This function provides initial values to be used when
fitting multivariate covariance generalized linear models by using
the function mcglm. In general the users do not need to use
this function, since it is already employed when setting the argument
control_initial = "automatic" in the mcglm function.
However, if the users want to change some of the initial values,
this function can be useful.
mc_initial_values(linear_pred, matrix_pred, link, variance, covariance, offset, Ntrial, contrasts, data)
| linear_pred | a list of formula see |
|---|---|
| matrix_pred | a list of known matrices to be used on the matrix
linear predictor. |
| link | a list of link functions names, see
|
| variance | a list of variance functions names, see
|
| covariance | a list of covariance link functions names, see
|
| offset | a list of offset values if any. |
| Ntrial | a list of the number of trials on Bernoulli
experiments. It is useful only for |
| contrasts | list of contrasts to be used in the
|
| data | data frame. |
Return a list of initial values to be used while fitting
in the mcglm function.
To obtain initial values for multivariate covariance
generalized linear models the function
mc_initial_values fits
a generalized linear model (GLM) using the function glm with
the specified linear predictor and link function for each response
variables considering independent observations. The family
argument is always specified as quasi. The link function depends
on the specification of the argument link.
The variance function is always specified as "mu" the only
excession appears when using variance = "constant" then the
family argument in the glm function is specified as
quasi(link = link, variance = "constant"). The estimated value
of the dispersion parameter from the glm function is used as
initial value for the first component of the matrix linear predictor,
for all other components the value zero is used.
For the
cases covariance = "inverse" and covariance = "expm"
the inverse and the logarithm of the estimated dispersion parameter
is used as initial value for the first component of the matrix linear
predictor. The value of the power parameter is always started at 1.
In the cases of multivariate models the correlation between response
variables is always started at 0.