Altman DG, De Stavola BL, Love SB, Stepniewska KAReview of survival analyses published in cancer journals. Br J Cancer 72: 511-518

Abstract

Survival analysis has found widespread applications in medicine in the last 10-15 years. However, there has been no published review of the use and presentation of survival analyses. We have carried out a systematic review of the research papers published between October and December 1991 in five clinical oncology journals. A total of 132 papers were reviewed. We looked at several aspects of study design, data handling, analysis and presentation of the results. We found that almost half of the papers did not give any summary of length of follow-up; that in 62% of papers at least one end point was not clearly defined; and that both logrank and multivariate analyses were frequently reported at most only as P-values [63/84 (75%) and 22/47 (47%) respectively]. Furthermore, although many studies were small, uncertainty of the estimates was rarely indicated [in 13/84 (15%) logrank and 16/47 (34%) multivariate results]. The procedure for categorisation of continuous variables in logrank analyses was explained in only 8/49 (16%) papers. The quality of graphs was felt to be poor in 43/117 (37%) papers which included at least one survival curve. To address some of the presentational inadequacies found in this review we include new suggested guidelines for the presentation of survival analyses in medical journals. These would complement the statistical guidelines recommended by several clinical oncology journals.

Full text

British
Journal
of
Cancer
(1995)
72,
511-518
©
1995
Stockton
Press
All
rights
reserved
0007-0920/95
$12.00
x
Review
of
survival
analyses
published
in
cancer
journals
DG
Altman,
BL
De
Stavola*,
SB
Love
and
KA
Stepniewska
Medical
Statistics
Laboratory,
Imperial
Cancer
Research
Fund,
London
WC2A
3PX,
UK.
Summary
Survival
analysis
has
found
widespread
applications
in
medicine
in
the
last
10-15
years.
However,
there
has
been
no
published
review
of
the
use
and
presentation
of
survival
analyses.
We
have
carried
out
a
systematic
review
of
the
research
papers
published
between
October
and
December
1991
in
five
clinical
oncology
journals.
A
total
of
132
papers
were
reviewed.
We
looked
at
several
aspects
of
study
design,
data
handling,
analysis
and
presentation
of
the
results.
We
found
that
almost
half
of
the
papers
did
not
give
any
summary
of
length
of
follow-up;
that
in
62%
of
papers
at
least
one
end
point
was
not
clearly
defined;
and
that
both
logrank
and
multivariate
analyses
were
frequently
reported
at
most
only
as
P-values
[63/84
(75%)
and
22/47
(47%)
respectively].
Furthermore,
although
many
studies
were
small,
uncertainty
of
the
estimates
was
rarely
indicated
[in
13/84
(15%)
logrank
and
16/47
(34%)
multivariate
results].
The
procedure
for
categorisa-
tion
of
continuous
variables
in
logrank
analyses
was
explained
in
only
8/49
(16%)
papers.
The
quality
of
graphs
was
felt
to
be
poor
in
43/117
(37%)
papers
which
included
at
least
one
survival
curve.
To
address
some
of
the
presentational
inadequacies
found
in
this
review
we
include
new
suggested
guidelines
for
the
presenta-
tion
of
survival
analyses
in
medical
journals.
These
would
complement
the
statistical
guidelines
recommended
by
several
clinical
oncology
journals.
Keywords:
survival
analysis;
review;
statistics
Survival
analysis
has
found
widespread
applications
in
medicine
in
the
last
10-15
years
(Andersen,
1991),
partic-
ularly
in
clinical
oncology,
and
its
correct
application
and
presentation
is
critically
relevant
for
much
of
the
cancer
literature.
Although
the
use
of
statistical
methods
in
medicine
has
been
subjected
to
much
scrutiny
(see
Altman,
1982,
1991),
we
believe
that
there
has
not
been
any
published
review
of
the
use
of
survival
analysis
methods
in
medical
journals.
Hence,
we
have
carried
out
a
systematic
review
of
the
appropriateness
of
the
application
and
presentation
of
survival
analyses
in
clinical
oncology
journals.
We
have
focused
on
the
size
of
the
studies
being
published,
the
ade-
quacy
of
the
description
of
the
data
analysed
(with
particular
interest
given
to
the
length
and
quality
of
follow-up
and
the
clarity
of
the
end
points
of
interest)
and
the
choice
and
quality
of
univariate,
multivariate
and
graphical
analyses.
In
the
light
of
disappointing
findings,
we
discuss
existing
guidelines
and
present
some
new
guidelines
aimed
in
partic-
ular
at
presentation.
Methods
We
examined
all
papers
published
in
British
Journal
of
Cancer,
European
Journal
of
Cancer,
Journal
of
Clinical
Oncology
and
American
Journal
of
Clinical
Oncology
between
October
and
December
1991
which
included
analyses
of
sur-
vival
data.
There
were
132
papers
which
reported
at
least
one
of
the
following:
Kaplan-Meier
or
actuarial
survival
curves;
logrank
or
related
tests;
parametric
or
semiparametric
sur-
vival
analyses.
Those
papers
with
survival
data
which
did
not
present
any
of
these
analyses,
and
thus
were
not
included,
were
largely
phase
I
or
II
clinical
trials.
The
132
papers
were
reviewed
using
a
standard
form
that
had
been
tested
in
a
small
pilot
study
of
20
papers
which
were
read
by
all
four
authors.
When
the
form
was
finalised
each
paper
was
read
by
two
of
the
authors
according
to
a
balanced
randomisation
scheme.
Disagreements
between
reviewers
were
resolved
in
paired
discussions
and
by
discus-
sion
between
all
four
reviewers
on
the
rare
occasions
when
it
was
necessary.
The
assessment
form
included
separate
sections
relating
to
distinct
aspects
of
each
paper.
It
also
included
the
time
taken
by
each
author
to
extract
the
information
from
the
paper
onto
the
form
as
an
indication
of
the
clarity
of each
paper.
The
form
is
available
from
the
authors
upon
request;
the
contents
of
the
form
are
summarised
below.
Sample
size
The
importance
that
can
be
attached
to
the
results
from
a
survival
analysis
depends
on
the
selection
and
number
of
subjects
included.
Hence,
for
each
paper
we
recorded
the
number
of
subjects
studied
and
the
maximum
and
minimum
number
of
subjects
analysed
by
survival
methods.
The
number
of
events
(e.g.
deaths),
which
determines
the
statis-
tical
power
of
a
survival
study,
was
also
recorded.
Follow-up
The
interpretation
of
the
results
of
a
survival
analysis
depends
in
great
measure
upon
the
time
frame
in
which
the
study
was
carried
out
and
the
completeness
of
follow-up
of
the
subjects
being
investigated.
Three
critical
dates
define
the
start
and
end
of
patient
accrual
and
the
cut-off
date
for
the
analysis
(Shuster,
1991).
We
checked
whether
these
dates
were
reported
and
also
whether
a
summary
of
the
length
of
follow-up
(such
as
a
median)
was
given.
We
also
noted
whether
the
authors
mentioned
if
any
subjects
were
lost
to
follow-up
and,
if
so,
whether
there
was
a
statement
on
how
these
were
treated
in
the
analysis.
Correspondence:
DG
Altman,
Medical
Statistics
Laboratory,
Imper-
ial
Cancer
Research
Fund,
PO
Box
123,
61
Lincoln's
Inn
Fields,
London
WC2A
3PX,
UK
*Current
address:
Epidemiology
Monitoring
Unit,
London
School
of
Hygiene
and
Tropical
Medicine,
London
WC1E
7HT
Received
12
May
1994;
revised
7
November
1994;
accepted
28
March
1995
End
points
Survival
analysis
is
based
on
the
time
measured
from
a
relevant
time
origin
to
a
particular
event
of
interest,
for
example
from
date
of
surgery
to
recurrence
of
disease.
How-
ever,
the
event
of
interest
may
not
be
observed
for
some
patients
because
of
end-of-study
censoring,
loss
to
follow-up
or
competing
events
(such
as
deaths
from
other
causes).
In
these
cases
the
patient's
survival
is
said
to
be
censored
since

Review
of
survival
analysis
in
cancer
journals
DG
Altman
et
al
his/her
actual
survival
time
is
known
to
be
larger
than
the
observed
one.
Problems
arise
when
reading
a
paper
if
the
end
point
of
interest
is
not
clearly
defined
or
losses
to
follow-up
and
competing
events
are
not
specified.
For
example,
analyses
of
time
to
death
may
be
based
on
either
deaths
from
any
cause
or
only
cancer-related
deaths.
Likewise,
in
the
analysis
of
relapse-free
survival
time,
deaths
without
apparent
relapse
may
be
either
treated
as
events
or
censored.
In
addition,
a
single
study
may
include
survival
analyses
of
more
than
one
end
point,
and
the
time
origins
for
these
may
differ.
Additional
problems
arise
when
analysing
relapse-free
sur-
vival
time,
partly
because
of
the
lack
of
standardised
definitions.
Relapse-free
survival,
disease-free
survival,
remis-
sion
duration
and
progression-free
survival
are
the
terms
most
commonly
used;
however,
they
are
rarely
defined
precisely.
The
first
three
imply
that
only
patients
who
res-
ponded
to
treatment
were
analysed
(although
this
is
not
always
the
case),
while
for
the
last
all
patients
are
generally
included
in
the
analysis.
Which
group
of
patients
is
actually
analysed
directly
affects
the
values
of
the
estimated
survival
probabilities,
and
this
has
important
implications
for
the
interpretation
of
the
results.
In
order
to
evaluate
the
quality
of
the
reviewed
papers,
we
therefore
recorded
the
number
of
end
points
studied
in
each
paper,
whether
their
time
origin
was
stated,
whether
censor-
ing
events
were
clearly
reported
and
when
relapse-free
sur-
vival
analyses
were
carried
out
whether
it
was
clear
which
patients
were
included.
Explanatory
variables
When
the
effects
of
prognostic
factors
are
examined
using
survival
analysis
the
results
may
be
affected
by
the
number
of
variables
examined,
their
coding
and
the
presence
of
missing
values.
We
therefore
recorded
the
total
number
of
variables
examined
in
univariate
analyses
and
the
maximum
number
of
variables
examined
in
multivariate
analyses,
whether
con-
tinuous
variables
were
recoded,
and
how
they
were
categ-
orised.
We
also
noted
if
missing
values
were
reported
and
discussed.
For
many
survival
analyses
it
is
important
that
the
variables
are
measured
at
or
before
the
time
origin,
otherwise
their
observation
depends
on
what
happens
to
the
patient
between
the
time
origin
and
the
measurement
of
the
variable.
An
example
is
response
to
treatment
when
the
time
origin
is
diagnosis.
Comparisons
of
the
survival
probabilities
of
res-
ponders
and
non-responders,
where
survival
is
measured
from
diagnosis,
are
still
seen
in
the
literature
despite
many
published
warnings
against
them
(Anderson
et
al.,
1985;
Simon
and
Wittes,
1985).
We
recorded
whether
such
incor-
rect
analyses
were
reported.
Continuous
explanatory
variables
need
to
be
categorised
to
produce
survival
plots
and
perform
some
types
of
analyses.
We
recorded
whether
variables
were
reduced
to
two
or
more
categories
and
whether
the
choice
of
cut-off
points
was
explained.
We
recorded
whether
cut-off
points
were
derived
from
the
data
by
minimising
the
associated
P-value.
This
so-called
'optimal'
cut-off
point
approach
is
seriously
flawed,
leading
to
overestimates
of
prognostic
importance
and
P-
values
that
are
far
too
small
(Altman
et
al.,
1994).
Graphical
presentation
An
accurate
description
of
a
data
set
in
terms
of
survival
is
provided
by
Kaplan-Meier
or
life
table
estimates,
which
are
usually
presented
in
the
form
of
a
graph.
We
recorded
the
type
of
survival
curve
and
number
of
survival
plots
present-
ed.
We
also
considered
the
quality
of
graphs;
how
points
in
the
graph
were
joined
(steps
are
appropriate);
whether
there
were
any
marks
indicating
censoring
times
and
if
so
whether
they
were
identified;
whether
the
number
of
subjects
at
risk
or
confidence
intervals
were
given
at
any
time;
and
whether
the
numerical
axes
were
reasonable.
When
survival
curves
were
presented
for
more
than
one
group
of
patients,
we
noted
if
different
line
types
were
used
and
if
each
line
was
labelled.
Univariate
analyses
We
defined
survival
analysis
as
univariate
where
length
of
survival
was
examined
in
relation
to
only
one
explanatory
variable
at
a
time,
hence
ignoring
the
simultaneous
effects
of
other
variables.
The
most
common
univariate
analysis
of
survival
data
compares
the
survival
in
two
or
more
groups;
these
could be
treatment
groups
in
clinical
trials
or
risk
groups
in
observational
studies.
The
most
familiar
method
is
the
logrank
test,
which
also
goes
under
several
other
names
including
Mantel,
Mantel-Haenszel,
generalised
Savage
and
Mantel-Cox.
There
are
also
a
class
of
tests
(referred
to
here
as
weighted
logrank)
which
allow
events
occurring
at
different
times
to
have
differing
weights
in
the
computation
(Harrington
and
Fleming,
1982),
the
best-known
names
being
the
generalised
Wilcoxon
and
the
Gehan.
We
therefore
recorded
which
tests
were
used
and,
where
applicable,
whether
the
use
of
unequal
weights
was
expl-
ained.
When
the
variable
examined
had
three
or
more
ordered
categories
we
also
checked
whether
the
more
appro-
priate
logrank
test
for
trend
(which
seeks
monotonic
relation-
ship
instead
of
just
heterogeneity)
was
used.
We
noted
which
papers
used
a
Cox
proportional
hazards
regression
model
(Cox,
1972)
with
a
single
explanatory
variable
in
place
of
or
in
addition
to
the
logrank
test.
We
recorded
the
type
of
information
that
was
presented
about
each
analysis,
including
P-values,
median
survival
times
or
survival
probabilities
(e.g.
for
surviving
2
years)
for
each
group
being
compared,
hazard
ratio
(within
survival
analysis
also
known
as
relative
risk)
estimates
and
confidence
intervals.
The
reporting
of
crude
rates
of
observed
events
(calculated
ignoring
the
differing
lengths
of
follow-up)
was
recorded.
Such
simple
indicators
are
easily
misinterpreted.
Multivariate
analyses
We
defined
a
survival
analysis
as
multivariate
where
survival
was
examined
in
relation
to
at
least
two
explanatory
variables
simultaneously.
We
recorded
if
such
an
analysis
involved
Cox
regression
analysis
with
baseline
covariates
(time-fixed),
Cox
regression
analysis
with
covariates
meas-
ured
over
time
(time-dependent),
the
fitting
of
a
Weibull
model,
multivariate
logistic
regression,
adjusted
Kaplan-
Meier
or
stratified
logrank
analyses.
In
addition,
we
noted
how
the
variables
included
in
the
full
model
were
selected.
We
noted
whether
the
authors
discussed
the
model-building
strategy
(e.g.
using
stepwise
analysis),
the
model
assumptions
and
the
goodness
of
fit
of
the
final
model.
We
also
recorded
the
type
of
information
presented
to
summarise
the
results,
including
estimated
regression
coefficients,
hazard
ratios,
P-
values
and
the
computation
of
prognostic
indices.
Subset
analyses
We
defined
a
subset
analysis
as
one
which
did
not
use
the
full
number
of
subjects
who
could
be
used.
We
were
interested
in
whether
any
subset
analyses
were
performed
and,
if
so,
whether
they
were
the
main
purpose
of
the
paper.
We
examined
in
particular
how
the
analyses
of
complementary
subsets
(e.g.
patients
with
different
stages
of
the
disease)
were
performed.
The
reporting
of
separate
analyses
on
each
subset
may
lead
to
erroneous
conclusions;
tests
for interaction
are
more
appropriate
(Simon,
1982;
Simon
and
Altman,
1994).
However,
tests
for
interaction
not
recommended
in
small
samples
owing
to
low
power.
Abstracts
The
abstract
has
many
functions,
the
most
important
for
this
study
being
the
correct
summary
of
results
and
conclusions.
We
recorded
whether
a
summary
of
follow-up
was
men-

Review
of
survival
analys
in
cancer
journals
DG
Altman
et
al
tioned
and
whether
and
how
univariate
and
multivariate
analyses
were
reported.
None
of
the
journals
in
the
study
used
structured
abstracts
(Haynes
et
al.,
1990)
at
the
time
of
the
review.
Miscellanea
The
majority
of
survival
analyses
are
performed
with
the
help
of
computer
programs.
We
noted
if
any
information
was
given
in
the
papers
about
the
commercial
software
used
to
analyse
the
data.
As
an
overall
appraisal
of each
paper
we
made
subjective
assessments
of
the
quality
(recorded
as
adequate
or
not)
of
the
analyses,
the
description
of
the
statistical
methods
and
the
style
of
the
graphical
presentation.
Since
these
assess-
ments
were
subjective,
'not
adequate'
was
noted
if
either
reviewer
recorded
this
option.
Further,
to
evaluate
whether
known
statistical
involvement
improved
the
quality
of
a
paper,
we
noted
if
any
author
was
a
member
of
a
department
of
statistics
or
epidemiology.
Finally,
we
recorded
any
peculiarities
which
we
noticed
in
the
papers
and
which
were
not
covered
by
questions
in
the
form.
They
are
reported
in
the
relevant
sections.
Results
The
132
papers
in
the
study
included
11
from
the
British
Journal
of
Cancer,
52
from
Cancer,
20
from
the
European
Journal
of
Cancer,
32
from
the
Journal
of
Clinical
Oncology
and
17
from
the
American
Journal
of
Clinical
Oncology.
They
represent,
respectively,
11%,
27%,
20%,
60%,
59%
of
all
papers
published
in
the
journal
between
October
and
December
1991.
Papers
from
all
journals
were
considered
together;
it
was
not
our
intention
to
compare
journals.
Most
of
the
papers
described
clinical
trials
(51%)
or
ret-
rospective
observational
studies
(45%).
Of
the
former,
only
18
papers
were
about
controlled
clinical
trials,
and
of
the
latter
only
eight
were
treatment
related.
The
remaining
six
papers
included
a
case-control
study,
the
data
from
a
pro-
spective
screening
study,
three
papers
each
describing
anal-
ysis
on
selections
of
patients
from
several
randomised
clinical
trials
and
one
paper
which
gave
no
information
on
sample
recruitment.
Reading
the
papers
and
completion
of
the
form
took
a
median
of
30
min
per
reader
(range
12-80
min).
Sample
size
The number
of
subjects
in
each
analysis
was
not
always
clear.
A
total
of
123
papers
out
of
132
stated
the
number
of
patients
included
in
the
study
and
the
numbers
examined
in
survival
analyses.
In
21
out
of
123
(17%)
papers
some
study
patients
were
excluded
from
all
survival
analyses.
Seventeen
included
in
their
survival
analyses
a
maximum
of
fewer
than
30
patients
and
three
papers
a
maximum
of
fewer
than
15.
Many
studies
(78/126
with
clear
information,
62%)
incl-
uded
one
or
more
additional
survival
analyses
on
a
selection
of
the
subjects
already
analysed.
Most
of
these
papers
(47/78)
included
subsidiary
analyses
of
less
than
half
of
the
patients
included
in
the
main
survival
analysis.
Moreover,
about
a
third
of
these
papers
(29/78)
included
analyses
of
fewer
than
30
subjects,
and
six
fewer
than
15
subjects.
Fewer
than
half
(45%)
of
the
papers
gave
the
number
of
events
for
each
end
point,
with
only
two-thirds
giving
the
number
of
events
for
at
least
one
of
the
end
points
analysed
within
the
paper.
Follow-up
Fewer
than
a
quarter
of
papers
reported
all
three
dates
of
start
and
end
of
accrual
and
cut-off
point
for
the
analysis
(Table
I).
The
majority
of
the
papers
gave
only
the
accrual
period,
and
nearly
half
of
these
(37/77)
did
not
include
any
summary
of
follow-up
time.
Several
of
the
papers
which
did
not
report
any
dates
also
did
not
summarise
follow-up
(9/16;
see
first
row
in
Table
I).
Almost
half
of
the
papers
did
not
give
any
summary
of
follow-up.
In
at
least
two
cases
the
event
of
interest
had
been
recorded
for
all
patients
by
the
time
of
the
analysis,
making
a
summary
of
follow-up
unnecessary.
For
those
which
did
give
a
summary,
the
median
follow-up
time
was
the
most
frequent
value
presented,
although
the
method
used
to
compute
it
was
rarely
specified
(16/52,
31%).
The
other
summaries
reported
in
these
papers
were
the
mean
follow-up
time,
which
is
inappropriate
in
most
cases
because
of
the
likely
skewness
of
the
observed
survival
times,
and
the
range
of
follow-up
times
(or
only
the
minimum),
which
only
reports
the
most
extreme
cases
and
therefore
is
not
very
informative.
Losses
to
follow-up
were
mentioned
in
34
papers,
of
which
12
declared
no
losses
and
the
remaining
22
reported
that
losses
had
occurred.
However,
12
out
of
these
22
did
not
state
how
losses
were
treated
in
the
analyses.
Endpoints
The
identification
of
end
points
was
one
of
the
hardest
aspects
of
the
assessment
of
papers.
In
many
papers
one
or
more
end
points
was
not
clearly
defined.
Many
papers
refer-
red
to
'overall
survival'.
We
took
this
to
imply
that
the
end
point
was
death
from
any
cause,
but
it
would
be
unsafe
to
assume
that
such
usage
was
the
rule,
as
in
at
least
one
paper
the
term
overall
survival
related
to
cancer
deaths
only.
With
the
exception
of
one
paper,
between
one
and
six
end
points
were
examined
in
univariate
analyses
(median
1)
and
between
one
and
seven
end
points
(median
1)
in
multivariate
analyses.
The
one
other
paper
examined
univariate
analyses
only,
considering
40
end
points.
Only
65%
(30/46)
of
papers
which
also
carried
out
a
multivariate
analysis
examined
the
same
end
points
in
the
univariate
and
multivariate
analyses.
Most
of
the
remaining
16
papers
considered
in
the
mul-
tivariate
analysis
only
a
selection
of
the
end
points
used
in
the
univariate
analysis.
The
most
common
end
point
was
Table
I
Study
dates
and
follow-up
summary
information
Dates
given
Start
End
End
Number
(%)
Summary
of
follow-up
of
accrual
of
accrual
of
follow-up
of
papers
Median
Other
None
No
No
No
16
(12)
5
2 9
Yes
No
No
3
(2)
2
1
0
No
Yes
No
0
(-)
0
0
0
No
No
Yes
3
(2)
1
1
1
Yes
Yes
No
77
(58)
28
12
37
Yes
No
Yes
2
(2)
0
1 1
No
Yes
Yes
0
(-)
0
0 0
Yes Yes
Yes
31
(23)
16
3
12
All
132
(100)
52
20
60

Review
of
survival
analysis
in
cancer
journals
DG
Altman
et
al
death,
which
was
used
in
106
(80%)
papers,
with
explicitly
cancer
death
being
used
in
21
(16%)
and
eight
papers
using
both.
Remission
duration
(progression-free
survival)
analysis
was
carried
out
in
64
(48%)
papers;
it
was
often
unclear
which
patients
were
included.
In
five
papers
time
to
disease
progression
was
examined
in
two
ways:
firstly,
death
from
any
cause
was
treated
as
an
event
and
then
the
analysis
was
rerun
censoring
survival
times
of
subjects
who
died
without
progressing.
In
62%
of
all
papers,
at
least
one
end
point
was
not
clearly
defined.
For
example,
among
the
106
papers
with
death
as
an
end
point,
only
50
explicitly
described
the
end
point
as
either
any
death
or
only
cancer
death.
Likewise,
treatment
of
deaths
was
unclear
in
39
out
of
64
(61%)
papers
which
studied
time
to
disease
progression.
In
48%
of
all
papers
the
time
origin
was
not
stated
for
at
least
one
end
point.
Explanatory
variables
Across
all
the
113
papers
using
univariate
analysis,
the
median
number
of
prognostic
factors
investigated
was
three
(range
1-28)
with
the
number
unclear
in
one
paper.
For
the
47
papers
using
multivariate
analysis,
the
median
investigated
in
the
multivariate
analysis
was
six
(range
2-15)
with
the
number
unclear
in
14
papers.
Table
II
shows
the
number
of
patients
used
in
survival
analyses
in
relation
to
the
number
of
variables
investigated
by
univariate
and
multivariate
met-
hods.
In
28
papers
the
number
of
missing
values
was
not
given
for
at
least
one
of
the
prognostic
factors
tested,
and
in
34
out
of
63
(54%)
papers
with
missing
values
there
was
no
discus-
sion
of
the
missing
data
in
the
text.
For
logrank
analysis,
continuous
variables
have
to
be
categorised.
Among
49
papers
using
categorised
continuous
variables
in
logrank
analysis,
76%
dichotomised
at
least
one
variable,
with
only
eight
giving
a
reason
for
the
cut-off
point
used.
In
Cox
regression,
continuous
variables
do
not
need
to
be
categorised,
but
of
the
38
papers
using
this
model
with
continuous
variables,
63%
categorised
at
least
one
con-
tinuous
variable;
75%
of
these
papers
gave
no
clear
explana-
tion
for
the
cut-off
points
used.
In
eight
papers,
no
inform-
ation
was
provided
about
how
any
of
the
continuous
variables
were
used
in
the
Cox
analysis.
Five
out
of
50
papers
used
the
so-called
optimal
cut-off
point
method
in
the
logrank
test
or
Cox
regression
analysis.
Some
papers
used
variables
which
were
not
known
at
the
time
origin.
Fifteen
of
them
compared
the
survival
of
res-
ponders
and
non-responders.
At
least
three
further
papers
used
other
variables
which
were
measured
after
the
time
origin.
Graphical
presentation
A
survival
curve
was
calculated
in
95%
of
all
papers.
The
majority
of
these
used
the
Kaplan-Meier
method,
other
methods
being
life
table,
actuarial
and
Nelson
estimates
(Nel-
son,
1969).
However,
almost
a
fifth
did
not
name
the
method
of
estimation
used.
At
least
one
plot
of
a
survival
curve
(median
2,
range
1-26)
was
presented
in
117
papers.
In
17%
of
these
papers
at
least
one
graph
used
slopes
to
connect
points
of
the
survival
curve.
Censored
observations
were
rarely
marked
(29%).
Few
papers
gave
numbers
of
patients
at
risk
at
given
times
(8%)
or
showed
confidence
intervals
or
standard
errors
(7%).
Poor
numerical
axis
scales
were
used,
in
at
least
one
graph,
in
31
out
of
117
papers
(26%).
The
most
common
deficiency
was
an
unhelpful
time
scale
-
for
example
using
intervals
of
10
or
20
months
or
even
2000
days.
In
one
paper
percentage
survival
of
120%
was
marked.
In
two
further
papers
the
logarithmic
transformation
of
survival
scale
was
used
without
any
comments
in
the
text.
There
were
98
papers
which
included
graphs
comparing
survival
in
two
or
more
groups.
Fifteen
of
them
did
not
use
different
symbols
or
line
types
to
distinguish
the
curves.
In
51
papers
curves
were
clearly
labelled
either
directly
or
using
a
special
key.
A
further
41
papers
described
the
curves
in
a
legend,
while
six
papers
used
a
mixture
of
these
three
methods.
Of
84
papers
presenting
more
than
one
graph,
20%
were
inconsistent
with
respect
to
at
least
one
feature.
Use
of
differing
line
types
and
labelling
of
curves
was
not
merely
attributable
to
the
policy
of
the
publishing
journal.
We
also
noticed
many
errors
caused mostly
by
authors'
carelessness,
nevertheless
making
the
understanding
of
papers
much
more
difficult.
We
were
not
explicitly
looking
for
this
type
of
error
so
numbers
quoted
here
underestimate
the
true
proportion
of
such
papers.
We
found
that
in
9
out
of
117
(8%)
papers
the
graphs
did
not
tally
with
the
data:
either
estimates
of
survival
quoted
in
the
text
were
different
from
those
presented
in
the
graph
(six
papers)
or
graphs
were
plotted
beyond
the
stated
maximum
follow-up
time
(four
papers).
In
a
further
three
papers,
titles
were
swapped
between
graphs,
and
in
one
paper
the
curves
were
incorrectly
labelled.
In
two
other
papers
the
graphs
suggested
that
the
technique
of
estimation
was
different
from
the
one
specified
in
the
methods
section.
Univariate
analyses
Overall,
113
(86%)
papers
included
at
least
one
univariate
analysis.
Of
the
remaining
19
papers,
one
reported
mul-
tivariate
analysis
only
and
the
other
18
were
uncontrolled
clinical
trials
which
were
included
in
our
review
because
they
presented
graphs
of
survival.
Table
III
shows
the
methods
of
analysis
used
in
the
113
papers.
The
majority
used
some
form
of
logrank
test;
in
13
papers
the
test
used
was
not
stated.
A
weighted
version
of
the
logrank
test
was
used
in
22
papers,
always
without
any
explanation.
The
logrank
test
for
trend
was
used
in
only
one
of
the
papers,
although
ordered
variables
were
used
in
at
least
36
further
papers.
Table
IV
summarises
the
information
Table
II
Number
of
variables
examined
in
relation
to
the
number
of
patients
in
the
study
Number
of
variables
Number
of
patients
1
2-5
6-10
11
+
Unclear
Total
Univariate
analysisa
<
50
12
7
3
1
1
23
51-100
12
16
2
5
035
101-200
7
6
4
4
0
21
>201
9
8
9
6
0
32
All
40
37
18
16
1
112
Multivariate
analysis
<50
-
4
0 0
0
4
51-100
-
4
2
0
5
11
101-200
-
1
62 2
11
>
201
-
7
5
2
7 21
All
-
16
13
4
14
47
aOne
paper
was
excluded
because
of
unclear
number
of
patients

Review
of
survival
analysis
in
cancer
journals
DG
Altman
et
al
presented
in
connection
with
logrank
tests;
the
majority
gave
only
P-values,
and
it
was
rare
for
estimates
of
survival
probabilities,
survival
rates
at
fixed
time
points
or
hazard
ratios
to
be
given
as
well
as
P-values.
Eleven
papers
presented
results
of
univariate
Cox
regres-
sion
analyses,
seven
in
addition
to
the
logrank.
Five
papers
reported
the
estimated
Cox
coefficients
or
hazard
ratios
for
at
least
one
of
the
univariate
models,
with
an
associated
measure
of
precision
in
four
of
them.
Confidence
intervals
are
especially
appropriate
when
pres-
enting
the
results
of
controlled
clinical
trials.
Only
4
of
the
18
controlled
trials
in
our
study
(22%)
provided
confidence
intervals
for
the
treatment
effect.
The
crude
rate
of
observed
events
was
quoted
in
38
(29%)
papers.
Multivariate
analyses
Of
the
113
papers
which
presented
the
results
of
univariate
analyses,
46
also
included
multivariate
results.
One
additional
paper
reported
only
a
multivariate
Cox
regression
analysis.
Multivariate
analyses
other
than
Cox
regression
were
pres-
ented
in
only
three
papers:
two
were
stratified
logrank
analyses
(both
also
presented
the
results
of
a
Cox
regression)
and
one
was
a
logistic
regression
model.
In
three
papers
it
was
not
clear
what
type
of
multivariate
analysis
had
been
done.
Three
papers
reported
variations
of
the
Cox
model.
Stratified
Cox
regression
models
were
used
in
two
papers
and
a
time-dependent
Cox
model
was
compared
with
the
stan-
dard
time-fixed
specification in
one
study.
The
assumptions
underlying
the
Cox
model
were
investigated
in
two
papers
out
of
43
(5%),
one
by
plots
of
the
logarithm
of
the
cumulative
hazard
and
one
by
comparisons
of
the
Cox
regression
estimates
with
those
from
fully
parametric
models
(though
the
Cox
model
was
presented).
None
of
the
papers
assessed
goodness
of
fit,
but
validation
of
the
final
model
was
carried
out
in
one
paper
using
split-sample
methods.
Few
papers
reported
examining
more
than
ten
variables
(Table
II).
The
choice
of
variables
to
examine
was
related
to
the
univariate
analysis
in
12
papers
(being
either
all
those
used
or
all
those
found
to
be
significant
in
the
univariate
analysis),
but
in
14
papers
no
explanation
was
given
for
the
set
of
variables
used.
In
a
further
13
papers
it
was
unclear
which
variables
had
been
examined
in
the
multivariate
analysis.
The
strategy
for
building
the
multivariate
model
was
unclear
in
25
papers out
of
47
(53%);
over
half
of
the
remaining
papers
(13)
used
a
stepwise
procedure.
Table
III
Type
of
univariate
analysis
(n
=
113)
Number
(%)
Type
of
analysis
of
papers
Median/n
year
%
survival
5
(4)
Log-rank
(equal
weights)
62
(55)
Weighted
log-rank
22
(19)
Cox
regression
4
(4)
Other
named
testa
7
(6)
Unnamed
test
13
(12)
aChi
squared,
Mann-
Whitney,
t-test,
Kaplan-Meier.
Table
IV
Presentation
of
log-rank
results
(n
=
84)
Presented
results
Number
(%)
Estimate
SE/CI
P-value
of
papers
No
No
Noa
6
(7)
Yes
No No
1
(1)
No
Yes
No
1
(1)
No
No
Yes
57
(68)
Yes
Yes
No
1
(1)
Yes
No
Yes
7
(8)
No
Yes
Yes
9
(11)
Yes
Yes
Yes
2
(2)
aAll
papers
stated
whether
variables
were
'significant'.
Table
V
shows
how
the
results
of
the
multivariate
analyses
were
presented.
Half
of
the
papers
gave
at
least
the
model
estimates,
but
six
papers
gave
no
results
despite
mentioning
that
a
multivariate
survival
analysis
had
been
performed.
A
prognostic
index
was
computed
in
only
two
papers,
and
plots
of
survival
probability
based
on
the
multivariate
model
were
presented
in
six
papers.
Subset
analyses
Subset
analyses
were
carried
out
in
52%
of
all
papers,
being
the
main
purpose
of
the
study
in
almost
a
fifth
and
with
a
reason
given
for
at
least
some
of
the
subset
analysis
in
43%
of
these
papers.
Independent
analysis
of
complementary
subsets
was
carried
out
in
39%
of
papers;
no
paper
reported
a
test
of
interaction.
Abstracts
Out
of
72
papers
which
gave
a
summary
of
length
of
follow-
up
in
the
body
of
the
paper,
30
gave
at
least
some
of
this
information
in
the
abstract
and
one
further
paper
gave
a
different
summary
in
the
abstract
and
in
the
body
of
the
paper.
The
situation
was
better
for
a
summary
of
survival,
with
42
out
of
the
63
who
gave
a
summary
in
the
body
of
the
paper
also
giving
at
least
one
summary
in
the
abstract.
However,
11
papers
that
gave
no
summary
of
survival
in
the
body
of
the
paper
reported
a
median
survival
or
n
year
percentage
in
the
abstract.
Of
the
46
papers
which
contained
both
univariate
and
multivariate
analysis,
nine
gave
no
multivariate
results
in
the
abstract
and
a
further
nine
gave
more
emphasis
to
univariate
than
to
multivariate
results.
Miscellanea
Only
19%
of
all
papers
mentioned
use
of
any
statistical
software,
most
of
these
using
BMDP
(17)
and
SAS
(6).
Software
was
mentioned
much
more
often
when
multivariate
analysis
was
performed,
in
38%
papers
compared
with
8%
papers
without
multivariate
analysis.
Our
subjective
assessment
of
the
papers
indicated
poor
quality
of
both
analysis
and
presentation
in
most
papers.
Only
in
57
papers
(43%)
were
univariate
analyses
felt
to
be
adequate
and
satisfactorily
presented.
For
22/67
papers
which
presented
only
univariate
results,
we
judged
that
mul-
tivariate
analyses
should
have
been
performed.
For
mul-
tivariate
analyses
we
felt
that
17/47
papers
performed
an
adequate
analysis
and
presented
results
satisfactorily.
All
plots
were
felt
to
be
acceptable
in
63%
papers
which
present-
ed
survival
graphs.
The
description
of
the
statistical
methods
was
judged to
be
adequate
in
64%
of
all
papers,
but
the
content
and
presentation
of
the
analyses
were
far
less
satis-
factory.
In
only
21%
of
all
papers
were
presentation
of
univariate
and
multivariate
analyses
and
graphs
considered
adequate.
The
inclusion
of
a
member
of
a
department
of
statistics
or
epidemiology
as
an
author
of
the
paper
was
not
related
to
the
quality
of
the
analyses,
although
it
did
seem
to
have
some
positive
effect
on
the
description
of
the
statistical
methods
and
the
quality
of
the
graphs
included
in
the
papers
(Table
VI).
Table
V
Presentation
of
multivariate
results
(n
=
47)
Presented
results
Number
(%)
Estimate
SE/CI
P-value
of
papers
No
No
Noa
7
(15)
No
No
Yes
15
(32)
Yes
Yes
No
4
(
9)
Yes
No
Yes
9
(19)
Yes Yes Yes
12
(26)
aHowever,
one
paper
stated
whether
variables
were
'significant'
or
not.
515

Review
of
survival
analysis
in
cancer
journals
DG
Altman
et
al
Table
VI
Subjective
assessment
of
the
quality
of
presentation
and
statistician's
involvement
Percentage
of
papers
with
statistician
among
authors
No
Unclear
Yes
Total
(%)
Adequate
(n
=
70)
(n
=
31)
(n
=
31)
(n
=
132)
Description
61
58
77
64
Graphs
51
58
81
60
Univariate
analysis
40
52
42
43
Multivariate
analysis
66
52
55
58
Discussion
Many
reviews
of
the
use
of
statistics
in
medical
journals
have
found
that
the
standard
of
the
statistical
component
of
pub-
lished
papers
is
poor
(see
Altman
1982,
1991,
for
references).
Most
reviews
have
been
either
of
a
particular
type
of
study
(especially
clinical
trials)
or
were
general
examinations
of
all
statistical
analyses
in
certain
journals.
A
few
authors
have
concentrated
on
the
use
of
specific
statistical
techniques,
but
we
are
not
aware
of
any
previous
investigation
of
the
use
of
survival
analyses
in
published
papers.
The
use
of
survival
analysis
has
increased
markedly
in
recent
years
(Altman,
1991;
Andersen,
1991),
and
it
plays
a
particularly
crucial
role
in
clinical
cancer
research.
The
papers
included
in
our
review
constituted
28%
of
all
papers
published
in
five
clinical
cancer
journals
in
the
3
months
of
the
study.
The
results
of
our
review
show
that
in
most
areas
examined
a
high
proportion
of
papers
were
deficient.
The
reporting
of
follow-up
was
very
poor,
with
only
23%
of
papers
giving
all
three
relevant
dates
(Table
I).
As
well
as
the
dates
of
the
study,
it
can
be
helpful
to
provide
the
median
follow-up
time.
There
are,
however,
several
ways
of
cal-
culating
the
median,
not
all
of
which
are
sensible
(Shuster,
1991).
The
median
follow-up
time
of
all
patients
is
of
ques-
tionable
value
because
it
is
directly
affected
by
the
times
of
the
observed
events.
Many
authors
provide
the
median
follow-up
time
of
the
survivors
only,
but
this
value
can
be
quite
unstable
if
the
number
of
survivors
is
small.
Two
more
plausible
measures
are
the
time
from
the
median
patient
entry
to
the
cut-off
date
of
the
study
and
the
median
time
to
censoring
derived
from
a
'reverse'
Kaplan-Meier
analysis
in
which
the
outcomes
'dead'
and
'censored'
are
exchanged
(Shuster,
1991).
Most
papers
in
our
study
which
reported
the
median
follow-up
time
did
not
explain
how
the
median
was
calculated.
More
seriously,
the
majority
of
papers
(62%)
gave
an
unclear
description
of
at
least
one
of
the
study
end
points.
Specific
difficulties
included
the
failure
to
specify
whether
non-cancer
deaths
were
treated
as
events
or
censored
and
failure
to
specify
how
deaths
without
relapse
were
treated
in
analyses
of
disease-free
survival
time.
This
issue
is
discussed
by
Peto
et
al.
(1977)
and
Gelber
and
Goldhirsch
(1992).
The
widespread
absence
of
adequate
information
about
end
points
is
rather
surprising.
It
is
hard
to
see
how
readers
can
adequately
assess
a
study
without
this
information.
The
large
majority
of
papers
reported
the
use
of
the
log-
rank
test
for
univariate
analyses,
with
12%
not
specifying
the
method
used.
The
most
notable
weakness
that
we
identified
was
the
almost
complete
failure
to
use
the
logrank
test
for
trend
when
survival
was
compared
in
three
or
more
ordered
groups.
Such
failure
may
greatly
reduce
the
statistical
power
of
the
analysis
(Peto
et
al.,
1977)
so
real
associations
may
be
missed.
Only
one
paper
used
this
method
out
of
37
which
analysed
such
data.
The
logrank
test
does
not
give
any
information
about
the
actual
survival
experience
of
the
patients
in
the
study.
The
P-value
is
not
a
measure
of
the
difference
in
survival
between
groups.
Further,
P-values
alone
do
not
indicate
the
direction
of
observed
differences.
Thus
it
is
desirable
also
to
quantify
the
survival
for
each
group
of
interest,
using
the
median
survival
time
or
the
percentage
surviving
a
given
number
of
years.
The
comparative
survival
experience
of
two
groups
can
be
usefully
assessed
by
the
hazard
ratio
(Peto
et
al.,
1977;
Gelber
and
Goldhirsch,
1992).
When
estimates
such
as
these
are
given
it
is
also
desirable
to
supply
confidence
intervals
(Simon,
1986;
Machin
and
Gardner,
1989).
Few
papers
in
our
sample
gave
such
information
-
of
the
84
papers
repor-
ting
the
results
of
logrank
tests,
less
than
a
sixth
reported
estimates
of
survival
or
gave
standard
errors
or
confidence
intervals
(Table
IV).
In
controlled
trials
it
is
especially
impor-
tant
to
provide
a
confidence
interval
for
the
treatment
effect:
in
our
sample
only
4
of
the
18
reports
of
controlled
trials
did
so.
About
a
third
of
the
papers
reported
the
results
of
mul-
tivariate
analyses,
about
half
of
which
included
estimates
of
some
sort
and
a
third
gave
standard
errors
or
confidence
intervals
(Table
V).
Although
estimates
were
provided
more
often
than
for
univariate
analysis,
only
16
of
the
47
papers
provided
an
adequate
summary.
Further,
it
was
often
unclear
exactly
how
the
multivariate
analyses
had
been
carried
out.
The
logrank
test
and
Cox
regression
analysis
make
up
the
vast
majority
of
published
survival
analyses,
with
both
methods
being
used
in
many
papers.
The
methods
of
analys-
ing
survival
data
have
been
described
in
various
medical
statistics
textbooks
(Collett,
1994;
Dawson-Saunders
and
Trapp,
1994)
and
a
few
helpful
journal
articles
(Peto
et
al.,
1977;
Tibshirani,
1982;
Elashoff,
1983;
Christensen,
1987).
Of
the
papers
that
reported
both
univariate
and
multivariate
analyses
18/46
(39%)
chose
to
highlight
the
univariate
results
in
the
abstract.
In
some
of
these
studies
the
univariate
results
of
interest
were
significant
while
the
multivariate
results
were
not.
The
selective
reporting
of
results
in
abstracts
has
been
noted
before
in
other
areas
of
research
(Pocock
et
al.,
1987;
G0tzsche,
1989).
Continuous
variables
were
frequently
analysed
by
creating
categories,
but
the
basis
for
the
choice
of
categories
was
often
missing.
Five
papers
reported
using
the
'optimal'
cut-
off
point
approach,
which
is
not
recommended
(Altman,
1992;
Altman
et
al.,
1994).
Many
of
the
studies
were
quite
small;
over
a
third
of
papers
included
analyses
of
fewer
than
30
patients.
When
such
analyses
compare
two
(or
more)
groups
there
is
very
poor
statistical
power
to
detect
any
real
differences,
although
this
weakness
is
disguised
unless
a
confidence
interval
is
supplied.
Over
half
of
the
papers
included
some
analysis
of
a
subset
of
the
patients,
but
the
reason
for
performing
these
analyses
was
given
in
less
than
half
of
them.
It
is
disappoin-
ting
that
so
many
studies
reported
the
results
of
separate
significance
tests
performed
on
complementary
subgroups.
Here,
too,
confidence
intervals
are
much
more
revealing
than
P-values.
It
is
also
disappointing
to
see
so
many
papers
(11%)
comparing
the
survival
of
patients
who
do
or
do
not
respond
to
treatment
or
comparisons
of
other
groups
defined
by
events
occurring
after
the
start
of
follow-up,
given
the
numerous
published
admonishments
against
such
analyses
(e.g.
Anderson
et
al.,
1985;
Simon
and
Wittes,
1985).
An
association
between
response
and
survival
is
generally
to
be
expected,
and
provides
no
information
about
treatment
efficacy.
Many
of
the
deficiencies
we
have
described
can
be
classified
as
poor
reporting
rather
than
errors
in
metho-
dology,
but
this
should
not
be
taken
to
suggest
that
the
identified
weaknesses
are
unimportant.
Omitted
or
ambi-
guous
descriptions
of
the
methods
used
makes
it
difficult
or
impossible
for
readers
to
know
what
was
done.
Similar
observations
were
made
by
Hokanson
et
.al.
(1986),
who
surveyed
the
statistical
methods
used
in
several
cancer
jour-
nals.
Authors
should
adhere
to
the
advice
of
the
Interna-
tional
Committee
of
Medical
Journal
Editors
(1991):
'Des-
cribe
statistical
methods
with
enough
detail
to
enable
a
knowledgeable
reader
with
access
to
the
original
data
to
verify
the
reported
results'.
Poor
reporting
is
one
of
the
failings
that
is
most
amenable
to
improvement.
We
hope
that
referees
and
editors
will
be
more
vigilant
on
this
aspect
in
future.
We
also
identified
many
presentational
problems.
In
par-
ticular,
there
was
a
general
tendency
to
present
results
as
516

Review
of
survival
analysis
in
cancer
journals
DG
Altman
et
al
*;
517
P-values
without
quantitative
results,
and
an
absence
of
confidence
intervals.
Many
of
the
survival
plots
were
of
poor
quality.
Most
of
the
problems
were
not
too
serious,
but
there
was
considerable
scope
for
improved
presentation
of
inform-
ation
in
graphs
(see
Greenberg
et
al.,
1983).
However,
the
use
of
sloping
lines
to
join
Kaplan-Meier
survival
curves
gives
incorrect
estimates
(Peto
et
al.,
1977).
Some
of
the
deficiencies
may
have been
due
to
limitations
of
the
software
used.
We
looked
at
the
advice
currently
given
by
each of
the
journals
in
our
survey
in
their
instructions
to
authors
regar-
ding
statistical
aspects
of
submitted
papers.
The
British
Jour-
nal
of
Cancer
and
the
European
Journal
of
Cancer
make
no
mention
of
statistics.
Cancer
and
the
Journal
of
Clinical
Oncology
recommend
authors
to
consult
the
guidelines
of
Simon
and
Wittes
(1985).
The
latter
also
recommends
the
guidelines
of
Zelen
(1983).
The
American
Journal
of
Clinical
Oncology
refers
to
the
'NCI
Methodologic
Guidelines
for
Reports
of
Clinical
Trials'
published
in
their
journal.
Eight
of
the
ten
points
in
this
unattributed
document
(Anonymous,
1986)
are
in
fact
taken
from
Simon
and
Wittes
(1985).
The
original
recommendation
about
quality
control
of
data
was
omitted.
The
most
quoted
guidelines
(Simon
and
Wittes,
1985)
are
cited
in
some
form
by
three
of
the
five
journals
in
our
study
(and
also
by
the
Journal
of
the
National
Cancer
Institute).
Their
nine
points
cover
quality
control
of
data;
accounting
for
patients
and
describing follow-up;
rates
of
ineligibility
or
missing
data;
the
need
for
an
intention
to
treat
analysis
in
randomised
trials;
sample
size,
power
and
confidence
inter-
vals;
planned
sample
size
and
reasons
for
stopping
patient
accrual;
dangers
of
using
non-randomised
controls
in
treat-
ment
comparisons
and
the
unacceptability
of
comparing
sur-
vival
of
responders
and
non-responders;
description
of
patients,
extrapolation
and
subset
analyses;
and
finally
a
statement
of
the
principle
of
adequate
description
already
referred
to
above.
Zelen
(1983)
made
a
similar
set
of
recom-
mendations,
and
also
discussed
other
issues
including
the
early
reporting
of
clinical
trials.
These
two
sets
of
sensible
guidelines
cover
many
of
the
issues
that
we
investigated
in
our
review.
It
is
clear
that
many
authors
have
not
taken
notice
of
these
recommendations,
and
that
referees
and
editors
have
not
insisted
that
authors
do
so.
These
sets
of
guidelines
do
not,
however,
include
much
advice
on
the
presentation
of
the
results
of
studies
of
sur-
vival.
There
seems
to
be
little
published
advice
elsewhere
on
summarising
the
data
and
presenting
the
results
of
the
analyses,
although
some
authors
have
considered
some
of
the
relevant
issues,
mostly
in
the
context
of
the
reporting
of
clinical
trials
(Greenberg
et
al.,
1983;
Meinert,
1986;
Gelber
and
Goldhirsch,
1993).
Our
findings
suggest
that
guidelines
for
presentation
of
survival
analysis
in
medical
journals
would
be
useful.
Some
suggested
guidelines
are
shown
in
the
appendix.
These
largely
follow
from
the
most
obvious
repor-
ting
deficiencies
described
above.
The
methodological
stan-
dard
of
papers
published
in
cancer
journals
could
be
imp-
roved
if
authors
were
made
to
adhere
to
these
or
similar
guidelines
for
presentation
as
well
as
the
other
guidelines
discussed
above
(Zelen,
1983;
Simon
and
Wittes,
1985).
Acknowledgements
We
thank
David
Collett,
Robert
Edwards,
David
Machin
and
Peter
Sasieni
for
comments
on
an
earlier
version
of
the
paper.
References
ALTMAN
DG.
(1982).
Statistics
in
medical
journals.
Stat.
Med.,
1,
59-71.
ALTMAN
DG.
(1991).
Statistics
in
medical
journals:
developments
in
the
1980s.
Stat.
Med.,
10,
1897-1913.
ALTMAN
DG.
(1992).
Categorising
continuous
variables.
Br.
J.
Cancer,
64,
975.
ALTMAN
DG,
LAUSEN
B,
SAUERBREI
W
AND
SCHUMACHER
M.
(1994).
Dangers
of
using
'optimal'
cutpoints
in
the
evaluation
of
prognostic
factors.
J.
Natl
Cancer
Inst.,
86,
829-835.
ANDERSEN
PK.
(1991).
Survival
analysis
1982-91:
the
second
decade
of
the
proportional
hazards
regression
model.
Stat.
Med.,
10,
1931-1941.
ANDERSON
JR,
CAIN
KC,
GELBER
RD
AND
GELMAN
RS.
(1985).
Analysis
and
interpretation
of
the
comparison
of
survival
by
treatment
outcome
variables
in
cancer
clinical
trials.
Cancer
Treat.
Rep.,
69,
1139-1144.
ANONYMOUS.
(1986).
Methodologic
guidelines
for
reports
of
clinical
trials.
Am.
J.
Clin.
Oncol.,
9,
276.
CHRISTENSEN
E.
(1987).
Multivariate
survival
analysis
using
Cox's
regression
model.
Hepatology,
7,
1346-1358.
COLLETT
D.
(1994).
Modelling
Survival
Data
in
Medical
Research.
Chapman
&
Hall:
London.
COX
DR.
(1972).
Regression
models
and
life-tables
(with
discussion)
J.R.
Stat.
Soc.
B,
39,
86-94.
DAWSON-SAUNDERS
B
AND
TRAPP
RG.
(1994).
Basic
and
Clinical
Biostatistics,
2nd
edn.
Appleton
&
Lange:
East
Norwalk,
CT.
ELASHOFF
JD.
(1983).
Surviving
proportional
hazards.
Hepatology,
3,
1031-1035.
GELBER
RD
AND
GOLDHIRSCH
A.
(1992).
Reporting
and
interp-
reting
adjuvant
therapy
clinical
trials.
J.
Natl
Cancer
Inst.
Monogr.,
11,
59-69.
G0TZSCHE
PC.
(1989).
Methodology
and
overt
and
hidden
bias
in
reports
of
196
double-blind
trials
of
nonsteroidal,
antiinflammatory
drugs
in
rheumatoid
arthritis.
Controlled
Clin.
Trials,
10,
31-56.
GREENBERG
ER,
BARON
JA
AND
COLTON
T.
(1983).
Reporting
the
results
of
a
clinical
trial.
In
Clinical
Trials:
Issues
and
Approaches,
Shapiro
SH
and
Louis
TA.
(eds)
pp.
191-204.
Marcel
Dekker:
New
York.
HARRINGTON
DP
AND
FLEMING
TR.
(1982).
A
class
of
rank
test
procedures
for
censored
survival
data.
Biometrika,
69,
553-566.
HAYNES
RB,
MULROW
CD,
HUTH
EJ,
ALTMAN
DG
AND
GARD-
NER
MJ.
(1990).
More
informative
abstracts
revisited.
Ann.
Intern.
Med.,
113,
69-76.
HOKANSON
JA,
LUTTMAN
DJ
AND
WEISS
GB.
(1986).
Frequency
and
diversity
of
use
of
statistical
techniques
in
oncology
journals.
Cancer
Treat.
Rep.,
70,
589-594.
INTERNATIONAL
COMMITTEE
OF
MEDICAL
JOURNAL
EDITORS.
(1991).
Uniform
requirements
for
manuscripts
submitted
to
biomedical
journals.
Br.
Med.
J.,
302,
338-341.
MACHIN
D
AND
GARDNER
MJ.
(1989).
Calculating
confidence
intervals
for
survival
time
analyses.
In
Statistics
with
Confidence
Gardner
MJ
and
Altman
DG.
(eds)
pp.
64-70.
British
Medical
Journal:
London.
MEINERT
C.
(1986).
Preparation
of
the
study
publication.
In:
Clinical
Trials.
Design,
Conduct,
and
Analysis,
pp.
264-270.
Oxford
University
Press:
New
York.
NELSON
W.
(1969).
Hazard
plotting
for
incomplete
failure
data.
J.
Qual.
Technol.,
1,
27-52.
PETO
R,
PIKE
MC,
ARMITAGE
P,
BRESLOW
NE.
et
al.
(1977).
Design
and
analysis
of
randomized
clinical
trials
requiring
prolonged
observation
of
each
patient.
II.
Analysis
and
examples.
Br.
J.
Cancer,
35,
1-39.
POCOCK
SJ,
HUGHES
MD
AND
LEE
RJ.
(1987).
Statistical
problems
in
the
reporting
of
clinical
trials.
A
survey
of
three
medical
journals.
N.
Engl.
J.
Med.,
317,
426-432.
SHUSTER
JJ.
(1991).
Median
follow
up
in
clinical
trials.
J.
Clin.
Oncol.,
9,
191-192.
SIMON
R.
(1982).
Patient
subsets
and
variation
in
therapeutic
efficacy.
Br.
J.
Clin.
Pharmacol.,
14,
473-482.
SIMON
R.
(1986).
Confidence
limits
for
reporting
results
of
clinical
trials.
Ann.
Intern.
Med.,
105,
429-435.
SIMON
R
AND
ALTMAN
DG.
(1994).
Statistical
aspects
of
prognostic
factor
studies
in
oncology.
Br.
J.
Cancer,
69,
979-985.
SIMON
R
AND
WITTES
RE.
(1985).
Methodologic
guidelines
for
reports
of
clinical
trials.
Cancer
Treat.
Rep.,
69,
1-3.
TIBSHIRANI
R.
(1982).
A
plain
man's
guide
to
the
proportional
hazards
model.
Clin.
Invest.
Med.,
5,
63-68.
ZELEN
M.
(1983).
Guidelines
for
published
papers
on
cancer
clinical
trials:
responsibilities
of
editors
and
authors.
J.
Clin.
Oncol.,
1,
164-
169.

Review
of
survival
analysis
in
cancer
journals
DG
Altman
et
al
Appendix:
Suggested
guidelines for
presentation
of
survival
analyses
Presentation
of
data
*
Describe
the
recruitment
and
analysis
dates.
*
Describe
the
reason
for
the
sample
size.
*
Report
a
summary
of
follow-up,
such
as
the
median
and
quartiles
computed
by
the
reverse
Kaplan-Meier
method.
*
Report
how
many
subjects
were
lost
to
follow-up
and
whether,
and
how,
they
had
been
included
in
the
analyses.
*
Report
the
number
of
events
for
each
end
point.
Presentation
of
methods
*
Give
a
clear
definition
of
each
end
point
being
considered,
i.e.
define
the
time
origin,
the
event
of
interest
and
the
circumstances
where
survival
times
are
censored.
*
Name
the
method
used
for
estimating
survival
pro-
babilities.
*
Name
any
test
used
in
the
analyses;
in
particular,
justify
the
use
of
weighted
logrank
tests.
*
Report
the
test
for
trend
when
ordered
categorical
variables
are
examined.
*
When
performing
univariate
or
multivariate
analyses,
report
all
the
covariates
examined,
their
frequency
of
miss-
ing
values
and
the
definition
of
the
categories
used
(if
any)
whether
the
covariate
is
significant
or
not.
*
When
Cox
regression
analyses
are
performed,
describe
the
criteria
used
to
select
the
variables
in
the
initial
model,
the
procedure
to
specify
the
final
model
and
describe
any
methods
used
to
assess
the
model
assumptions.
*
Name
the
software
used.
Presentation
of
results
*
Give
a
summary
of
overall
survival:
preferably
median
and/or
percent
surviving
n
years.
*
If
study
is
a
randomised
clinical
trial,
give
separate
sum-
maries
of
survival
for
each
treatment
group.
When
reporting
the
results
of
any
test,
give
the
test
statis-
tic,
the
degrees
of
freedom
(when
applicable)
and
the
exact
P-value.
*
When
presenting
results
of
a
logrank
test
also
report
the
numbers
of
observed
and
expected
events
in
each
group
(desirable).
*
When
comparing
survival
in
two
or
more
groups,
give
an
estimate
of
the
survival
in
each
group,
e.g.
median
sur-
vival
time,
survival
probabilities
for
a
particular
time
point,
hazard
ratio.
*
When
presenting
the
results
of
a
Cox
regression
analysis,
report
the
estimated
coefficients
(or
estimated
hazard
ratios),
measures
of
their
precision
(i.e.
standard
errors
or
confidence
intervals)
and/or
the
associated
P-values.
*
Do
not
use
crude
rates
to
summarise
the
data.
Graphs
*
Use
meaningful
time
intervals.
*
Use
a
step
function
to
join
Kaplan-Meier
survival
estimates.
*
Mark
the
survival
time
of
censored
observations
(desir-
able).
*
If
several
curves
are
reported
in
the
same
plot
use
different
lines
type
(desirable).
*
Give
number
of
patients
at
risk
at
selected
time
points
(desirable).
*
Mark
confidence
intervals
or
standard
errors
for
some
of
the
selected
time
points
(desirable).
Abstract
*
Include
in
the
abstract
summaries
of
follow-up
and
survival
(separately
by
treatment
group
if
applicable)
and
the
final
results
of
both
univariate
and
multivariate
analyses
(when
applicable).
518