To carry out a score test for a GLM, we first fit a "base" model using
the standard iteratively reweighted least squares (IRLS) algorithm and
then carry out a score test for addition of further terms. This
function sets various control parameters for this.
Usage
glm.test.control(maxit, epsilon, R2Max)
Arguments
maxit
Maximum number of IRLS steps
epsilon
Convergence threshold for IRLS algorithm
R2Max
R-squared limit for aliasing of new terms
Details
Sometimes (although not always), an iterative scheme is necessary to fit
the "base" generalized linear model (GLM) before carrying out a score
test for effect of adding new term(s). The maxit parameter sets
the maximum number of iterations to be carried out, while the
epsilon parameter sets the criterion for determining
convergence. After fitting the base model, the new terms are added, but
terms judged to be "aliased" are omitted. The method for determining
aliasing is as follows (denoting the "design" matrix for the additional
terms by Z):
Step 1Regress each column of Z on the base model matrix,
using the final GLM weights from the base model fit, and replace
Z with the residuals from these regressions.
Step 2Consider each column of the new Z matrix in turn,
regressing it on the previous columns (again using the weights
from the base model fit). If the proportion of the weighted sum of
squares "explained" by this regression exceeds R2Max, the term
is dropped and not included in the test,
The aim of this procedure to avoid wasting degrees of freedom on columns
so strongly aliased that there is little power to detect their effect.
Value
Returns the parameters as a list in the expected order