Clinical Trials
Clinical trials design is part of the broad area of the design of experiments. As with all
experiments, well-done clinical trials have: (1) an hypothesis; (2) background; (3)
methods; (4) results; and (5) conclusions. We will concentrate o

We recall from the last lecture the power of our test could be calculated from:
Z 1.65 ln 0.045 ln 0.067
Power P
1
# events
Technically, in order to set T, and n (the per time unit accrual rate, we should be
expressing the power function as:
e
nT
ln 0

Planning a Study Comparing Two Factors
We now know how to plan a cohort trial so that we can assess whether the cohort has changed
from some fixed standard we agree characterizes the established treatment standard. These
studies are less common than studi

Planning a Study Comparing Two Factors
We now know how to plan a cohort trial so that we can assess whether the cohort has changed
from some fixed standard we agree characterizes the established treatment standard. These
studies are less common than studi

Non-Parametric Comparison
Generally, the log rank (Savage scores) test can is calculated as
d
D
U
D
j 1
2
E d Aj
2
~ A (1)
nAj nBj d j n j d j
j 1
Aj
n 2 n j 1
j
If we are willing to accept that the null and alternative hypotheses can be expressed as:

. #delimit ;
delimiter now ;
. set more off;
. use c:\ddrive\pm552\c782_c.dta, clear;
. stset efs, failure(censor=1) id(cid) scale(365.25);
id:
failure event:
obs. time interval:
exit on or before:
t for analysis:
cid
censor = 1
(efs[_n-1], efs]
failure
t

Stratified Tests
The problem with testing tumor size is that the size of the bone limits how big a tumor can be.
The fibula and tibia are, at most of the length of the femur and humerus. This will limit the
size of tumors that occur in the most distal reg

Rank Tests
Formally, we can approach the non-parametric testing by replacing the observed value of
the failure time by its rank. This preserves as much information as possible in the sample
without actually using the true failure time.
We will assume the

Estimating 0(t)
We found out how to estimate the coefficients in the model:
Z
t | Z 0 t e
We were not interested to this point in the baseline hazard, but it may be interesting to
determine whether the baseline is closely approximated by a parametric mo

These are non-parametric tests because their distribution is known when the Z are not
related to outcome. When the Z are related, then the choice of c and C influence the
power. If the AFT model is correct, the choice of scores can affect efficiency. If t

Tied Failure Times
Because the data we have been working with so far are considered absolutely continuous,
we know that we cannot have tied failure times. Our approach will be that, if there are
failure times, the problem is that we just didn't collect th

Designing a Trial to Compare Two Therapies (Phase III Study)
We have all the tools now to design a study that compares two therapies in the situation
where our primary outcome measure is determined only after some follow-up and
individuals may be censored

Treatment Assignment
Randomization: A process based on a realization of a random process unrelated to
accumulating data regarding the study population in terms of patient characteristics or
outcome. Usually, if there are K treatments, the probability of a

Sample Size
Sample size can be calculated by noting that:
, where we have the model
. As noted in the
ln
t , S1 t S0 t
ln
~ AN 0,1
1 1
d1 d 2
Sposto and Sather article, this means that sample size can be calculated with a model for
the survival, a tar

Design Considerations for Various Types of RCTs
Two Treatments, No Implicit Intensity Order In Treatments
AEWS0031 Design
R
A
N
D
O
M
I
Z
E
Usual Approach:
Regimen A
VCR
IFOS
Dox
ETOP
CPM
q3w x 2
Local
Control
VCR
Dox
CPM
q3w x 5
IFO
ET
Regimen B
VCR
IFOS

From the last interim monitoring example, we had:
K= 4
= 0.05
Power Family: t 2 was used as spending Function.
Times
0.25
0.5
0.75
1
Lower Bounds
-2.9552
-2.5593
-2.3008
-2.0919
Upper Bounds
2.9552
2.5593
2.3008
2.0919
[i]-[i-1]
0.003125
0.009375
0.01562

Planning a Cohort Study
If we assume a parametric model for time-to-event, we can use the methodology we have
learned to date to plan a future trial. The elements we will need to complete such a
process are:
1. A statement of the goal of the new study
2.

Checking the Fit of The Model
Cox-Snell residuals are often used to check the fit of the model. These provide a
graphical means to check the model fit, although these are not "analytic".
For each ti in the dataset, the corresponding Cox-Snell residual is

Statistical Testing
In observational studies, we are interested in determining whether our data are consistent
with hypotheses clinical hypotheses regarding the true state of nature. Further, this true
state of nature can be represented in our model. We c

Cancer Clinical Trials
We will examine design issues through the current paradigm of cancer clinical trials.
Although the context is specialized, many of the scientific problems, and their statistical
solutions can be applied to other areas of research. I

Some Reminders from Last Time
Phase I trials - H0: Proposed therapy can be delivered feasibly.
Quantitative evaluation made for each patient that addresses H0: The discrete indicator
random variable - Whether or not a serious tolerability event occurs wit

Comparison of Phase I Algorithms
N+M Design (3+3)
Algorithm is completely specified before the
trial starts.
Does not fully exploit accumulating toxicity
data at levels other than the current level
Most published studies use an up then down
strategy for d

A Brief Statistical Digression
All good study protocols have a
detailed statistical section outlining the
parameters of the design of the study
and how it will be analyzed
Pediatric Clinical Trials Conference
The Study: What is the Goal?
Sample Versus Tar

Clinical Trials
Clinical trials design is part of the broad area of the design of experiments. As with all
experiments, well-done clinical trials have: (1) an hypothesis; (2) background; (3)
methods; (4) results; and (5) conclusions. We will concentrate o

Cancer Clinical Trials
We will examine design issues through the current paradigm of cancer clinical trials.
Although the context is specialized, many of the scientific problems, and their statistical
solutions can be applied to other areas of research. I

Some Reminders from Last Time
Phase I trials - H0: Proposed therapy can be delivered feasibly.
Quantitative evaluation made for each patient that addresses H0: The discrete
indicator random variable - Whether or not a serious tolerability event
occurs wit

Some Reminders from Last Time
Contrasts Between the Algorithmic Approach and the Bayesian Approaches
(CRM and EWOC)
Both approaches to determining a dose of a therapy (be it drug therapy or
radiation therapy or any other type of intervention) attempt to i

Measurement of Effect
A participant characteristic that is required to be present in all participants at the time of
enrollment and that can be objectively measured at times throughout the trial as selected
by the investigator.
1. Serum prostate specific

If the null hypothesis is true, we want to limit the possibility of incorrectly concluding
there is evidence of effect. Usually this is set to 5%-10%. This is called the size of the
test ().
If the alternative hypothesis is true, we want to have a high pr