grouprar 0.2.0
New features
- Multi-arm efficient randomized adaptive design (ERADE): the
DBCD-family designs gain
allocation = "ERADE" and
erade.alpha (Hu, Zhang and He, 2009, for two arms;
Alkhnefr, Hu and Zhai, 2025, for k arms).
- k-arm optimal target allocations with a lower bound
lower.bound on each proportion:
"OptimalNeyman" and "OptimalRSIHR" for binary
responses (Tymofyeyev, Rosenberger and Hu, 2007) and
"OptimalNeyman" for continuous responses. "ZR"
now works for any number of arms and, for two arms, follows rule (7) of
Zhang and Rosenberger (2006).
- Group sequential monitoring for two-arm designs
(
DBCD_Bin(), DBCD_Cont(),
Group.DBCD_Bin(), Group.DBCD_Cont()) through
the new monitor argument, sqMonitor() and
sqBoundary(), with O’Brien-Fleming, Pocock and linear alpha
spending (Zhu and Hu, 2010). Results include the stopping probability at
each look and the expected sample size.
nextAlloc() computes the allocation probabilities (and
optionally the assignments) of the next patient or group in an ongoing
trial.- All designs gain
test.fun (user-supplied test returning
a p-value), typeI (also estimates the type I error under
the null) and seed.
- Results are objects of class
"grouprar" with
print() and summary() methods. They now
include the allocation sequence of every simulated trial
(data: allocation) and the SD of the allocation proportion
of every arm (sd of propotion, previously arm 1 only).
Delayed designs also report the trial duration and the enrollment
duration.
Bug fixes
- The Hu and Zhang (2004) allocation function behind all DBCD designs
now normalizes each arm jointly against all k arms. Previously each arm
was compared with the pooled remaining arms, which agrees with Hu and
Zhang
(2004) only for two arms. Results for two-arm designs are unchanged.
Thanks to Stina Zetterstrom, David S. Robertson and Sofía S. Villar for
the report (StinaZet/RAR-software-bugs#7).
dyldDBCD_Bin() updated its estimates with the responses
of the wrong patients and ignored late responses of the initial
patients. It also failed for more than two arms.Bai.Hu.Shen.Urn() updated the urn with the true success
rates. It now implements the proposed design of Bai, Hu and Shen (2002),
which uses the estimated success rates.dyldDBCD_Cont() ignored the r argument,
failed for more than two arms, did not mark missing responses as
unobserved, and did not update the allocation when a response of one of
the initial patients arrived.dyldDBCD_Bin() and dyldDBCD_Cont() now
recompute the allocation probability for every patient with the current
allocation proportions (previously only when a new response
arrived).Group.dyldDBCD_Bin() and
Group.dyldDBCD_Cont() drew an extra group when the group
sizes added up exactly to ssn.Group.dyldDBCD_Bin() used the wrong column of the
response time matrix, so response times did not depend on the observed
response as documented.DBCD_Bin() ignored mRate, and
Group.DBCD_Cont() failed when mRate was
set.- With missing responses, the allocation proportions used by the
allocation function now count all enrolled patients, as in Section 2.3
of Zhai et al. (2024).
Group.DBCD_Bin() and
Group.DBCD_Cont() previously counted only patients with
observed responses. The reported allocation proportions
(propotion) of all designs with mRate also
count all enrolled patients now; the failure rate and the test still use
the observed responses only.
DBCD_Bin() and DBCD_Cont() now use the
chi-squared test for more than two arms, and the chi-squared test for
continuous responses uses the sample variance of each arm.- The tests no longer fail for two arms (
PolyaUrn(),
WeiUrn()), for arms without patients, or for arms without
variability. For binary responses with more than two arms, arms without
variability no longer make the test NA: the Wald test then uses the
adjusted variances of Agresti and Caffo (2000).
GDLRule() returns one failure rate per simulation, and
GDLRule(), DLRule() and
BirthDeathUrn() no longer fail when aK or
Y0 are not whole numbers (balls are drawn in proportion to
the positive part of the urn).rspT.dist = "uniform" was rejected because of a
typo.- Allocation proportions are no longer misaligned when an arm has no
patients.
- Equal allocation is used while the target cannot be estimated
(e.g. fewer than two responses in an arm of a continuous design).
RPWRule() stops with an error for k != 2,
the DBCD-family designs check that n0 >= k and
ssn > n0, and the group designs check that
gsize.param > 0 (a rate of 0 made the initial phase loop
forever). All designs check k >= 2, Y0,
aK and the length of rspT.param.
Documentation
- Corrected return values, argument descriptions, references and
examples in all help pages. Added a citation file.
Other changes
- Removed the unused
ggplot2, gridExtra,
methods and tidyr imports, and the dependency
on stringr, whose current version needs R >= 4.1.
- Added unit tests and a GitHub Actions workflow for R CMD check.