Biomedical Engineering Reference
In-Depth Information
13.4.2
Validation of the \Interval" Package
Although the base R distribution is fairly stable and has its own validation
system (R Foundation for Statistical Computing, 2008), each package needs to
be checked individually for its validity. The intervalR package has been peer
reviewed in the Journal of Statistical Software (Fay and Shaw, 2010). Fay
and Shaw (2010) discuss the package and validation checks, but we mention
a few here. First, the calculation of the scores ci
i
require the estimation of
the NPMLE, S
0
. The interval package uses an E-M algorithm, then checks
the Kuhn{Tucker conditions to make sure the self-consistent estimate is in
fact the NPMLE (see Gentleman and Geyer, 1994). The NPMLE was checked
against theIcenspackage, and for the right-censored case it was checked that
the NPMLE gives the Kaplan{Meier estimates. The permutation part of the
software was checked against thecoinpackage as well as StatXact. Note that
the results for exact tests for right-censored data were tested, but the results
are not expected to match exactly unless the scores are defined precisely the
same way (see Callaert, 2003; Fay and Shaw, 2010). Additionally, because
there are eight methods available, some data sets were checked to see that
the results are reasonably close among all eight methods. Finally, the exact
methods (\exact.ce" and \exact.network") were checked to see that they gave
the same answers on small data sets.
13.5
Simulation with Regular Assessment
Fay and Shih (2012) simulated many situations with both regular and irregular
assessments, with small to moderate sample sizes, and with or without ATD
but all with TIA or CIA. They studied all the tests we study in this section
except the Freidlin et al. (2007) method, and their conclusions for the others
are similar to our conclusions described below. We perform new simulations
Search WWH ::
Custom Search