Overview

This document is intended to evaluate statistical significance for graphical multiplicity control when used with group sequential design (Maurer and Bretz 2013). In particular, we demonstrate design and analysis of a complex oncology trial. There are many details building on the necessarily simple example provided by Maurer and Bretz (2013). The combination of tools provided by the gMCP and gsDesign packages is non-trivial, but developed in a way that is meant to be re-used in a straightforward fashion. This has been found to be particularly valuable to provide a prompt and verifiable conclusion in multiple trials such as Burtness et al. (2019) where 14 hypotheses were evaluated using a template such as this.

Given the complexity involved, substantial effort has been taken to provide methods to check hypothesis testing.

The table of contents above lays out the organization of the document. In short, we begin with 1) design specification followed by 2) results entry which includes event counts and nominal p-values for testing, 3) carrying out hypothesis testing, and 4) verification of the hypothesis testing results.

Design

For the template example, there are 3 endpoints and 2 populations resulting in 6 hypotheses to be tested in the trial. The endpoints are:

The populations to be studied are:

For simplicity, we design assuming the control group has an exponential time to event with a median of 12 months for OS and 5 months for PFS. We design under a proportional hazards assumption. ORR for the control group is assumed to be 15%. Some of the choices here are arbitrary, but the intent is to fully specify how patients will be enrolled and followed for \(\alpha\)-controlled study analyses.

The following design characteristics are also specified to well-characterize outcomes for all subjects by the end of the trial:

The sample size for the trial will be driven by an adequate sample size and targeted events in the subgroup to ensure 90% power for the OS endpoint assuming a hazard ratio of 0.65. For group sequential designs, we assume 1-sided testing.

To reveal code blocks for the remainder of the document, press the code buttons indicated throughout. The initial code block sets options and loads needed packages; no modification should be required by the user.

### THERE SHOULD BE NO NEED TO MODIFY THIS CODE SECTION
options(scipen = 999)
# Colorblind palette
cbPalette <- c("#999999", "#E69F00", "#56B4E9", "#009E73", "#F0E442", "#0072B2", "#D55E00", "#CC79A7")
# 3 packages used for data storage and manipulation: dplyr, tibble
library(dplyr)
library(tibble)
# 2 packages used for R Markdown capabilities: knitr, kableExtra
library(knitr)
library(kableExtra)
library(gt)
library(ggplot2) # For plotting
library(gsDesign) # Group sequential design capabilities
library(gMCP) # Multiplicity evaluation

Multiplicity diagram for hypothesis testing

Following is the multiplicity graph for the trial design. We have arbitrarily split Type I error equally between the subgroup and overall populations. Most \(\alpha\) is allocated to OS and the least to ORR, with PFS receiving an intermediate amount. This reflects the priority of the endpoints as well as the practicality to detect clinically significant differences in each population. Reallocation for each endpoint proceeds from the subgroup to the overall population. If the overall population hypothesis is rejected for a given endpoint, the reallocation is split between the two populations for another endpoint. The choice for allocation and reallocation illustrated here is to demonstrate a complex multiplicity scenario; when actually applying this method the allocation and realloation choices should be carefully considered.

### THIS CODE NEEDS TO BE MODIFIED FOR YOUR STUDY
# If needed, see help file for gsDesign::hGraph() for explanation of parameters below
# Hypothesis names
nameHypotheses <- c(
  "H1: OS\n Subgroup",
  "H2: OS\n All subjects",
  "H3: PFS\n Subgroup",
  "H4: PFS\n All subjects",
  "H5: ORR\n Subgroup",
  "H6: ORR\n All subjects"
)
# Number of hypotheses to be tested
nHypotheses <- length(nameHypotheses)
# Transition weights for alpha reallocation (square matrix)
m <- matrix(c(
  0, 1, 0, 0, 0, 0,
  0, 0, .5, .5, 0, 0,
  0, 0, 0, 1, 0, 0,
  0, 0, 0, 0, .5, .5,
  0, 0, 0, 0, 0, 1,
  .5, .5, 0, 0, 0, 0
), nrow = 6, byrow = TRUE)
# Initial Type I error assigned to each hypothesis (one-sided)
alphaHypotheses <- c(.01, .01, .004, 0.000, 0.0005, .0005)
fwer <- sum(alphaHypotheses)
# Make a ggplot representation of the above specification and display it
g <- gsDesign::hGraph(6,
  alphaHypotheses = alphaHypotheses, m = m, nameHypotheses = nameHypotheses,
  halfWid = 1, halfHgt = .35, xradius = 2.5, yradius = 1, offset = 0, trhw = .15,
  x = c(-1.25, 1.25, -2.5, 2.5, -1.25, 1.25), y = c(2, 2, 1, 1, 0, 0),
  trprop = 0.4, fill = as.character(c(2, 2, 4, 4, 3, 3))
) + scale_fill_manual(values = cbPalette)
print(g)

This testing scheme can result in what might be referred to as time travel for passing of \(\alpha\). That is, if PFS hypotheses are not rejected at a given analysis (say final PFS analysis) and OS hypotheses are rejected at the final analysis, then the previously evaluated PFS tests at the interim and final PFS analysis can be compared to updated bounds based on reallocated Type I error. While this practice was not encouraged by Maurer and Bretz (2013), it was acknowledged to control Type I error as previous discussed in (Liu and Anderson 2008). Given the stringent Type I error control for multiple hypotheses, the ability to acknowledge clinically significant differences as statistically significant can be important in describing treatment benefits in regulatory labeling for a treatment.

Group sequential designs for each hypothesis

For the example, we assume 1-sided testing or a non-binding futility bound as required for a Maurer and Bretz (2013) design using group sequential design with graphical multiplicity control. Each is demonstrated in the example code for respective hypotheses. Efficacy \(\alpha\)-spending for all group sequential designs uses the Lan and DeMets (1983) spending function approximating an O’Brien-Fleming bound.

This section needs to be modified by the user to match the study design under consideration. Those uncomfortable with coding may wish to design using the gsDesign Shiny app which provides not only a point and click interface, but also a code tab that generates R code that can be copied and plugged in for designs below.

H1: OS, Subgroup

We assume 50% of the population is in the subgroup of interest. A sample size of 378 is driven by overall survival (OS) in the subgroup where we assume a hazard ratio of 0.65. Here we assume a one-sided group sequential design with no futility bound.

osmedian <- 12 # Median control survival
# Derive group sequential design for OS in the targeted subgroup
ossub <- gsDesign::gsSurv(
  k = 3, # 3 analyses for OS
  test.type = 1, # Efficacy bound only (no futility)
  alpha = alphaHypotheses[1], # Allocated alpha from design hypothesis group
  beta = 0.1, # Type 2 error (1 - power)
  hr = 0.65, # Assumed hazard ratio for power calculation
  timing = c(0.61, 0.82), # Choose these to match targeted calendar timing of analyses
  sfu = sfLDOF, # Spending function to approximate O'Brien-Fleming bound
  lambdaC = log(2) / osmedian, # Exponential control failure rate
  eta = 0.001, # Exponential dropout rate
  gamma = c(2.5, 5, 7.5, 10), # Relative enrollment rates by time period
  R = c(2, 2, 2, 12), # Duration of time periods for rates in gamma
  T = 42, # Planned study duration for OS
  minfup = 24 # Planned minimum follow-up after end of enrollment
)
tab <- gsDesign::gsBoundSummary(ossub)
rownames(tab) <- 1:nrow(tab)
cat(summary(ossub))

One-sided group sequential design with 3 analyses, time-to-event outcome with sample size 378 and 284 events required, 90 percent power, 1 percent (1-sided) Type I error to detect a hazard ratio of 0.65. Enrollment and total study durations are assumed to be 18 and 42 months, respectively. Efficacy bounds derived using a Lan-DeMets O’Brien-Fleming approximation spending function with none = 1.

The above text was automatically generated and could be edited appropriately for description of the design. Following is a summary table describing study bounds.

# tab %>% kable(caption = "Design for OS in the subgroup.") %>% kable_styling()
tab %>%
  gt() %>%
  tab_header(title = "Design for OS in the Subgroup") %>%
  cols_align(align = "left", columns = Value) %>%
  tab_footnote(
    footnote = "Cumulative boundary crossing probability includes crossing probability at earlier analysis.",
    locations = cells_body(columns = "Value", rows = c(9, 10, 14, 15))
  ) %>%
  tab_footnote(
    footnote = "Approximate hazard ratio at bound.",
    locations = cells_body(columns = "Value", rows = c(3, 8, 13))
  )
Design for OS in the Subgroup
Analysis Value Efficacy
IA 1: 61% Z 3.0981
N: 378 p (1-sided) 0.0010
Events: 173 ~HR at bound1 0.6243
Month: 24 P(Cross) if HR=1 0.0010
P(Cross) if HR=0.65 0.3986
IA 2: 82% Z 2.6404
N: 378 p (1-sided) 0.0041
Events: 233 ~HR at bound1 0.7073
Month: 32 P(Cross) if HR=12 0.0044
P(Cross) if HR=0.652 0.7469
Final Z 2.3825
N: 378 p (1-sided) 0.0086
Events: 284 ~HR at bound1 0.7535
Month: 42 P(Cross) if HR=12 0.0100
P(Cross) if HR=0.652 0.9000
1 Approximate hazard ratio at bound.
2 Cumulative boundary crossing probability includes crossing probability at earlier analysis.

H2: OS, All

The total sample size is assumed to be twice the above, N=756. The power and hazard ratio can be adjusted to appropriately size the trial rather than starting with adjusting sample size to reach a targeted power. For this example, we consider altering power (beta) while fixing the hazard ratio at 0.75, representing an increase in median OS from 12 months in the control group to 16 months in the experimental group. For this design, we consider a non-binding futility bound where the trial may be stopped early in the overall population if the bound is crossed. We use a Hwang, Shih, and De Cani (1990) bound with \(\gamma = -3.25\). Study designers should carefully consider implication for parameter choices, particularly if the futility bounds provide sensible guidance for stopping the trial. Since the futility bounds are non-binding, the efficacy bound is computed assuming the futility bound is ignored which will control Type I error at the targeted level even if a futility bound is crossed and the trial is continued.

hr <- .75
beta <- .14
os <- gsDesign::gsSurv(
  k = 3, test.type = 4, alpha = 0.01, beta = beta, hr = hr,
  timing = c(0.62, 0.83), sfu = sfLDOF,
  sfl = sfHSD, sflpar = -3.25,
  lambdaC = log(2) / 12, eta = 0.001, S = NULL,
  gamma = c(2.5, 5, 7.5, 10), R = c(2, 2, 2, 12),
  T = 42, minfup = 24
)
tab <- gsDesign::gsBoundSummary(os)
rownames(tab) <- 1:nrow(tab)
cat(summary(os))

Asymmetric two-sided group sequential design with non-binding futility bound, 3 analyses, time-to-event outcome with sample size 756 and 589 events required, 86 percent power, 1 percent (1-sided) Type I error to detect a hazard ratio of 0.75. Enrollment and total study durations are assumed to be 18 and 42 months, respectively. Efficacy bounds derived using a Lan-DeMets O’Brien-Fleming approximation spending function with none = 1. Futility bounds derived using a Hwang-Shih-DeCani spending function with gamma = -3.25.

tab %>%
  kable(caption = "Design for OS in all subjects") %>%
  kable_styling()
Design for OS in all subjects
Analysis Value Efficacy Futility
IA 1: 62% Z 3.0699 0.9611
N: 756 p (1-sided) 0.0011 0.1682
Events: 365 ~HR at bound 0.7250 0.9042
Month: 24 P(Cross) if HR=1 0.0011 0.8318
P(Cross) if HR=0.75 0.3750 0.0367
IA 2: 83% Z 2.6231 1.7013
N: 756 p (1-sided) 0.0044 0.0444
Events: 489 ~HR at bound 0.7886 0.8573
Month: 32 P(Cross) if HR=1 0.0047 0.9585
P(Cross) if HR=0.75 0.7160 0.0782
Final Z 2.3857 2.3857
N: 756 p (1-sided) 0.0085 0.0085
Events: 589 ~HR at bound 0.8214 0.8214
Month: 42 P(Cross) if HR=1 0.0094 0.9906
P(Cross) if HR=0.75 0.8600 0.1400

We can also plot different design characteristics. Here, we plot the approximate hazard ratio to cross each bound which may be helpful for design team discussions.

plot(os, plottype = "HR", xlab = "Events")

H3: PFS, Subgroup

For progression free survival (PFS) we assume a shorter median time to event of 5 months. With an assumed hazard ratio of 0.65, we adjust beta and timing to match the targeted sample size and interim analysis timing. We assume a larger dropout rate for PFS than we did for OS. Here we set up a futility bound for safety. This is an asymmetric 2-sided design with both futility and efficacy boundary crossing probabilities under the null hypothesis. The parameter astar = 0.1 specifies total lower bound spending of 10%. The lower Hwang, Shih, and De Cani (1990) spending bound with \(\gamma = -8\) is intended to be conservative in terms of futility at the interim, but still provide a safety bound for the PFS result in this targeted population. This would have to be carefully evaluated by the study design team at the time of design.

hr <- .65
beta <- .149
pfssub <- gsDesign::gsSurv(
  k = 2, test.type = 6, astar = 0.1, alpha = 0.004, beta = beta, hr = hr,
  timing = .87, sfu = sfLDOF,
  sfl = sfHSD, sflpar = -8,
  lambdaC = log(2) / 5, eta = 0.02, S = NULL,
  gamma = c(2.5, 5, 7.5, 10), R = c(2, 2, 2, 12),
  T = 32, minfup = 14
)
tab <- gsDesign::gsBoundSummary(pfssub)
rownames(tab) <- 1:nrow(tab)
cat(summary(pfssub))

Asymmetric two-sided group sequential design with non-binding futility bound, 2 analyses, time-to-event outcome with sample size 378 and 296 events required, 85.1 percent power, 0.4 percent (1-sided) Type I error to detect a hazard ratio of 0.65. Enrollment and total study durations are assumed to be 18 and 32 months, respectively. Efficacy bounds derived using a Lan-DeMets O’Brien-Fleming approximation spending function with none = 1. Futility bounds derived using a Hwang-Shih-DeCani spending function with gamma = -8.

tab %>%
  kable(caption = "Design for PFS in the subgroup") %>%
  kable_styling()
Design for PFS in the subgroup
Analysis Value Efficacy Futility
IA 1: 87% Z 2.8734 -1.8077
N: 378 p (1-sided) 0.0020 0.9647
Events: 258 ~HR at bound 0.6988 1.2529
Month: 24 P(Cross) if HR=1 0.0020 0.0353
P(Cross) if HR=0.65 0.7240 0.0000
Final Z 2.7062 -1.2904
N: 378 p (1-sided) 0.0034 0.9016
Events: 296 ~HR at bound 0.7299 1.1620
Month: 32 P(Cross) if HR=1 0.0040 0.1000
P(Cross) if HR=0.65 0.8510 0.0000

H4: PFS, All

Finally, we design for PFS in all subjects. In this case, we simplify to a one-sided design. A futility bound could be considered, if appropriate.

hr <- .74
beta <- .15
pfs <- gsDesign::gsSurv(
  k = 2, test.type = 1, alpha = 0.004, beta = beta, hr = hr,
  timing = .86, sfu = sfLDOF,
  lambdaC = log(2) / 5, eta = 0.02, S = NULL,
  gamma = c(2.5, 5, 7.5, 10), R = c(2, 2, 2, 12),
  T = 32, minfup = 14
)
tab <- gsDesign::gsBoundSummary(pfs)
rownames(tab) <- 1:nrow(tab)
tab %>%
  kable(caption = "Design for PFS in the overall population") %>%
  kable_styling()
Design for PFS in the overall population
Analysis Value Efficacy
IA 1: 86% Z 2.8924
N: 756 p (1-sided) 0.0019
Events: 522 ~HR at bound 0.7761
Month: 24 P(Cross) if HR=1 0.0019
P(Cross) if HR=0.74 0.7092
Final Z 2.7032
N: 756 p (1-sided) 0.0034
Events: 606 ~HR at bound 0.8028
Month: 32 P(Cross) if HR=1 0.0040
P(Cross) if HR=0.74 0.8500

H5 and H6: ORR

For objective response rate (ORR), we assume an underlying control rate of 15%. In the subgroup population, we have almost 90% power to detect a 20% improvement.

nBinomial(p1 = .35, p2 = .15, alpha = .0005, n = 378)
#> [1] 0.8911724

In the all subjects population, we have approximately 95% power to detect an improvement in ORR from 15% to 30%.

nBinomial(p1 = .3, p2 = .15, alpha = .0005, n = 756)
#> [1] 0.9530369

Design list

Now we associate designs with hypotheses in an ordered list corresponding to the order in the multiplicity graph setup. Since ORR designs are not group sequential, we enter NULL values for those in the last 2 entries of the design list; hit code button to reveal code for this.

### THIS NEEDS TO BE MODIFIED TO MATCH STUDY
gsDlist <- list(ossub, os, pfssub, pfs, NULL, NULL)

Spending plan and spending time

While it was relatively straightforward above to set up timing of analyses to match for the different hypotheses, accumulation of endpoints can vary from plan in a variety of ways. Planning on how to deal with this is critical at the time of protocol development to avoid later amendments or inappropriate \(\alpha\)-allocation to early analyses. Before going into examples, we review the concept of \(\alpha\)-spending and what we will refer to as spending time.

For a given hypothesis, we will assign a non-decreasing spending function \(f(t)\) defined for \(t\ge 0\) with \(f(0)=0\) and \(f(t)=\alpha\) for \(t\ge 1\). We will assume \(K\) analyses with observed event counts \(n_k\) at analysis \(k=1,2,\ldots,K\) and a targeted final event count of \(N_k\). The \(\alpha\)-spending at analysis \(k\) was originally defined (Lan and DeMets 1983) as \(f(t_k=n_k/N_K)\). The values \(n_k/N_K\) will be referred to as the information fraction, \(k=1,\ldots,K\). This is used to pre-specify the cumulative amount of Type I error for a hypothesis at each analysis. In Lan and DeMets (1989) they noted that calendar time was another option for \(t_k\) values, \(k=1,\ldots,K.\) Proschan, Lan, and Wittes (2006) noted further that as long as \(t_k\) is increasing with \(k\), it can be used to define spending; this is subject to the requirement that under the null hypothesis, the timing must be selected in a way that is not correlated with the test statistic (e.g., blinded). We will refer to \(t_k\), regardless of its definition, as the spending time for a hypothesis. Note that the joint distribution of interim and final tests for a hypothesis is driven by \(n_k\), \(k=1,\ldots,K\). This is equivalent to basing correlation on the information fraction \(n_k^{(actual)}/n_K^{(planned)}\), \(1\le k\le K\). Thus, both spending time and information fraction are required to compute bounds for group sequential testing. Our general objectives here will be to:

  • Spend all Type I error for each hypothesis in its combined interim and final analyses; this requires the spending time to be 1 for the final analysis of a hypothesis.
  • Ensure spending time is well defined for each analysis of each hypothesis.
  • We will assume that both follow-up duration and event counts may be of interest in determining timing of analyses; e.g., for immuno-oncology therapies there have been delayed treatment effects and the tail of the time-to-event distribution has been important to establish benefit. Thus, we will assume here that over-spending at interim analysis is to be avoided.

Here we assume that the subgroup prevalence was over-estimated in the study design and indicating how spending time can be used to deal with this deviation from plan.

Results entry at time of analysis

Results for each analysis performed should be entered here. We begin by documenting timing and event counts of each analysis. Then we proceed to enter nominal 1-sided testing p-values for each analysis of each hypothesis.

Timing of analyses and resulting event counts and spending times

Recall that the design assumed 50% prevalence of the subgroup. Here we assume that the observed prevalence is 40% and that, by specification stated above, we enroll until the targeted subpopulation of 378 is achieved. This is assumed to occur after 22 months with a total enrollment of 940. Timing of analyses is now targeted as follows:

  • The first interim is scheduled 28 months, 6 months after final enrollment.
  • The second interim is scheduled at the later of 14 months after final enrollment (22 + 14 = 36 months after start of enrollment) or the targeted final PFS event count of 297 events. We assume the event count is reached at 34 months and that the achieved final event count is 320 in the subgroup at 36 months.
  • The final analysis is scheduled at 24 months after final enrollment (month 22 + 24 = 46) or when 284 events have been observed in the subgroup, whichever comes first; there is also the qualification that the final analysis will be no more than 30 months after final enrollment (6 months after targeted time). We assume the targeted event count is not reached by 6 months after the targeted final analysis time and, thus, the final analysis cutoff is set at month 22 + 30 = 52 and that at that time 270 OS events have been observed in the subgroup.

All of the above leads to event counts and spending for PFS and OS as follows:

### THIS NEEDS TO BE MODIFIED TO MATCH YOUR STUDY
# PFS, overall population
pfs$n.I <- c(675, 750)
# PFS, subgroup
pfssub$n.I <- c(265, 310)
# OS, overall population
os$n.I <- c(529, 700, 800)
# OS, subgroup
ossub$n.I <- c(185, 245, 295)

Nominal p-values for each analysis

For analyses not yet performed enter dummy values, including a p-value near 1 (e.g., .99). No other entry is required by the user in any other section of the document. Calendar timing is also associated with PFS hypotheses for use in spending functions. Spending time for OS spending will be input as NULL so that spending will be based on event counts for OS hypotheses.

### THIS NEEDS TO BE MODIFIED TO MATCH YOUR STUDY
inputResults <- tibble(
  H = c(rep(1, 3), rep(2, 3), rep(3, 2), rep(4, 2), 5, 6),
  Pop = c(
    rep("Subgroup", 3), rep("All", 3),
    rep("Subgroup", 2), rep("All", 2),
    "Subgroup", "All"
  ),
  Endpoint = c(rep("OS", 6), rep("PFS", 4), rep("ORR", 2)),
  # Example with some rejections
  nominalP = c(
    .03, .0001, .000001,
    .2, .15, .1,
    .2, .001,
    .3, .2,
    .00001,
    .1
  ),
  # Example with no rejections
  # nominalP = rep(.03, 12),
  Analysis = c(1:3, 1:3, 1:2, 1:2, 1, 1),
  events = c(ossub$n.I, os$n.I, pfssub$n.I, pfs$n.I, NA, NA),
  spendingTime = c(
    ossub$n.I / max(ossub$n.I),
    ossub$n.I / max(ossub$n.I),
    pfssub$n.I / max(pfssub$n.I),
    pfssub$n.I / max(pfssub$n.I),
    NA, NA
  )
)
kable(inputResults, caption = "DUMMY RESULTS FOR IA2.") %>%
  kable_styling() %>%
  add_footnote("Dummy results", notation = "none")
DUMMY RESULTS FOR IA2.
H Pop Endpoint nominalP Analysis events spendingTime
1 Subgroup OS 0.030000 1 185 0.6271186
1 Subgroup OS 0.000100 2 245 0.8305085
1 Subgroup OS 0.000001 3 295 1.0000000
2 All OS 0.200000 1 529 0.6271186
2 All OS 0.150000 2 700 0.8305085
2 All OS 0.100000 3 800 1.0000000
3 Subgroup PFS 0.200000 1 265 0.8548387
3 Subgroup PFS 0.001000 2 310 1.0000000
4 All PFS 0.300000 1 675 0.8548387
4 All PFS 0.200000 2 750 1.0000000
5 Subgroup ORR 0.000010 1 NA NA
6 All ORR 0.100000 1 NA NA
Dummy results

Testing hypotheses

Compute sequential p-values for each hypothesis

Sequential p-value computation is done in one loop in an attempt to minimize chances for coding errors. We delay showing these until after display of the sequence of multiplicity graphs generated by hypothesis rejection is shown.

### USER SHOULD NOT NEED TO MODIFY THIS CODE
EOCtab <- NULL
EOCtab <- inputResults %>%
  group_by(H) %>%
  slice(1) %>%
  ungroup() %>%
  select("H", "Pop", "Endpoint", "nominalP")
EOCtab$seqp <- .9999
for (EOCtabline in 1:nHypotheses) {
  EOCtab$seqp[EOCtabline] <-
    ifelse(is.null(gsDlist[[EOCtabline]]), EOCtab$nominalP[EOCtabline], {
      tem <- filter(inputResults, H == EOCtabline)
      sequentialPValue(
        gsD = gsDlist[[EOCtabline]], interval = c(.0001, .9999),
        n.I = tem$events,
        Z = -qnorm(tem$nominalP),
        usTime = tem$spendingTime
      )
    })
}
EOCtab <- EOCtab %>% select(-"nominalP")
# kable(EOCtab,caption="Sequential p-values as initially placed in EOCtab") %>% kable_styling()

Evaluate hypothesis rejection using gMCP

We need to set up a graph object as implemented in the gMCP package.

# Make a graph object
rownames(m) <- nameHypotheses
graph <- matrix2graph(m)
# Add weights to the object based on alpha allocation
graph <- setWeights(graph, alphaHypotheses / fwer)
rescale <- 45
d <- g$layers[[2]]$data
rownames(d) <- rownames(m)
# graph@nodeAttr$X <- rescale * d$x * 1.75
# graph@nodeAttr$Y <- -rescale * d$y * 2

Now we add the sequential p-values and evaluate which hypotheses have been rejected.

result <- gMCP(graph = graph, pvalues = EOCtab$seqp, alpha = fwer)
result@rejected
#>      H1: OS\n Subgroup  H2: OS\n All subjects     H3: PFS\n Subgroup 
#>                   TRUE                  FALSE                   TRUE 
#> H4: PFS\n All subjects     H5: ORR\n Subgroup H6: ORR\n All subjects 
#>                  FALSE                   TRUE                  FALSE
# now map back into EOCtable (CHECK AGAIN!!!)
EOCtab$Rejected <- result@rejected
EOCtab$adjPValues <- result@adjPValues

Verification of hypotheses rejected

# Number of graphs is used repeatedly
ngraphs <- length(result@graphs)
# Set up tibble with hypotheses rejected at each stage
rejected <- NULL
for (i in 1:length(result@graphs)) {
  rejected <- rbind(
    rejected,
    tibble(
      H = 1:nHypotheses, Stage = i,
      Rejected = as.logical(result@graphs[[i]]@nodeAttr$rejected)
    )
  )
}
rejected <- rejected %>%
  filter(Rejected) %>%
  group_by(H) %>%
  summarize(graphRejecting = min(Stage) - 1, .groups = "drop") %>% # Last graph with weight>0 where H rejected
  arrange(graphRejecting)

# Get final weights
# for hypotheses not rejected, this will be final weight where
# no hypothesis could be rejected
lastWeights <- as.numeric(result@graphs[[ngraphs]]@weights)
lastGraph <- rep(ngraphs, nrow(EOCtab))

# We will update for rejected hypotheses with last positive weight for each
if (ngraphs > 1) {
  for (i in 1:(ngraphs - 1)) {
    lastWeights[rejected$H[i]] <- as.numeric(result@graphs[[i]]@weights[rejected$H[i]])
    lastGraph[rejected$H[i]] <- i
  }
}
EOCtab$lastAlpha <- fwer * lastWeights
EOCtab$lastGraph <- lastGraph
EOCtabx <- EOCtab
names(EOCtabx) <- c(
  "Hypothesis", "Population", "Endpoint", "Sequential p",
  "Rejected", "Adjusted p", "Max alpha allocated", "Last Graph"
)
# Display table with desired column order
# Delayed following until after multiplicity graph sequence
# EOCtabx %>% select(c(1:4,7,5:6,8)) %>% kable() %>% kable_styling()

Multiplicity graph sequence from gMCP

### THERE SHOULD BE NO NEED TO MODIFY THIS CODE SECTION
for (i in 1:ngraphs) {
  mx <- result@graphs[[i]]@m
  rownames(mx) <- NULL
  colnames(mx) <- NULL
  g <- hGraph(
    nHypotheses = nHypotheses,
    alphaHypotheses = result@graphs[[i]]@weights * fwer,
    m = mx,
    nameHypotheses = nameHypotheses,
    halfWid = 1, halfHgt = .35, xradius = 2.5, yradius = 1, offset = 0, trhw = .15,
    x = c(-1.25, 1.25, -2.5, 2.5, -1.25, 1.25), y = c(2, 2, 1, 1, 0, 0),
    trprop = .4, fill = as.character(c(2, 2, 4, 4, 3, 3))
  ) +
    scale_fill_manual(values = cbPalette)
  cat(" \n")
  cat("####", paste(" Graph", as.character(i), " \n\n"))
  print(g)
  cat(" \n\n\n")
}

Graph 1

Graph 2

Graph 3

Graph 4

Comparison of sequential p-values to multiplicity graphs

We can compare sequential p-values to available \(\alpha\) in each graph. In the column ‘Last Graph’ we can see one of 2 things:

  1. For rejected hypotheses, the maximum \(\alpha\) allocated to the hypothesis. For example, hypothesis one was allocated \(\alpha=0.01\) in the first graph above (select using first tab). We see that the sequential p-value of 0.0001 is smaller than \(\alpha=0.01\) and thus the hypothesis is rejected. We can then proceed to the second graph and see that hypothesis 5 was rejected. The last hypothesis rejected is hypothesis 3 in the third graph.
  2. For the remaining hypotheses (H2, H4, H6) the maximum \(\alpha\) allocated is in the fourth graph; since each sequential p-value is greater than the allocated \(\alpha\) for the corresponding hypothesis, none of these hypotheses were rejected.
EOCtabx %>%
  select(c(1:4, 7, 5:6, 8)) %>%
  kable() %>%
  kable_styling()
Hypothesis Population Endpoint Sequential p Max alpha allocated Rejected Adjusted p Last Graph
1 Subgroup OS 0.0001000 0.0100 TRUE 0.0002500 1
2 All OS 0.1232177 0.0200 FALSE 0.1540221 4
3 Subgroup PFS 0.0011310 0.0040 TRUE 0.0070685 3
4 All PFS 0.2355583 0.0040 FALSE 0.2453732 4
5 Subgroup ORR 0.0000100 0.0005 TRUE 0.0005000 2
6 All ORR 0.1000000 0.0010 FALSE 0.2453732 4

Bounds at final \(\alpha\) allocated for group sequential tests

As a separate validation, we examine group sequential bounds for each hypothesis updated with 1) the maximum \(\alpha\) allocated above, 2) the number of events at each analysis, and 3) the cumulative spending at each analysis above. The nominal p-value for at least one of the analyses performed for each rejected hypotheses should be less than or equal to the nominal p-value in the group sequential design. For each hypothesis not rejected, all nominal p-values are greater than the its corresponding bound. For hypotheses tested without a group sequential design, the nominal p-value for the test of that hypothesis can be compared to the maximum alpha allocated in the above table.

for (i in 1:nHypotheses) {
  # Set up tab for hypothesis in output
  cat("####", paste(" Hypothesis", as.character(i), " \n"))
  # Get results for hypothesis
  hresults <- inputResults %>% filter(H == i)
  # Print out max alpha allocated
  xx <- paste("Max alpha allocated from above table: ",
    as.character(EOCtab$lastAlpha[i]),
    sep = ""
  )
  d <- gsDlist[[i]]
  # If not group sequential for this hypothesis, print the max alpha allocated
  # and the nominal p-value
  if (is.null(d)) {
    cat("Maximum alpha allocated: ")
    cat(EOCtab$lastAlpha[i])
    cat("\n\n")
    cat("Nominal p-value for hypothesis test: ")
    cat(hresults$nominalP)
    cat("\n\n")
  }
  # For group sequential tests, print max alpha allocated and
  # corresponding group sequential bounds
  if (!is.null(gsDlist[[i]])) {
    cat("Nominal p-values at each analysis for comparison to bounds in table below:\n\n")
    cat(hresults$nominalP)
    cat("\n\n")
    # Get other info for current hypothesis
    n.I <- hresults$events
    usTime <- hresults$spendingTime
    n.Iplan <- max(d$n.I)
    if (length(n.I) == 1) {
      n.I <- c(n.I, n.Iplan)
      usTime <- c(usTime, 1)
    }
    # If no alpha allocated, just print text line to note this along with the 0 alpha allocated
    if (EOCtab$lastAlpha[i] == 0) {
      cat("Maximum alpha allocated: 0\n\n")
      cat("No testing required\n\n")
    }
    if (EOCtab$lastAlpha[i] > 0) {
      dupdate <- gsDesign::gsDesign(
        alpha = EOCtab$lastAlpha[i],
        k = length(n.I),
        n.I = n.I,
        usTime = usTime,
        maxn.IPlan = n.Iplan,
        n.fix = d$n.fix,
        test.type = 1,
        sfu = d$upper$sf,
        sfupar = d$upper$param
      )
      tabl <- gsDesign::gsBoundSummary(dupdate,
        Nname = "Events",
        exclude = c(
          "B-value", "CP", "CP H1", "Spending",
          "~delta at bound", "P(Cross) if delta=0",
          "PP", "P(Cross) if delta=1"
        )
      )
      kable(tabl, caption = xx, row.names = FALSE) %>%
        kable_styling() %>%
        cat()
      cat("\n\n")
    }
  }
}

Hypothesis 1

Nominal p-values at each analysis for comparison to bounds in table below:

0.03 0.0001 0.000001

Max alpha allocated from above table: 0.01
Analysis Value Efficacy
IA 1: 65% Z 3.0503
Events: 185 p (1-sided) 0.0011
IA 2: 86% Z 2.6238
Events: 245 p (1-sided) 0.0043
Final Z 2.3861
Events: 295 p (1-sided) 0.0085

Hypothesis 2

Nominal p-values at each analysis for comparison to bounds in table below:

0.2 0.15 0.1

Max alpha allocated from above table: 0.02
Analysis Value Efficacy
IA 1: 90% Z 2.7157
Events: 529 p (1-sided) 0.0033
IA 2: 119% Z 2.3386
Events: 700 p (1-sided) 0.0097
Final Z 2.1098
Events: 800 p (1-sided) 0.0174

Hypothesis 3

Nominal p-values at each analysis for comparison to bounds in table below:

0.2 0.001

Max alpha allocated from above table: 0.004
Analysis Value Efficacy
IA 1: 90% Z 2.9023
Events: 265 p (1-sided) 0.0019
Final Z 2.7017
Events: 310 p (1-sided) 0.0034

Hypothesis 4

Nominal p-values at each analysis for comparison to bounds in table below:

0.3 0.2

Max alpha allocated from above table: 0.004
Analysis Value Efficacy
IA 1: 111% Z 2.9023
Events: 675 p (1-sided) 0.0019
Final Z 2.6840
Events: 750 p (1-sided) 0.0036

Hypothesis 5

Maximum alpha allocated: 0.0005

Nominal p-value for hypothesis test: 0.00001

Hypothesis 6

Maximum alpha allocated: 0.001

Nominal p-value for hypothesis test: 0.1

Session information

You can use sessionInfo() to document the versions of R and R packages used to render this document. Note, in particular, that version 3.1 or later of the gsDesign package is needed.

References

Bretz, Frank, Willi Maurer, and Martin Posch. 2009. “A Graphical Approach to Sequentially Rejective Multiple Test Procedures.” Statistics in Medicine 28: 586–604.
Burtness, Barbara, Kevin J Harrington, Richard Greil, Denis Soulières, Makoto Tahara, Gilberto de Castro Jr, Amanda Psyrri, et al. 2019. “Pembrolizumab Alone or with Chemotherapy Versus Cetuximab with Chemotherapy for Recurrent or Metastatic Squamous Cell Carcinoma of the Head and Neck (KEYNOTE-048): A Randomised, Open-Label, Phase 3 Study.” The Lancet 394 (10212): 1915–28.
Hwang, Irving K, Weichung J Shih, and John S De Cani. 1990. “Group Sequential Designs Using a Family of Type i Error Probability Spending Functions.” Statistics in Medicine 9 (12): 1439–45.
Lan, K. K. G., and David L. DeMets. 1983. “Discrete Sequential Boundaries for Clinical Trials.” Biometrika 70: 659–63.
———. 1989. “Group Sequential Procedures: Calendar Versus Information Time.” Statistics in Medicine 8: 1191–98.
Liu, Qing, and Keaven M Anderson. 2008. “On Adaptive Extensions of Group Sequential Trials for Clinical Investigations.” Journal of the American Statistical Association 103 (484): 1621–30.
Maurer, Willi, and Frank Bretz. 2013. “Multiple Testing in Group Sequential Trials Using Graphical Approaches.” Statistics in Biopharmaceutical Research 5: 311–20.
Proschan, Michael A., K. K. Gordon Lan, and Janet Turk Wittes. 2006. Statistical Monitoring of Clinical Trials: A Unified Approach. New York, NY: Springer.