Auxiliary function providing a user interface for
mra fitting. Typically only used when calling one of the
*.estim routines.
mra.control(algorithm=1, maxfn=1000, cov.meth=1, trace=0, tol=1e-07 )Integer specifying the maximization algorithm to use. If algorithm = 1,
the VA09AD algorithm from the HSL library is used. The VA09AD algorithm is very reliable,
has been tested extensively (same algorithm as Program MARK), and will almost always find the
maximum likelihood estimates. This parameter was added to allow easy addition of other
algorithms in the future, but no other options are currently implemented because VA09AD
works so well. Anything other than 1 throws a warning and resets the value
to 1.
Maximum number of likelihood evaluations allowed during the fitting process.
This determines when the minimization routine stops and concludes that the problem
will not converge. The routine stops after the likelihood has been evaluated
maxfn times within the minimization routine (but not the hessian or gradient routine).
Decreasing this value will cause the fitting routine to give up on badly-behaved
models quicker and return faster. In some simulation where ill-conditioned data
can be generated, decreasing this parameter can speed up the
simulation tremendously.
Increasing this value may allow marginally-behaved (slow to converge) models to converge.
Integer specifying the covariance estimation method. cov.meth = 1
takes numerical 2nd partial derivatives. cov.meth = 2 inverts the Hessian
of the maximization. Method 1 (numeric 2nd derivatives) is the preferred method, but
is computationally expensive and therefore slower than method 2. The difference in
speed is minimal when number of parameters < 15. Method 2 variances are
generally very close to method 1 variances, and could be used during initial model
fitting stages when exact estimation of variance may not be necessary.
Integer controlling output of intermediate results. If trace != 0,
a few lines will be written to the R console and a log file (named mra.log) will be
opened in the current directory and details of the fitting process will
be written there. The level of detail written to the log depends on the value of trace.
If trace == 1, details written to the log include TEST2 and TEST3 results used to determine
c-hat, details from the fitting algorithm, the variance covariance matrix,
singular values of the VC matrix used to determine df, and other
statistics. If trace == 2, all trace == 1 details plus the likelihood output is written
to the log. If trace == 3, all trace == 2 details plus the gradient is written
to the log.
When using algorithm = 1 (VA09AD), trace has additional meaning. If trace > 3,
the coefficient and gradient vectors are written every |trace| iterations and also on exit.
Output is of the form: Function value, beta(1), beta(2), ..., beta(n), G(1), ...G(n).
Coefficient and gradient values are suppressed if trace < 0 (only final results printed).
Intermediate values from within VA09AD are suppressed if trace > maxit + 1, but
other intermediate values are output.
trace > 1 probably produces more output than the casual (non-geek statistician) wants
to look at. But, geek statisticians may find trace > 3 useful for assessing stability and
determining when the routine gets stuck in a local minima.
Vector or scalar of tolerance(s) for coefficients in the model. Minimization stops and concludes it has converged when |delta.b(i)| < tol(i) for all i, where delta.b(i) is change in parameter i on an iteration.
A list with the input arguments as components.
# NOT RUN {
data(dipper.histories)
ct <- as.factor( paste("T",1:ncol(dipper.histories), sep=""))
attr(ct,"nan")<-nrow(dipper.histories)
dipper.cjs <- F.cjs.estim( ~tvar(ct,drop=c(1,2)), ~tvar(ct,drop=c(1,6,7)),
dipper.histories, control=mra.control(trace=1, maxfn=200) )
# }
Run the code above in your browser using DataLab