Worst Fraction of Missing Information

DESCRIPTION:

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")