Various parameters that control fitting of beta regression models
using betareg
.
betareg.control(phi = TRUE, method = "BFGS", maxit = 5000,
gradient = NULL, hessian = FALSE, trace = FALSE, start = NULL,
fsmaxit = 200, fstol = 1e-8, quad = 20, ...)
A list with the arguments specified.
logical indicating whether the precision parameter
phi should be treated as a full model parameter (TRUE
, default)
or as a nuisance parameter.
characters string specifying the method
argument
passed to optim
. Additionally, method = "nlminb"
can be used to employ nlminb
, instead.
integer specifying the maxit
argument (maximal number
of iterations) passed to optim
.
logical or integer controlling whether tracing information on
the progress of the optimization should be produced (passed to optim
).
logical. Should the analytical gradient be used for optimizing
the log-likelihood? If set to FALSE
a finite-difference approximation
is used instead. The default of NULL
signals that analytical gradients
are only used for the classical "beta"
distribution but not for
"xbetax"
or "xbeta"
.
logical. Should the numerical Hessian matrix from the optim
output
be used for estimation of the covariance matrix? By default the analytical solution is employed.
For details see below.
an optional vector with starting values for all parameters (including phi).
integer specifying maximal number of additional (quasi) Fisher scoring iterations. For details see below.
numeric tolerance for convergence in (quasi) Fisher scoring. For details see below.
numeric. The number of quadrature points for numeric
integration in case of dist = "xbetax"
is used in the beta regression.
arguments passed to optim
.
All parameters in betareg
are estimated by maximum likelihood
using optim
with control options set in betareg.control
.
Most arguments are passed on directly to optim
, and start
controls
how optim
is called.
After the optim
maximization, an additional (quasi) Fisher scoring
can be perfomed to further enhance the result or to perform additional bias reduction.
If fsmaxit
is greater than zero, this additional optimization is
performed and it converges if the threshold fstol
is attained
for the cross-product of the step size.
Starting values can be supplied via start
or estimated by
lm.wfit
, using the link-transformed response.
Covariances are in general derived analytically. Only if type = "ML"
and
hessian = TRUE
, they are determined numerically using the Hessian matrix
returned by optim
. In the latter case no Fisher scoring iterations are
performed.
The main parameters of interest are the coefficients in the linear predictor of the
model and the additional precision parameter phi which can either
be treated as a full model parameter (default) or as a nuisance parameter. In the latter case
the estimation does not change, only the reported information in output from print
,
summary
, or coef
(among others) will be different. See also examples.
betareg
options(digits = 4)
data("GasolineYield", package = "betareg")
## regression with phi as full model parameter
gy1 <- betareg(yield ~ batch + temp, data = GasolineYield)
gy1
## regression with phi as nuisance parameter
gy2 <- betareg(yield ~ batch + temp, data = GasolineYield, phi = FALSE)
gy2
## compare reported output
coef(gy1)
coef(gy2)
summary(gy1)
summary(gy2)
Run the code above in your browser using DataLab