Chapter 1 Introduction
3.7 Inference for the marginal model
Inference for the fixed effects βcan be based on the Wald test, t-test,F-test, robust inference or the likelihood ratio (LR) test. Inference for the variance components is based on the Wald test and the LR test. The information cri- teria can generally be useful for making inference about the marginal model.
The estimate ofβ is β(α) = (ˆ
n
X
i=1
Xi0WiXi)−1
n
X
i=1
Xi0Wiyi (3.13) withαbeing replaced by its ML or REML estimate according Harville (1974) and Laird and Ware (1982). Conditional on α, ˆβ(α) is multivariate normal
with mean β and covariance Var( ˆβ) = (
n
X
i=1
Xi0WiXi)−1(
n
X
i=1
Xi0Wi(VarYi)WiXi)(
n
X
i=1
Xi0WiXi)−1
=
n
X
i=1
Xi0WiXi
!−1
(3.14) provided that Wi =Vi−1 where Vi = Var(Yi) =ZiGZi+ Σi.
3.7.1 Approximate Wald test
LetLbe a known contrast or transformation matrix and consider testing the hypothesis
H0 :Lβ=0 versus
HA :Lβ 6=0 (3.15)
Then the Wald test statistic is given by WT =βˆ0L0[L
n
X
i=1
Xi0Vi−1( ˆα)Xi
!
L0]−1Lβˆ
The asymptotic sum distribution ofWT is chi-square distributed with rank(L) degrees of freedom. Thus using the statisticWT inference on fixed effects can be made via the transformation Lβ.
3.7.2 Approximate t-test and F-test
It should be noted that the Wald test is based on var( ˆβ) =
n
X
i=1
Xi0Wi(α)Xi
!−1
The deficiency with the Wald test statistic is that the variability introduced by replacingα by some estimate (ML or REML) is not taken into account in
the subsequent test. Therefore Wald tests will only provide valid inferences in sufficiently large samples. In practice, this is often resolved by replacing the χ2 distribution by an appropriate F distribution. Thus to test the hypothesis H0 versus HA in Eq. (3.14), the above statistic becomes
FT =
βˆ0L0[L Pn
i=1Xi0Vi−1( ˆα)Xi
L0]−1Lβˆ rank(L)
The approximate null distribution of FT is F with numerator degrees of freedom equal to rank(L). The denominator degrees of freedom have to be estimated from the data using common methods such as the containment method, the Sattherwaite approximation and the Kenward and Roger ap- proximation. In the context of longitudinal data, all methods typically lead to large degrees of freedom, and therefore also very similar p-values. For univariate hypotheses, rank(L)=1 and in this case the F-test is equivalent reduces to a t-test. Linear hypotheses of the form given by Eq. (3.14) can be tested in SAS using a CONTRAST statement. The option “chisq” in the CONTRAST statement is needed in order to obtain a Wald test. SAS Proc Mixed also allows the estimation and testing of linear combinations of the elements in β using an ESTIMATE statement. Using similar arguments as for approximate Wald tests, t-tests, and F-tests, approximate confidence intervals can be obtained for such linear combinations, also implemented in the ESTIMATE statement. Specification of L remains the same as for the CONTRAST statement.
3.7.3 Robust Inference
Given the estimate for β in Eq. (3.13) with αreplaced by its ML or REML estimates then conditional on α, ˆβ has the expected value given by,
E[ ˆβ(α)] =
n
X
i=1
Xi0WiXi
!−1 n X
i=1
XiWiE(Yi)
=
n
X
i=1
Xi0WiXi
!−1 n
X
i=1
XiWiXiβ
= β
provided that theE(Yi) =Xiβ. Hence in order for ˆβto be unbiased, it is only sufficient that the mean of the response is correctly specified. Conditional on α, ˆβ has covariance, var( ˆβ) =Pn
i=1(Xi0WiXi)−1 as derived in Eq. (3.14) Var( ˆβ) = (
n
X
i=1
Xi0WiXi)−1
n
X
i=1
(Xi0WiVar(Yi)WiXi)(
n
X
i=1
Xi0WiXi)−1
= (
n
X
i=1
Xi0WiXi)−1
Note that this assumes that the covariance matrix is correctly modelled as Var(Yi) = Vi = ZiGZi0 + Σi and Wi = Vi−1. This form of the covari- ance estimate is therefore often called the ‘naive’ estimate. The so-called robust estimate for Var( ˆβ) which does not assume the covariance matrix to be correctly specified is obtained by replacing Var(Yi) by
Var(Y^i) = [Yi−Xiβ][Yi−Xiβ]0
rather than Vi . The only condition for Var(Y^i) to be unbiased for Var(Yi) is that the mean is correctly specified. The ‘robust’ variance estimate also called the sandwich estimate is now given by
Var( ˆβ) = (
n
X
i=1
Xi0WiXi)−1(
n
X
i=1
Xi0WiVar(Y^i)WiXi)(
n
X
i=1
Xi0WiXi)−1
Based on this sandwich estimate, robust versions of the Wald test as well as of the t-test and the F-test can be obtained. This signifies the point that as long as interest is only in the inferences in the mean structure, little effort should be spent in modelling the exact covariance structure, provided that the data set is sufficiently large. An extreme point of view involves the use of OLS with robust standard errors. Nevertheless appropriate covariance modelling may still be of interest, firstly for the purpose of interpretation of random variation in the data, secondly for gaining efficiency and thirdly because in the presence of missing data, robust inference is only valid under very severe assumptions about the underlying missingness process. Issues of missingness were discussed briefly in Chapter 2 and will be revisited in more detail in Chapter 9. Robust inference for the fixed effects can be obtained by adding the option ‘empirical’ in the PROC MIXED statement in SAS namely
proc mixed data=data1 method=reml empirical;
assuming the data set is ‘data 1’. It is quite possible that for some parameters, the robust standard error is smaller than the naive, model based one. For others the opposite can be true. Thus interpretation of both standard errors should be done with caution.
3.7.4 Likelihood ratio test
The likelihood ratio tests are used to compare nested models with different mean structures, but equal covariance structures. The null hypothesis of interest can therefore be stated as
H0 :β∈Θβ,0
for some subspace Θβ,0 of the parameter space Θβ of the fixed effects β.
Let the notations LM L,ˆθM L,0 and ˆθM L respectively denote the maximum likelihood (ML) function, the maximum likelihood estimator(MLE) underH0 and under the general model. Then the test statistic under the LR method is
−2lnλn =−2ln
"
LM L(ˆθM L,0) LM L(ˆθM L)
# .
The asymptotic distribution of the statistic under the null distribution is χ2 with degrees of freedom (df) equal to the difference in dimension of Θβ and Θβ,0 that is
dimΘβ−dimΘβ,0.
It should be noted that LR tests for the mean structure are not valid under REML. A negative LR test statistic is a very possible outcome under REML.
The reason is as follows: under REML the response Y is transformed into error contrasts U =A0Y, for some matrixA withA0X = 0. Afterwards ML estimation is performed based on error contrasts. Models with different mean structures lead to different sets of error contrasts. Hence the corresponding REML likelihoods are based on different observations, which makes them no longer comparable.