Specify a Mixed Models to be fit with Generalized Estimating Equations

DESCRIPTION:

Createa a ranDesign object for random effect Z*b where Z is a covariate matrix of {1, z2, z2, ..., zn} and b is a vactor of random-effects parameters coming from a normal distribution (0, D), where D is a diagonal matrix.

USAGE:

ranDesign(formula, variance = "independent", method = "LS", fixed = NULL)

REQUIRED ARGUMENTS:

formula
a one-sided formula of the form ~z1+...zn specifying the model for the random effects.

OPTIONAL ARGUMENTS:

variance
a character string specifing the variance structure of random variables. Presently, it is set to "independent" only.
fixed
a numeric vector if the variance of random effects are fixed, or NULL if the variance of random effects need to be estimated. The default is set to NULL.
method
a character string, either "LS" for least square method or "moment" for moment method, specifying the method to estimate the scale parameters and variances of random effects. The default is set to "LS".

VALUE:

an object of class ranDesign is returned which includes the following components:
`terms'
a terms object constructed from the formula.
`variance'
a character string of "independent" telling the variance structure of random variables.
`fixed'
a numeric vector for fixed variance of random effects or NULL if the variance of random effects need to be estimated.
`method'
a character string telling the estimation method for scale and variances of random effects.

SIDE EFFECTS:

For mixed models, GEE methods work on the marginal mean and variance. The marginal mean might have an extra offset added or multiplied to the linear predictor of the conditional mean. The marginal variance consists of a set of variance components. GEE is not a likelihood method. Except Gaussian case, applying GEE methods to mixed models might not result in the same fit as applying with the glme . See Zeger, Liang and Albert (1988) for more details. Currently, this function is applicable only to log, logit and probit links for Poisson and binomial data with limited structure of variance components, e.g. independent variance components. The best way to check its validity is through a simulation study.

REFERENCES:

Zeger, Linag and Albert (1988). Models for longitudinal data: a generalized estimating equation approach. Biometrics, 44, 1049-1060.

SEE ALSO:

, .

EXAMPLES:

r1 <- ranDesign(~z1+z2)

seizure.re <- ranDesign(~Time)

seizure.Des <- geeDesign(y~Time, cluster=Subject, variance="glm.scale", 
  correlation="exchangeable", random=seizure.re, family=poisson, data=Seizure)

gee.fit(seizure.Des)