Methods to estimate the worst fraction of missing information, and the
direction of the worst linear function of parameters for class
missmodel objects with a
paramIter component inheriting from
classes
cgm,
Gauss,
Loglin, and produced by an EM
algorithm.
USAGE:
worstFraction.cgm(object, method = c("crude", "power"),
tol = c(1e-12, 1e-06), lim = c(20, 20), h = 0.01, d = 10)
worstFraction.Gauss(object, method = c("crude", "power"),
tol = c(1e-12, 1e-06), lim = c(20, 20), h = 0.01, d = 10)
worstFraction.Loglin(object, method = c("crude", "power"),
tol = c(1e-12, 1e-06), lim = c(20, 20), h = 0.01, d = 10)
REQUIRED ARGUMENTS:
object
a class
"missmodel" object produced by an EM algorithm.
OPTIONAL ARGUMENTS:
method
character, the method to be used for the computations.
The
"crude" method is the
default. In this method, the worst
fraction of missing information is the
median of the elementwise rates of convergence (the ratio of
two consecutive parameter differences, calculated at
convergence of the EM algorithm).
The direction of the worst linear function of
the parameters is computed as the normalized difference between the last
two iterates in the EM algorithm.
For method
"power", the worst linear function of the parameters is computed as the
largest eigenvector for the matrix of first partial derivatives of the EM
operator. In this method the jacobian is never actually computed.
Rather, the eigenvector and eigenvalue are computed using the power
method by repeated direct application of the EM operator in the
direction of the estimated eigenvalue.
tol
vector of length 2 of tolerance parameters.
tol[1] is the tolerance
used when computing products of the EM operator with arbitrary
vectors.
tol[2] is used to assess convergence of the estimated
eigenvalue. When the relative change in the eigenvalue is less than
tol[2], convergence is assumed. The default values are usually
satisfactory. In general, be sure that
tol[1] << tol[2].
lim
the number of extrapolation steps allowed in computing products of the
EM operator. The default values are usually satifactory.
h
The initial fractional step to be used in computing products of the EM
operator. The default value is usually satisfactory. If the
following error message occurs, then try using a smaller value of
h.
"Estimate outside the parameter space. Check prior."
d
the next smaller value of h to be tried in computing products of the
EM operator. The default value is usually satisfactory.
VALUE:
see the help file for
for a discussion of the returned
values and details.
SEE ALSO:
,
,
.
EXAMPLES:
fit <- emLoglin(object = crime, frequency = count)
worstFraction(fit)
worstFraction(fit, method = "power")