sfLinear: 4.6: Piecewise Linear and Step Function Spending Functions

Description Usage Arguments Value Note Author(s) References See Also Examples

Description

The function sfLinear() allows specification of a piecewise linear spending function. The function sfStep() specifies a step function spending function. Both functions provide complete flexibility in setting spending at desired timepoints in a group sequential design. Normally these function will be passed to gsDesign() in the parameter sfu for the upper bound or sfl for the lower bound to specify a spending function family for a design. When passed to gsDesign(), the value of param would be passed to sfLinear() or sfStep() through the gsDesign() arguments sfupar for the upper bound and sflpar for the lower bound.

Note that sfStep() allows setting a particular level of spending when the timing is not strictly known; an example shows how this can inflate Type I error when timing of analyses are changed based on knowing the treatment effect at an interim.

Usage

1
2
sfLinear(alpha, t, param)
sfStep(alpha, t, param)

Arguments

alpha

Real value > 0 and no more than 1. Normally, alpha=0.025 for one-sided Type I error specification or alpha=0.1 for Type II error specification. However, this could be set to 1 if for descriptive purposes you wish to see the proportion of spending as a function of the proportion of sample size or information.

t

A vector of points with increasing values from 0 to 1, inclusive. Values of the proportion of sample size or information for which the spending function will be computed.

param

A vector with a positive, even length. Values must range from 0 to 1, inclusive. Letting m <- length(param/2), the first m points in param specify increasing values strictly between 0 and 1 corresponding to interim timing (proportion of final total statistical information). The last m points in param specify non-decreasing values from 0 to 1, inclusive, with the cumulative proportion of spending at the specified timepoints.

Value

An object of type spendfn. The cumulative spending returned in sfLinear$spend is 0 for t <= 0 and alpha for t>=1. For t between specified points, linear interpolation is used to determine sfLinear$spend.

The cumulative spending returned in sfStep$spend is 0 for t<param[1] and alpha for t>=1. Letting m <- length(param/2), for i=1,2,...m-1 and param[i]<= t < param[i+1], the cumulative spending is set at alpha * param[i+m] (also for param[m]<=t<1).

Note that if param[2m] is 1, then the first time an analysis is performed after the last proportion of final planned information (param[m]) will be the final analysis, using any remaining error that was not previously spent.

See Spending function overview for further details.

Note

The manual is not linked to this help file, but is available in library/gsdesign/doc/gsDesignManual.pdf in the directory where R is installed.

Author(s)

Keaven Anderson keaven\[email protected]

References

Jennison C and Turnbull BW (2000), Group Sequential Methods with Applications to Clinical Trials. Boca Raton: Chapman and Hall.

See Also

Spending function overview, gsDesign, gsDesign package overview

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
# set up alpha spending and beta spending to be piecewise linear
sfupar <- c(.2, .4, .05, .2)
sflpar <- c(.3, .5, .65, .5, .75, .9)
x <- gsDesign(sfu=sfLinear, sfl=sfLinear, sfupar=sfupar, sflpar=sflpar)
plot(x, plottype="sf")
x

# now do an example where there is no lower-spending at interim 1
# and no upper spending at interim 2
sflpar<-c(1/3,2/3,0,.25)
sfupar<-c(1/3,2/3,.1,.1)
x <- gsDesign(sfu=sfLinear, sfl=sfLinear, sfupar=sfupar, sflpar=sflpar)
plot(x, plottype="sf")
x

# now do an example where timing of interims changes slightly, but error spending does not
# also, spend all alpha when at least >=90 percent of final information is in the analysis
sfupar=c(.2,.4,.9,((1:3)/3)^3)
x <- gsDesign(k=3,n.fix=100,sfu=sfStep,sfupar=sfupar,test.type=1)
plot(x,pl="sf")
# original planned sample sizes
ceiling(x$n.I)
# cumulative spending planned at original interims
cumsum(x$upper$spend)
# change timing of analyses; 
# note that cumulative spending "P(Cross) if delta=0" does not change from cumsum(x$upper$spend)
# while full alpha is spent, power is reduced by reduced sample size
y <- gsDesign(k=3, sfu=sfStep, sfupar=sfupar, test.type=1,
              maxn.IPlan=x$n.I[x$k], n.I=c(30,70,95),
              n.fix=x$n.fix)
# note that full alpha is used, but power is reduced due to lowered sample size
gsBoundSummary(y)

# now show how step function can be abused by 'adapting' stage 2 sample size based on interim result
x <- gsDesign(k=2,delta=.05,sfu=sfStep,sfupar=c(.02,.001),timing=.02,test.type=1)
# spending jumps from miniscule to full alpha at first analysis after interim 1
plot(x, pl="sf")
# sample sizes at analyses:
ceiling(x$n.I)
# simulate 1 million stage 1 sum of 178 Normal(0,1) random variables 
# Normal(0,Variance=178) under null hypothesis
s1 <- rnorm(1000000,0,sqrt(178))
# compute corresponding z-values
z1 <- s1/sqrt(178)
# set stage 2 sample size to 1 if z1 is over final bound, otherwise full sample size
n2 <- array(1,1000000)
n2[z1<1.96]<- ceiling(x$n.I[2])-ceiling(178)
# now sample n2 observations for second stage
s2 <- rnorm(1000000,0,sqrt(n2))
# add sum and divide by standard deviation
z2 <- (s1+s2)/(sqrt(178+n2))
# By allowing full spending when final analysis is either 
# early or late depending on observed interim z1, 
# Type I error is now almost twice the planned .025
sum(z1 >= x$upper$bound[1] | z2 >= x$upper$bound[2])/1000000
# if stage 2 sample size is random and independent of z1 with same frequency, 
# this is not a problem
s1alt <- rnorm(1000000,0,sqrt(178))
z1alt <- s1alt / sqrt(178)
z2alt <- (s1alt+s2)/sqrt(178+n2)
sum(z1alt >= x$upper$bound[1] | z2alt >= x$upper$bound[2])/1000000

keaven/gsDesign documentation built on Sept. 12, 2017, 9:24 p.m.