literature review september–december 2004
TRANSCRIPT
PHARMACEUTICAL STATISTICS
Pharmaceut. Statist. 2005; 4: 77–79
Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/pst.159
Literature Review September–December 2004
Simon Day1,*,y and Meinhard Kieser2
1Medicines and Healthcare products Regulatory Agency, Room 13-205, Market Towers,
1 Nine Elms Lane, London SW8 5NQ, UK2Department of Biometry, Dr Willmar Schwabe Pharmaceuticals, Karlsruhe, Germany
INTRODUCTION
This review covers the following journals received during the
period from the middle of September 2004 to the end of
December 2004:
* Biometrical Journal, volume 46, part 5.* Biometrika, volume 91, part 3.* Biostatistics, volume 5, part 4.* Clinical Trials, volume 1, part 5.* Communications in Statistics – Simulation and Computation,
volume 33, part 3.* Communications in Statistics – Theory and Practice, volume
33, parts 7–10.* Computational Statistics & Data Analysis, volume 47, parts
1–4.* Drug Information Journal, volume 38, part 4.* Journal of Biopharmaceutical Statistics, volume 14, parts 3, 4.* Journal of the American Statistical Association, volume 99,
part 3.* Journal of the Royal Statistical Society, Series A, volume
167, part 4.* Statistics in Medicine, volume 23, parts 20–24.* Statistical Methods in Medical Research, volume 13,
parts 5, 6.* The American Statistician, volume 58, part 4.
SELECTED HIGHLIGHTS FROM THE
LITERATURE
The themes of Statistical Methods in Medical Research were:
* Part 5: Cluster analysis (pages 343–408).* Part 6: Developing and comparing population models for
early detection of cancer (pages 419–538).
Part 24 of Statistics in Medicine contains papers from the
joint meeting of the International Society for Clinical Biosta-
tistics and the Society for Clinical Trials, held in London in July
2003. Two papers are worthy of mention here. The first is by
Garcia et al. who have reviewed the literature on published
cross-over trials and consider both the relative efficiency of a
cross-over compared to a parallel group design and the
reporting standards of cross-over trials compared to the
CONSORT statement. The second paper is by Stephen Senn
– although nothing to do with cross-over studies. This was the
President’s invited keynote lecture. It is full of insight, history
and general interest:
* Garcia R, Benet M, Arnau C, Cobo E. Efficiency of the
cross-over design: an empirical investigation. Statistics in
Medicine 2004; 23:3773–3780.* Senn S. Added values. Controversies concerning randomi-
zation and additivity in clinical trials. Statistics in Medicine
2004; 23:3729–3753.
Phase I
The paper by Whitehead et al. evaluates Bayesian procedures
for dose escalation studies with two binary responses. In a
comprehensive simulation study, various designs are considered
differing by prior opinion, criterion for optimal allocation of
doses, policy for early stopping, and dose schedule. The results
lead to general recommendations for the design of such studies.
* Whitehead J, Zhou Y, Stevens J, Blakey G. An evaluation
of a Bayesian method of dose escalation based on bivariate
binary responses. Journal of Biopharmaceutical Statistics
2004; 14:969–983.
Multiplicity
The following two papers consider active control trials where
the goal is to show non-inferiority or equivalence for multiple
endpoints. Kong et al. derive a test to demonstrate non-
inferiority or equivalence on each component of the endpoint
vector and conduct simulation studies to investigate its
performance. Tamhane and Logan address a slightly different
Copyright # 2005 John Wiley & Sons, Ltd.Received \60\re /teci
yE-mail: [email protected]
*Correspondence to: Simon Day, Medicines and Healthcareproducts Regulatory Agency, Room 13-205, Market Towers,1 Nine Elms Lane, London SW8 5NQ, UK.
test problem, namely to show that the experimental treatment is
not inferior on any of the endpoints and superior on at least one
(or some specified number of) endpoint(s). They propose a new
test and compare it with existing methods with respect to Type I
error control and power performance. The new test is especially
convenient if, after rejection of the global null hypothesis,
follow-up decisions on the single hypotheses concerning the
endpoints shall be made, which is usually the case.
* Kong L, Kohberger RC, Koch GG. Type I error and power
in noninferiority/equivalence trials with correlated multiple
endpoints: an example from vaccine development trials.
Journal of Biopharmaceutical Statistics 2004; 14:893–907.* Tamhane AC, Logan BR. A superiority-equivalence
approach to one-sided tests on multiple endpoints in clinical
trials. Biometrika 2004; 91:715–727.
Sample size calculation and recalculation
There is often debate on whether or not ‘responder analyses’ are
an adequate method to address the question of clinical
relevance. Stephen Senn and John Lewis exchanged the
arguments for and against in recent issues of this journal.
However, no matter how one answers this question, such
additional analyses are often required. To guard against any
unpleasant surprises in the analysis, it is essential that the study
has sufficient power not only for the significance test based on
the primary variable but also for the fulfilment of the responder
criterion defined for the dichotomized variable. The following
paper presents sample size calculation methods for this
situation and the cases of one or multiple endpoint(s) which
may be normally distributed or not.
* Kieser M, R .oohmel J, Friede T. Power and sample size
determination when assessing the clinical relevance of trial
results by ‘responder analysis’. Statistics in Medicine 2004;
23:3287–3305.
Ruvuna addresses the potential loss in power that may occur
in multicentre trials due to unequal centre sizes. For the
unweighted (Type III) analysis, a measure of imbalance is
derived. Employing this coefficient in the sample size calcula-
tion prevents an underestimation of the required number of
subjects in the case of extreme centre size imbalances.
* Ruvuna F. Unequal center sizes, sample size, and power in
multicenter clinical trials. Drug Information Journal 2004;
38:387–394.
The following paper could mark the beginning of a
wonderful friendship between statisticians and data managers.
Based on statistical considerations, sample size formulae are
derived that determine the number of items to be checked
during a database audit to ensure that the original data are
converted with acceptable quality. Simple random sampling
(item selection without any restriction) and cluster sampling (all
items from selected case record form books) are discussed, and
sample size tables and SAS programs, respectively, are given.
Implementation will obviate the need for data managers to
carry out a 100% audit of the database and will provide high-
quality data for the statisticians’ analyses – a classic win–win
situation!
* Zhang P. Statistical issues in clinical trial data audit. Drug
Information Journal 2004; 38:371–386.
Study design
Raab et al. consider timing (and frequency) of follow-up
observations for estimating median and mean survival times
when interval censoring is taking place. The focus is on equally
spaced (in time) observation points. If one of the observation
times is planned around the anticipated median, this may be
optimal. If the total follow-up time is kept constant but the
number of sampling times is changed, and the new sampling
times remain equally spaced, then the one near the anticipated
median will move and efficiency of the estimated median may
be reduced. This even happens if the number of sampling times
increases.
* Raab GM, Davies JA, Salter AB. Designing follow-up
intervals. Statistics in Medicine 2004; 23:3125–3137.
Data analysis issues
Fedorov et al. investigate the performance of three fixed effect
estimates of the treatment difference in multicentre trials under
random enrolment. In their simulation study, five different
enrolment schemes are considered that mimic a wide spectrum
of situations encountered in practice. A comforting result of
their work is that the estimator derived from the simplest model
works well (and even better than the more complex ones) in
many situations.
* Fedorov V, Jones B, Jones M, Zhigljavsky A. Estimation of
the treatment difference in multicenter trials. Journal of
Biopharmaceutical Statistics 2004; 14:1037–1063.
Darken reviews methods to evaluate the validity of least
squares analysis with the general linear model. Furthermore,
robustness with respect to these assumptions is investigated in a
simulation study and regulatory issues are discussed.
* Darken PF. Evaluating assumptions for least squares
analysis using the general linear model: a guide for the
pharmaceutical industry statistician. Journal of Biopharma-
ceutical Statistics 2004; 14:803–816.
The combination of p-values from independent tests is,
among others, an issue in meta-analysis, multiple testing and
group sequential adaptive designs. Loughin compares six such
methods with respect to their rejection region and power. The
Copyright # 2005 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2005; 4: 77–79
LITERATURE REVIEW78
results are summarized in recommendations concerning which
method to use in which situation.
* Loughin TM. A systematic comparison of methods for
combining p-values from independent tests. Computational
Statistics & Data Analysis 2004; 47:467–485.
When using random effects in modelling, the distribution of
the effects is often chosen for computational convenience.
Agresti et al. show three cases (random effects model for
proportions and log-odds ratio and frailty model for survival)
in which misspecification of random effects distributions can
reduce the efficiency considerably. Proposals are reviewed that
describe how to address misspecification issues.
* Agresti A, Caffo B, Ohman-Strickland P. Examples in
which misspecification of a random effects distribution
reduces efficiency, and possible remedies. Computational
Statistics & Data Analysis 2004; 47:639–653.
Paul Meier et al. examine the performance of the Kaplan–
Meier survival estimator relative to parametric models. As
reflected in the article’s title, the results indicate that the price to
be paid when using the nonparametric approach is generally
small, if any.
* Meier P, Karrison T, Chappell R, Xie H. The price of
Kaplan–Meier. Journal of the American Statistical Associa-
tion 2004; 99:890–896.
Lin and Wang propose a new test for the comparison of two
survival curves. This test has greater power than the commonly
used log-rank test and the Wilcoxon test if the two groups do
not have proportional hazard ratios but when the curves are
close at the beginning and then separate.
* Lin X, Wang H. A new testing approach for comparing the
overall homogeneity of survival curves. Biometrical Journal
2004; 46:489–496.
In contrast to asymptotic tests, unconditional exact tests
comparing two independent proportions never exceed the
preassigned nominal level. Furthermore, these tests are
generally more powerful than conditional exact tests. Skipka
et al. compare various unconditional exact tests for proving
non-inferiority or superiority with respect to power, size and
computational time. The SAS code for all tests can be obtained
from the authors upon request, which may further stimulate the
application of these tests in two-armed trials with binary
outcomes.
* Skipka G, Munk A, Freitag G. Unconditional exact tests
for the difference of binomial probabilities – contrasted and
compared. Computational Statistics & Data Analysis 2004;
47:757–773.
Should the full analysis set contain all randomized subjects?
Eriwan Radio would say: ‘In principle, yes’. However, as we
know, decisions are often difficult in actual trials. Stewart gives
an overview of problems surrounding intention-to-treat and per
protocol analysis and proposes a general approach when
selecting the full analysis set.
* Stewart WH. Basing intention-to-treat on cause and effect
criteria. Drug Information Journal 2004; 38:361–369.
Miscellaneous
The American Statistician contains some interesting statistical
computing software reviews covering WinBUGS, R, and
software for analysing correlated survival data.
* Cowles MK. Review of WinBUGS 1.4. The American
Statistician 2004; 58:330–336.* Kelly PJ. A review of software packages for analyzing
correlated survival data. The American Statistician 2004;
58:337–342.* Horton NJ, Brown ER, Qian L. Use of R as a toolbox for
mathematical statistics exploration. The American Statisti-
cian 2004; 58:343–357.
The genomics revolution radiates more and more in the
world of biostatisticians. The Journal of Biopharmaceutical
Statistics contains 10 papers dealing with the analysis
of microarray data. Authors from a regulatory agency,
university and industry highlight the areas of normaliza-
tion, gene identification, and data integration from their
perspective.
* Chen JJ. Guest editorial: Microarrays in pharmacoge-
nomics. Journal of Biopharmaceutical Statistics
2004; 14:535–537 (10 articles on this topic on pages
539–721).
A personal perspective on trials in the genomic area is given
by Richard Simon. The title is a bit of a pun, supposedly being
an agenda for the journal Clinical Trials, but it is obviously as
much an agenda for clinical trials – do companies, regulators
and patients want wide inclusion criteria and broad applic-
ability, or narrow inclusion criteria and a narrow market? It is
not a simple choice.
* Simon RM. An agenda for Clinical Trials: clinical trials in
the genomic era. Clinical Trials 2004; 1:468–470.
LITERATURE REVIEW 79
Copyright # 2005 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2005; 4: 77–79