Computes the various types of residuals available for
gee
objects.
USAGE:
residuals.gee(object, type="mean")
This is a method for the function
residuals() for objects inheriting
from class
"gee".
Note that as several types of residuals are
available for
gee objects,
there is an additional optional argument
type.
REQUIRED ARGUMENTS:
object
an object inheriting from class
"gee" representing a fitted model.
OPTIONAL ARGUMENTS:
type
a character string indicatung the desired type of residual, with choices
"response" (default),
"pearson",
"symmetric",
"model" or
"score".
VALUE:
Selected residuals of the
gee model.
DETAILS:
type="response"
The response residuals are simply
the
residuals component of
the
"gee" object,
y
minus the fitted values for the response.
They are used in constructing the other types of residuals.
type="pearson"
The Pearson residuals are standardized residuals created by dividing
the response residuals by the estimated observation variance, V.
When the scale is estimated, the sum of squared Pearson residuals adds up to
the degrees of freedom for the regressors (N-p).
type="symmetric",type="model"
The symmetric and model residuals are based on multiplying the residuals
in each cluster by a symmetric matrix
P or a model derived matrix Q,
respectively, where the matrix product P P or Q Q'
is the inverse of the working cluster variance.
P and Q will differ
from cluster to cluster unless the design is balanced.
Q is derived from the definition used in
gee
for the empirical correlation, and Q = S diag(W) where
S S = inverse(R), the inverse of the working correlation matrix,
where 1/(W(i)*W(i)) = V(i).
When V is constant for each cluster the symmetric and
model residuals produce the same values. This occurs when the predictors
are constant for a given subject or the link is identity.
type="score"
The score residuals, S, are derived from the GEE score equation
for the regression coefficients. They are defined so that the sum of
X' S vanishes over all the clusters, where X
is the covariate matrix.
type="deviance"
Deviance residuals are not yet implemented.
SEE ALSO:
,
,
,
EXAMPLES:
Seizure.Subject <- recordDesign(cluster = "Subject", data = Seizure)
gee.out <- gee(y~Time+group, cluster=cbind(clusterID,recordID),
variance="glm.scale", correlation="exchangeable",
family = poisson, link=log, data=Seizure.Subject)
# produce a plot of fitted versus model residuals
plot(fitted(gee.out), residuals(gee.out,type = "model"),
ylab="model residuals")