Determination of Power and Sample
Size in the Design of
Clinical Trials with Failure-Time
Endpoints and Interim Analyses
There are two
programs, dssp2.exe (Design of Survival Studies: Power, Version 2), which is
used to simulate power and expected duration of the trial, and dess1.exe
(Design of Survival Studies: Sample Size, Version 1), which finds the sample
size of a test having the pre-specified power at given baseline and alternative
distributions. Both programs require the following input information on the
clinical trial and the test statistics.
1. Interim
analysis information
(a) Number of
interim analyses: an integer from 1 to 12. If 1 is entered, this means that
there is only one analysis at the prescheduled end of the trial (i.e., trial
cannot terminate prior to that date).
(b) Times of
interim analysis: if k is entered in (1), then enter k positive numbers (separated
by spaces), in increasing order of magnitude, to denote the time (in years) of
the first, second, ..., kth interim analysis.
2. Basic
information on the test and simulations
(a) type I-error:
Input a number between 0 and 0.1.
(b) One-sided or
two-sided test: Input 1 or 2.
(c) Number of
simulations: We recommend 1000 for the power simulation program, and 500 for
the sample size program.
(d) Seed for random
number generation: The built-in seed is -123457. The user can input 0 for this
built-in seed, or specify and a larger negative integer as the seed used.
3. Accrual
period and accrual rate
The accrual period
is the length of time in years (starting from the beginning of the trial)
during which accrual occurs. The accrual rate is the number of subjects accrued
per year. Their product, therefore, is the total sample size, which should not
exceed 7000 (which is the value m set in the parameter line of the FORTRAN
source code. For those who want to conduct simulation with accrual period * accrual
rate > 7000, you can download the FORTRAN source code dssp2.f, change the
parameter value m, compile the program and run the program with the newly
generated executable file).
For the power
simulation program, the user inputs the accrual rate so that the program
simulates the power corresponding to that rate. However, for the sample size
determination program, the user inputs lower and upper bounds on the accrual
rate. There are usually practical constraints on the sample size in clinical
trials. For example, budget considerations would lead one to a natural
upper bound on the
sample size, while publication and administrative considerations would lead one
to a natural lower bound on the sample size. After dividing by the accrual
period, one therefore obtains upper and lower bounds on the accrual rate. The
product of the upper bound and the accrual length should not exceed 7000 in our
executable version of the program because of the pre-specified dimension of the
data vectors. (For those who want to conduct simulation with accrual period *
upper bound > 7000, you can download the FORTRAN source code dsss1.f, change
the parameter value m, compile the program and run the program with the newly
generated executable file.)
4. Group
sequential boundary
There are 4 methods
of choosing the boundary. The choice is made by typing 1, 2, 3, or 4.
Method 1 is that of
Slud and Wei, for which the user has to input the values e(1), ..., e(k), where
k is the number of interim analyses and e(i) is the error to be spent at the
ith interim analysis. The e(i) are positive numbers whose sum must equal to the
specified Type I error.
Method 2 is the
``use function'' method of Lan and DeMets. The user has to input the use
function choice by typing 1, 2, or 3, in which 1 denotes the O'Brien-Fleming
use function, 2 denotes the Pocock use function and 3 denotes the linear use
function.
Method 3 is the
Haybittle-type method, in which the user has to input a common boundary value
b, which is a positive number, for the first k - 1 interim analyses. The user
who is not sure how to choose b can simply type in 0, and each of the two
programs will automatically reset. Moreover, either program will reset b
suitably if the user has chosen too low a value of b.
Method 4 is the
user-specified boundary. Here the user has to input the boundary values b(1),
..., b(k) for the k interim analyses. Each b(i) is a positive number
representing the threshold whose exceedance by the normalized statistic (or by
its absolute value for a two-sided test) signals stopping.
5. Choice of
test statistics
(a) The statistics
are to be chosen from the Beta family, which has two parameters, rho and tau.
For the logrank statistic, rho = 0 = tau. For the Peto-Prentice generalized
Wilcoxon statistic, rho = 1 and tau = 0. More generally, the Harrington-Fleming
family of statistics has tau = 0.
(b) For the sample
size determination program, the user chooses one statistic and the program
determines the sample size of the test using the statistic chosen. For the
power simulation program, the user can simultaneously simulate the power
associated with no more than five statistics chosen. The user has to input (i)
the number of statistics used, (ii) the rho's of these statistics (separated by
spaces), and (iii) the tau's of these statistics.
6. Baseline
survival function
The baseline (or
control-group) survival function S(t) is specified by a non-increasing
piecewise log-linear function as follows.
(a) The user has to
input first the number, J, of intervals on which log S(t) is determined by
linear interpolation from the values at the endpoints. This number can be any integer
between 1 and 20.
(b) The left
endpoint of the first interval is 0, with S(0) = 1. The user has to input the
right endpoints of the J intervals. These are J positive numbers (separated by
blank spaces) in increasing order of magnitude, with the Jth number at least as
large as the time of the last interim analysis (see 6(b) above).
(c) The user then
inputs the survival probabilities at the J right endpoints in the input (b).
These are J numbers between 0 and 1, in decreasing order of magnitude.
7. Censoring
distributions
The survival
functions of the censoring (due to loss to follow-up) distributions for the
control and treatment groups are specified by piecewise log-linear functions in
the same way as the baseline survival function. If there is no loss in
follow-up, the user can set S(T) = 1, where T is the prescheduled end of the
trial, after setting number of intervals = 1 and right endpoint of the interval
= T.
8. Alternative
survival function
There are two ways
to specify alternative (or treatment-group under the alternative hypothesis)
survival function. One is to specify it as a piecewise log-linear function in
the same way as the baseline survival function. The user inputs 1 if this way
is chosen. The second way, for which the user inputs 2, is to specify it via
the hazard ratio, which is the ratio of the alternative hazard function to the
baseline survival function.
(a) In either way,
the user has to input first the number, I, of intervals, followed by the right
endpoints of these intervals.
(b) For the first
way, the user inputs the survival probabilities at these right endpoints.
(c) For the second
way, the user inputs for each interval the hazard ratio (i.e., hazard function
of the alternative distribution divided by that of the baseline distribution),
which is assumed to be constant over each interval. Thus, the user inputs I
positive numbers (separated by blank spaces) for the hazard ratios of the
successive (from left to right) intervals.
9. Noncompliance
rates
The noncompliance
(or crossover) rates are specified by the drop-out rate (from treatment group
to control group) and drop-in rate (from the control group to treatment group).
The specification of the drop-out rate is similar to that of the drop-in rate.
It assumes that crossovers can only occur at specified time points.
(a) To specify the
drop-in rate, the user first specifies the number of time point(s) at which
drop-in occurs. This number, K, is an integer ranging from 1 to 20.
(b) The user then
inputs the K time-points at which drop-in occurs, separated by blank spaces and
in increasing order of magnitude.
(c) This is
followed by inputting the drop-in rates at these times. The drop-in rate, which
is the proportion of subjects that switch from the control group to the
treatment group, is a number between 0 and 1.
(d) If there is no
drop-in at all, the user can set K = 1, drop-in time = 0, and drop-in rate at
that time = 0.
10. Desired
power
For the sample size
determination program dsss1.exe, the user has to input the desired power so
that the program will find the appropriate accrual rate for the desired power.
Please read the last few lines of the output file to see if the desired power
has been reached. This information can be used to modify either the bounds on
the accrual rate or the desired power in a rerun of the program.