performance of the one-step
estimators. - There are many ways to estimate the variance
parameters - Regardless of how the estimate is obtained, use
it in the GLS estimates. - See Section 3.5 for more details.
14Pooling test
- Test whether the intercepts take on a common
value. That is, do we have to account for
subject-specific effects? - Using notation, we wish to test the null
hypothesis H0 sa2 0. - This is an extension of a Lagrange multiplier
statistic due to Breusch and Pagan (1980). - This can be done using the following procedure
- Run the model yit xit b eit to get residuals
eit . - For each subject, compute an estimator of sa2
- Compute the test statistic,
- Reject H0 if TS exceeds a quantile from an c2
(chi-square) distribution with one degree of
freedom.
153.3 Mixed models
- The linear mixed-effects model is yit zit ?i
xit ? ?it . - This is short-hand notation for the model
- yit ?i1 zit1 ... ?iq zitq ?1 xit1... ?K
xitK ?it - The matrix form of this model is yi Zi ?i Xi
? ?i - The responses between subjects are independent,
yet we allow for temporal correlation through Var
?i Ri. - Further, we now assume that the subject-specific
effects ?i are random with mean zero and
variance-covariance matrix D. - We assume E ?i 0 and Var ?i D , a q ? q
(positive definite) matrix. - Subject-specific effects and the noise term are
assumed to be uncorrelated, that is, Cov (?i ,
?i ) 0. - Thus, the variance of each subject can be
expressed as - Var yi Zi D Zi Ri Vi(?).
16Observables Representation of the Linear Mixed
Effects Model
- E yi Xi ß.
- xit,1, ... , xit,K and zit,1, ... , zit,q are
nonstochastic variables. - Var yi Zi D Zi? Ri Vi(t) Vi.
- yi are independent random vectors.
- yi are normally distributed.
17Repeated measures design
- This is a special case of the linear mixed
effects model. - Here we have i1, ..., n subjects. A response for
each subject is measured based on each of T
treatments. The order of treatments is
randomized. The mathematical model is - The main research question of interest is H0 b1
b2 ... bT, no treatment differences. - Here, the order of treatments is randomized and
no serial correlation is assumed.
18Random coefficients model
- Here is another important special case of the
panel data mixed model. - Take zit xit . In this case the panel data
mixed model reduces to a random coefficients
model, of the form - yit xit(ai b) eit xit bi eit ,
- where bi are random variables with mean b,
independent of eit. - Two-stage interpretation
- 1. Sample subject to get bi
- 2. Sample observations with
- E(yi bi ) Xi bi and Var(yi bi ) Ri.
- This yields E yi Xi b and Var yi Xi D Xi
Ri Vi.
19Variations
- Take columns of Zi to be a strict subset of the
columns of Xi. - Thus, certain components of bi associated with Zi
are stochastic whereas the remaining components
that are associated with Xi but not Zi are
nonstochastic. - Two-stage interpretation
- Use variables Bi such that E bi Bi b.
- Then, we have,
- E yi Xi Bi b and Var yi Ri Xi D Xi.
- This is the random effects model replacing Xi by
Xi Bi and Zi by Xi
20More special cases
- Inclusion of group effects. Take q 1 and zit 1
and consider - yit ai dg xgit b egit ,
- for g 1, ..., G groups, i1, ..., ng subjects
in each group and t1, ..., Tgi observations of
each subject. - Here, ai represent random, subject-specific
effects and dg represent fixed differences
among groups. - This model is not estimable when ai are fixed
effects. - Time-constant variables. We may split the
explanatory variables associated with the
population parameters into those that vary by
time and those that do not (time-invariant).
Thus, we can write our panel data mixed model as - yit zit ai x1i b1 x2it b2 eit
- This model is a generalization of the group
effects model. - This model is not estimable when ai are fixed
effects. - Sec Chapter 5 on multilevel models
21Mixed Linear Models
- Not all models of interest fit into the linear
mixed effects model framework, so it is of
interest to introduce a generalization, the mixed
linear model. - This model is given by y Z ? X ? ? .
- Here, for the mean structure, we assume E (y
?)? Z ? X ? and E ? 0, so that E y ? X ?. - For the covariance structure, we assume
- Var ? R, Var ? D and Cov (? , ? ) 0.
- This yields Var y Z D Z R V.
- This model does not require independence between
subjects. - Much of the estimation can be accomplished
directly in terms of this more general model.
However, the linear mixed effects model provides
a more intuitive platform for examining
longitudinal data.
22Mixed linear model Special cases
- Linear mixed effects model
- Take y (y1,..., yn) , e (e1,..., en), a
(a1, ..., an), X (X1,..., Xn) and Z
block diagonal (Z1,..., Zn) . - With these choices, the model y Z ? X ? ?
is equivalent to the model yi Zi ?i Xi ? ?i
- The two-way error components model is an
important panel data model that is not a specific
type of linear mixed effects model although it is
a special case of the mixed linear model. - This model can be expressed as
- yit ai ?t xit b eit
- This is similar to the error components model but
we have added a random time component, ?t .
233.4 Regression coefficient inference
- The GLS estimator of b takes the same form as in
the error components model with a more general
variance covariance matrix V. - The GLS estimator of b is
- Recall Vi Vi(?) Zi D Zi Ri.
- The variance is
- Interpret bGLS as a weighted average of
subject-specific gls estimators. - bi,GLS is the least squares estimator based
solely on the ith subject - bi,GLS (Xi Vi-1 Xi )-1 Xi Vi-1 yi ,
Wi,GLS Xi Vi-1 Xi
24Matrix inversion formula
- To simplify the calculations, here is a formula
for inverting Vi(t). This matrix has dimension Ti
Ti . - Vi(t) -1 (Ri Zi D Zi ) -1
- Ri -1 - Ri -1 Zi (D-1 Zi Ri -1 Zi ) -1 Zi
Ri -1 - This is easier to compute if
- the temporal covariance matrix Ri has an easily
computable inverse and - the dimension q is smaller than Ti . Because the
matrix (D-1 Zi Ri -1 Zi ) -1 is only a q q
matrix, it is easier to invert than Vi(t) , a Ti
Ti matrix. - For the error components model, this is
25Maximum likelihood estimation
- The log-likelihood of a single subject is
- Thus, the log-likelihood for the entire data set
is - L(b, t ) Si li(b, t ) .
- The values of b, t that maximize L(b, t ) are the
maximum likelihood estimators. - The score vector is the vector of derivatives
with respect to the parameters. - For notation, let the vector of parameters be
- q (b , t).
- With this notation, the score vector is
. - If this score has a root, then the root is the
maximum likelihood estimator.
26Computing the score vector
- To compute the score vector, we first take
derivatives with respect to b and find the root.
That is, - This yields
- That is, for fixed covariance parameters t, the
maximum likelihood estimators and the generalized
least squares estimators are the same.
27Robust estimation of standard errors
- An alternative, weighted least squares estimator,
is - where the weighting matrix Wi,RE depends on the
application at hand. If Wi,RE Vi-1, then bW
bGLS. - Basic calculations show that it has variance
- Thus, a robust estimator of the standard error
is
28Testing hypotheses
- The interest may be in testing H0 ßj ßj,0,
where the specified value ßj,0 is often (although
not always) equal to 0. - Use
- Two variants
- One can replace se(bj,GLS) by se(bj,W) to get
so-called robust t-statistics. - One can replace the standard normal distribution
with a t-distribution with the appropriate
number of degrees of freedom - SAS default is the containment method.
- We typically will have large number of
observations and will be more concerned with
potential heteroscedasticity and serial
correlation and thus will use robust
t-statistics.
29Likelihood ratio test procedure
- Using the unconstrained model, calculate maximum
likelihood estimates and the corresponding
likelihood, denoted as LMLE. - For the model constrained using H0 C ß d ,
calculate maximum likelihood estimates and the
corresponding likelihood, denoted as LReduced. - Compute the likelihood ratio test statistic,
- LRT 2 (LMLE - LReduced).
- Reject H0 if LRT exceeds a percentile from a c2
(chi-square) distribution with p degrees of
freedom. The percentile is one minus the
significance level of the test. - See Appendix C.7 for more details on the
likelihood ratio test.
303.5 Variance component estimation
- Maximum Likelihood
- Iterative estimationNewton-Raphson and Fisher
Scoring - Restricted maximum likelihood (REML)
- Starting values
- Swamys method
- Raos MIVQUE estimators
31Maximum likelihood estimation
- The concentrated log-likelihood is
- Here, the error sum of squares is
- In some cases, one can obtain closed forms
solutions. - In general, this must be maximized iteratively.
32Variance components estimation
- Thus, we now maximize the log-likelihood as a
function of t only. Then we calculate bMLE (t) in
terms of t. - This can be done using either the Newton-Raphson
or the Fisher scoring method. - Newton-Raphson. Let L L(bMLE (t) , t ) , and
use the iterative method - Here, the matrix
- is called the sample
information matrix. - Scoring. Define the expected information matrix
H(t) E ( ) and use -
33Motivation for REML
- Maximum likelihood often produces biased
estimator of variance components. - To illustrate, consider the basic cross-sectional
regression model - Let yi xi b ei , i1, ..., N, where b is a
p ? 1 vector, ei are i.i.d. N(0, s2). - The mle of s2 is (Error SS)/ N, where Error SS is
the error sum of squares from the model fit. - This estimate has expectation s2 (N /(N -p)) and
thus is a biased estimate of s2.
34Further motivation for REML
- As another example, consider our basic fixed
effects panel data model - yit ai xi b eit , where b is a K ? 1
vector, eit are i.i.d. N(0, s2). - As above, the mle of s2 is (Error SS)/N, where
Error SS is the error sum of squares from the
model fit. - This estimator has expectation s2 (N-(nK))/ N
and thus is a biased estimate of s2. - The bias is not asymptotically negligible. To
illustrate, in the balanced design case, we have
NnT and - bias s2 (nT-(nK))/ (nT) - s2
- - s2 (nK)/(nT) _at_ - s2 /T, for large n.
35REML
- The acronym REML stands for restricted maximum
likelihood. - The idea is to consider only linear combinations
of responses y that do not depend on the mean
parameters. - To illustrate, consider the following generic
situation - the responses are denoted by the vector y, are
normally distribution and have mean E y X b and
variance-covariance matrix Var y V(t). - The dimension of y is N ? 1 and the dimension of
X is N ? p . - Suppose that we wish to estimate the parameters
of the variance component, t .
36REML estimation
- Define the projection matrix Q I - X (X X)-1
X. - If X has dimension N ? p, then the projection
matrix Q has dimension N ? N. - Consider the linear combination of responses Q y.
- Some straightforward calculation show that this
has mean 0 and variance-covariance matrix Var y
Q V Q. - Because (i) Q y is normally distributed and (ii)
the mean and variance do not depend on b , this
means that the entire distribution of Q y does
not depend on b. - We could also use any linear transform of Q, such
as A Q . - That is, the distribution of A Q y is also
normally distributed with with a mean and
variance that does not depend on b.
37Modified likelihood
- These observations led Patterson and Thompson
(1971) and Harville (1974) to modify our
likelihood calculations by considering the
restricted maximum log-likelihood - a function of t.
- Here, the error sum of squares is
- For comparison, the usual log-likelihood is
- The only difference is the term ln det(X V(t) X
) thus, methods of maximization are the same
(that is, using Newton-Raphson or scoring).
38Properties of REML estimates
- For the case V s2 I, then the REML estimate
yields the unbiased estimate of s2. - When p, the number of regressors is small, the
MLE and REML estimates of variance components are
similar. - When p, the number of regressors is large, REML
estimates tend to outperform MLE estimates. - The additional term for the longitudinal data
mixed model is
39Starting Values
- Both Swamy and Raos procedures provide useful,
non-recursive, variance components estimates - Raos MIVQUE estimators are available for a
larger class of models (handling serial
correlation, for example) - A version of MIVQUE is the default option in SAS
PROC MIXED for starting values.
40REML versus MLE
- Both are likelihood based estimators
- They applied to a wide variety of models
- They rely on a parametric specification
- For likelihood ratio tests, one should not use
REML. - Use instead maximum likelihood estimators
- Appendix 3A.3 demonstrates the potentially
disastrous consequences of using REML estimators
for likelihood ratio tests.