Here, we present the calculations for the initial design of the DEFUSE3 trial based on (Lai, Lavori, and Liao 2014) and (Lai, Lavori, and Tsang 2015). The trial parameters are fixed as follows.
library(ASSISTant)
##Fix randomization vector N, errors, eps
trialParameters <- list(N = c(200, 340, 476), type1Error = 0.025,
eps = 1/2, type2Error = 0.1)The design parameters are the following for various scenarios.
designParameters <- list(
nul0 = list(prevalence = rep(1/6, 6), mean = matrix(0, 2, 6),
sd = matrix(1, 2, 6)),
alt1 = list(prevalence = rep(1/6, 6), mean = rbind(rep(0, 6),
c(0.5, 0.4, 0.3, 0, 0, 0)),
sd = matrix(1, 2, 6)),
alt2 = list(prevalence = rep(1/6, 6), mean = rbind(rep(0, 6),
c(0.5, 0.5, 0, 0, 0, 0)),
sd = matrix(1,2, 6)),
alt3 = list(prevalence = rep(1/6, 6), mean = rbind(rep(0, 6), rep(0.36, 6)),
sd = matrix(1,2, 6)),
alt4 = list(prevalence = rep(1/6, 6), mean = rbind(rep(0, 6), rep(0.30, 6)),
sd = matrix(1,2, 6)),
alt5 = list(prevalence = rep(1/6, 6), mean = rbind(rep(0, 6),
c(0.4, 0.3, 0.2, 0, 0, 0)),
sd = matrix(1,2, 6)),
alt6 = list(prevalence = rep(1/6, 6), mean = rbind(rep(0, 6),
c(0.5, 0.5, 0.3, 0.3, 0.1, 0.1)),
sd = matrix(1,2, 6))
)defuse3 <- DEFUSE3Design$new(trialParameters = trialParameters,
numberOfSimulations = 500,
designParameters = designParameters$nul0,
showProgress = FALSE)
print(defuse3)## Design Parameters:
## Number of Groups: 6
## Prevalence:
##
## Group1 Group2 Group3 Group4 Group5 Group6
## ---------- ---------- ---------- ---------- ---------- ----------
## 0.1666667 0.1666667 0.1666667 0.1666667 0.1666667 0.1666667
##
## Using Discrete Rankin scores? FALSE
##
## Normal Rankin Distribution means (null row, alt. row):
##
##
## Group1 Group2 Group3 Group4 Group5 Group6
## ----- ------- ------- ------- ------- ------- -------
## Null 0 0 0 0 0 0
## Alt 0 0 0 0 0 0
##
## Normal Rankin Distribution SDs (null row, alt. row):
##
##
## Group1 Group2 Group3 Group4 Group5 Group6
## ----- ------- ------- ------- ------- ------- -------
## Null 1 1 1 1 1 1
## Alt 1 1 1 1 1 1
##
## Trial Parameters:
## List of 6
## $ N : num [1:3] 200 340 476
## $ type1Error : num 0.025
## $ eps : num 0.5
## $ type2Error : num 0.1
## $ effectSize : num 0.104
## $ originalEffectSize: num 0.0858
##
## Boundaries:
##
##
## btilde b c
## ---------- --------- ---------
## -1.885434 2.590764 2.758372result <- defuse3$explore(numberOfSimulations = 500,
showProgress = FALSE,
rngSeed = 28912)
analysis <- defuse3$analyze(result)
print(defuse3$summary(analysis))## P(Reject H0_ITT) = 0.002000; P(Reject H0_subgp) = 0.014000; P(Reject H0) = 0.016000
## P(Early stop for efficacy [futility]) = 0.008000 [0.662000]
## Mean [SD] Randomized N = 355.760000 [99.759743]
##
## Stage at exit (proportion)
##
##
## exitStage proportion
## ---------- -----------
## 1 0.208
## 2 0.462
## 3 0.330
##
## Mean [SD] Lost N = 131.356000 [66.434089]
## Mean [SD] Analyzed N = 224.404000 [84.921093]
##
## Mean loss by futility stage and subgroup
##
##
## FutilityStage selectedGroup mean sd
## -------------- -------------- ---------- ----------
## 1 1 167.20359 4.986172
## 1 2 132.76471 6.999185
## 1 3 99.04082 6.304625
## 1 4 67.68750 5.735875
## 1 5 33.23636 5.319300
## 2 1 280.93333 7.362712
## 2 2 226.78947 6.827875
## 2 3 171.25000 11.804963
## 2 4 116.37500 8.309633
## 2 5 55.08000 5.589574
## 3 1 400.50000 7.187953
## 3 2 304.75000 18.300729
## 3 3 233.50000 7.368853
## 3 4 158.33333 10.016653
## 3 5 80.66667 7.679173
##
## Chance of each subpopulation rejected
##
##
## group count proportion
## ------ ------ -----------
## 1 2 0.004
## 2 2 0.004
## 3 2 0.004
## 4 1 0.002
## 6 1 0.002
##
## Counts by futility stage and subgroup choice
##
##
## FutilityStage G1 G2 G3 G4 G5
## -------------- ---- --- --- --- ---
## 1 167 68 49 48 55
## 2 15 19 8 16 25
## 3 4 4 6 3 12
##
## CI Statistics:
## Overall coverage and coverage for rejections:
##
## overall rejection
## -------- ----------
## 0.984 0
##
## P(theta_test is in the confidence interval)
##
##
## coverage selectedCount rejectedCount
## ---------- -------------- --------------
## 0.9892473 186 2
## 0.9780220 91 2
## 0.9682540 63 2
## 0.9850746 67 1
## 1.0000000 92 0
## 0.0000000 1 1
## NULLresult1 <- defuse3$explore(numberOfSimulations = 500,
trueParameters = designParameters$alt1,
showProgress = FALSE,
rngSeed = 737218)
analysis1 <- defuse3$analyze(result1)
print(defuse3$summary(analysis1))## P(Reject H0_ITT) = 0.284000; P(Reject H0_subgp) = 0.558000; P(Reject H0) = 0.842000
## P(Early stop for efficacy [futility]) = 0.436000 [0.018000]
## Mean [SD] Randomized N = 395.776000 [97.950203]
##
## Stage at exit (proportion)
##
##
## exitStage proportion
## ---------- -----------
## 1 0.132
## 2 0.322
## 3 0.546
##
## Mean [SD] Lost N = 135.228000 [109.645375]
## Mean [SD] Analyzed N = 260.548000 [102.724449]
##
## Mean loss by futility stage and subgroup
##
##
## FutilityStage selectedGroup mean sd
## -------------- -------------- ---------- ----------
## 1 1 167.60417 4.569741
## 1 2 134.41176 7.912850
## 1 3 102.96875 5.474188
## 1 4 68.50000 3.728270
## 1 5 32.28571 3.352327
## 2 1 281.89474 5.404676
## 2 2 226.16667 9.318829
## 2 3 168.87879 8.513803
## 2 4 115.60000 7.604093
## 2 5 57.00000 7.958224
## 3 1 396.46154 6.678515
## 3 2 319.29630 7.700224
## 3 3 236.03125 11.188101
## 3 4 162.05556 10.032138
## 3 5 80.42105 8.408433
##
## Chance of each subpopulation rejected
##
##
## group count proportion
## ------ ------ -----------
## 1 69 0.138
## 2 72 0.144
## 3 105 0.210
## 4 23 0.046
## 5 10 0.020
## 6 142 0.284
##
## Counts by futility stage and subgroup choice
##
##
## FutilityStage G1 G2 G3 G4 G5
## -------------- --- --- --- --- ---
## 1 48 34 32 6 7
## 2 19 24 33 10 4
## 3 13 27 64 18 19
##
## CI Statistics:
## Overall coverage and coverage for rejections:
##
## overall rejection
## -------- ----------
## 0.158 0
##
## P(theta_test is in the confidence interval)
##
##
## coverage selectedCount rejectedCount
## ---------- -------------- --------------
## 0.1375000 80 69
## 0.1529412 85 72
## 0.1860465 129 105
## 0.3235294 34 23
## 0.6666667 30 10
## 0.0000000 142 142
## NULLresult2 <- defuse3$explore(numberOfSimulations = 500,
trueParameters = designParameters$alt2,
showProgress = FALSE,
rngSeed = 928812)
analysis2 <- defuse3$analyze(result2)
print(defuse3$summary(analysis2))## P(Reject H0_ITT) = 0.226000; P(Reject H0_subgp) = 0.620000; P(Reject H0) = 0.846000
## P(Early stop for efficacy [futility]) = 0.438000 [0.014000]
## Mean [SD] Randomized N = 402.488000 [89.064712]
##
## Stage at exit (proportion)
##
##
## exitStage proportion
## ---------- -----------
## 1 0.086
## 2 0.366
## 3 0.548
##
## Mean [SD] Lost N = 165.346000 [118.444978]
## Mean [SD] Analyzed N = 237.142000 [103.804569]
##
## Mean loss by futility stage and subgroup
##
##
## FutilityStage selectedGroup mean sd
## -------------- -------------- ---------- ----------
## 1 1 166.85294 5.105745
## 1 2 134.47674 6.722616
## 1 3 100.66667 9.232448
## 1 4 67.75000 4.267820
## 1 5 32.00000 2.943920
## 2 1 282.35294 4.872643
## 2 2 227.91935 8.218927
## 2 3 170.80000 7.450578
## 2 4 117.00000 6.164414
## 2 5 60.66667 8.386497
## 3 1 400.30769 8.606050
## 3 2 316.87368 10.026817
## 3 3 242.84211 7.581140
## 3 4 165.00000 8.194074
## 3 5 80.55556 5.052502
##
## Chance of each subpopulation rejected
##
##
## group count proportion
## ------ ------ -----------
## 1 54 0.108
## 2 216 0.432
## 3 27 0.054
## 4 11 0.022
## 5 2 0.004
## 6 113 0.226
##
## Counts by futility stage and subgroup choice
##
##
## FutilityStage G1 G2 G3 G4 G5
## -------------- --- --- --- --- ---
## 1 34 86 15 8 4
## 2 17 62 10 4 3
## 3 13 95 19 8 9
##
## CI Statistics:
## Overall coverage and coverage for rejections:
##
## overall rejection
## -------- ----------
## 0.154 0
##
## P(theta_test is in the confidence interval)
##
##
## coverage selectedCount rejectedCount
## ---------- -------------- --------------
## 0.1562500 64 54
## 0.1111111 243 216
## 0.3863636 44 27
## 0.4500000 20 11
## 0.8750000 16 2
## 0.0000000 113 113
## NULLThe discretized scenarios are designed to generally mimic the trends above in the alternatives. However, we have a problem: we cannot simulatenously match the mean and sd of the alternatives above. (Actually, we can, but not with Rankin scores 0 through 6. The software can easily be modified to generate discrete values where the values are 0 to 6 divided by the standard deviation of the respective distribution, for example.)
Also in future versions, I need to allow for more general support values for the scores, not just 0 through 6. Easy to do, but not done yet.
Some types of distributions:
null.uniform <- rep(1, 7L) ## uniform on 7 support points
hourglass <- c(1, 2, 2, 1, 2, 2, 1)
inverted.hourglass <- c(2, 1, 1, 2, 1, 1, 2)
bottom.heavy <- c(2, 2, 2, 1, 1, 1, 1)
bottom.heavier <- c(3, 3, 2, 2, 1, 1, 1)
bottom.loaded <- c(4, 4, 3, 3, 2, 1, 1)
top.heavy <- c(1, 1, 1, 1, 2, 2, 2)
top.heavier <- c(1, 1, 1, 2, 2, 3, 3)
top.loaded <- c(1, 1, 2, 3, 3, 4, 4)It is instructive to see what the means and standard deviations are.
knitr::kable(
sapply(list(null = null.uniform,
hourglass = hourglass,
inv.hourglass = inverted.hourglass,
bot.heavy = bottom.heavy,
bot.heavier = bottom.heavier,
bot.loaded = bottom.loaded,
top.heavy = top.heavy,
top.heavier = top.heavier,
top.loaded = top.loaded),
computeMeanAndSD)
)| null | hourglass | inv.hourglass | bot.heavy | bot.heavier | bot.loaded | top.heavy | top.heavier | top.loaded | |
|---|---|---|---|---|---|---|---|---|---|
| mean | 3 | 3.000000 | 3.000000 | 2.400000 | 2.153846 | 2.111111 | 3.600000 | 3.846154 | 3.888889 |
| sd | 2 | 1.858641 | 2.144761 | 1.959592 | 1.874778 | 1.760331 | 1.959592 | 1.874778 | 1.760331 |
With this in mind, we can reel off some runs. Phil, you mentioned you wanted \(J = 2\), which I adhere to, below.
designParameters <- list(
nul0 = list(prevalence = rep(1, 2),
ctlDist = null.uniform,
trtDist = cbind(null.uniform,
null.uniform)),
alt1 = list(prevalence = rep(1, 2),
ctlDist = null.uniform,
trtDist = cbind(top.loaded,
null.uniform)),
alt2 = list(prevalence = rep(1, 2),
ctlDist = null.uniform,
trtDist = cbind(null.uniform,
top.loaded))
)discDefuse3 <- DEFUSE3Design$new(trialParameters = trialParameters,
numberOfSimulations = 5000,
discreteData = TRUE,
designParameters = designParameters$nul0,
showProgress = FALSE)
print(discDefuse3)## Design Parameters:
## Number of Groups: 2
## Prevalence:
##
## Group1 Group2
## ------- -------
## 0.5 0.5
##
## Using Discrete Rankin scores? TRUE
##
## Null Rankin Distribution:
##
## Prob.
## --- ----------
## 0 0.1428571
## 1 0.1428571
## 2 0.1428571
## 3 0.1428571
## 4 0.1428571
## 5 0.1428571
## 6 0.1428571
## Null Distribution Mean: 3.000000, SD: 2.000000
##
## Alternative Rankin Distribution:
##
##
## Group1 Group2
## --- ---------- ----------
## 0 0.1428571 0.1428571
## 1 0.1428571 0.1428571
## 2 0.1428571 0.1428571
## 3 0.1428571 0.1428571
## 4 0.1428571 0.1428571
## 5 0.1428571 0.1428571
## 6 0.1428571 0.1428571
## Alternative Mean and SD
##
## Group1 Group2
## ----- ------- -------
## mean 3 3
## sd 2 2
##
## Trial Parameters:
## List of 6
## $ N : num [1:3] 200 340 476
## $ type1Error : num 0.025
## $ eps : num 0.5
## $ type2Error : num 0.1
## $ effectSize : num 0.106
## $ originalEffectSize: num 0.0858
##
## Boundaries:
##
##
## btilde b c
## ---------- -------- ---------
## -1.875887 2.48671 2.467753result <- discDefuse3$explore(numberOfSimulations = 50,
showProgress = FALSE,
rngSeed = 3783)
analysis <- discDefuse3$analyze(result)
print(discDefuse3$summary(analysis))## P(Reject H0_ITT) = 0.000000; P(Reject H0_subgp) = 0.020000; P(Reject H0) = 0.020000
## P(Early stop for efficacy [futility]) = 0.000000 [0.800000]
## Mean [SD] Randomized N = 305.600000 [106.616995]
##
## Stage at exit (proportion)
##
##
## exitStage proportion
## ---------- -----------
## 1 0.44
## 2 0.36
## 3 0.20
##
## Mean [SD] Lost N = 116.780000 [38.398709]
## Mean [SD] Analyzed N = 188.820000 [86.822054]
##
## Mean loss by futility stage and subgroup
##
##
## FutilityStage selectedGroup mean sd
## -------------- -------------- --------- ---------
## 1 1 99.5000 7.452413
## 2 1 165.4286 9.589180
## 3 1 233.6667 4.041452
##
## Chance of each subpopulation rejected
##
##
## group count proportion
## ------ ------ -----------
## 1 1 0.02
##
## Counts by futility stage and subgroup choice
##
##
## FutilityStage G1
## -------------- ---
## 1 40
## 2 7
## 3 3
##
## CI Statistics:
## Overall coverage and coverage for rejections:
##
## overall rejection
## -------- ----------
## 1 1
##
## P(theta_test is in the confidence interval)
##
##
## coverage selectedCount rejectedCount
## --------- -------------- --------------
## 1 50 1
## NULLresult1 <- discDefuse3$explore(numberOfSimulations = 50,
trueParameters = designParameters$alt1,
showProgress = FALSE,
rngSeed = 28912)
analysis1 <- discDefuse3$analyze(result1)
print(discDefuse3$summary(analysis1))## P(Reject H0_ITT) = 0.520000; P(Reject H0_subgp) = 0.400000; P(Reject H0) = 0.920000
## P(Early stop for efficacy [futility]) = 0.560000 [0.000000]
## Mean [SD] Randomized N = 374.640000 [103.343643]
##
## Stage at exit (proportion)
##
##
## exitStage proportion
## ---------- -----------
## 1 0.18
## 2 0.38
## 3 0.44
##
## Mean [SD] Lost N = 83.040000 [97.380164]
## Mean [SD] Analyzed N = 291.600000 [89.833860]
##
## Mean loss by futility stage and subgroup
##
##
## FutilityStage selectedGroup mean sd
## -------------- -------------- --------- ----------
## 1 1 100.5556 9.951270
## 2 1 169.2500 9.673848
## 3 1 233.6364 15.298841
##
## Chance of each subpopulation rejected
##
##
## group count proportion
## ------ ------ -----------
## 1 20 0.40
## 2 26 0.52
##
## Counts by futility stage and subgroup choice
##
##
## FutilityStage G1
## -------------- ---
## 1 9
## 2 4
## 3 11
##
## CI Statistics:
## Overall coverage and coverage for rejections:
##
## overall rejection
## -------- ----------
## 1 1
##
## P(theta_test is in the confidence interval)
##
##
## coverage selectedCount rejectedCount
## --------- -------------- --------------
## 1 24 20
## 1 26 26
## NULLresult2 <- discDefuse3$explore(numberOfSimulations = 50,
trueParameters = designParameters$alt2,
showProgress = FALSE,
rngSeed = 931)
analysis2 <- discDefuse3$analyze(result2)
print(discDefuse3$summary(analysis2))## P(Reject H0_ITT) = 0.560000; P(Reject H0_subgp) = 0.000000; P(Reject H0) = 0.560000
## P(Early stop for efficacy [futility]) = 0.380000 [0.340000]
## Mean [SD] Randomized N = 336.080000 [106.155035]
##
## Stage at exit (proportion)
##
##
## exitStage proportion
## ---------- -----------
## 1 0.30
## 2 0.42
## 3 0.28
##
## Mean [SD] Lost N = 68.340000 [86.598442]
## Mean [SD] Analyzed N = 267.740000 [129.424757]
##
## Mean loss by futility stage and subgroup
##
##
## FutilityStage selectedGroup mean sd
## -------------- -------------- ------ ---------
## 1 1 100.8 8.011103
## 2 1 171.0 5.567764
## 3 1 242.4 5.856620
##
## Chance of each subpopulation rejected
##
##
## group count proportion
## ------ ------ -----------
## 2 28 0.56
##
## Counts by futility stage and subgroup choice
##
##
## FutilityStage G1
## -------------- ---
## 1 10
## 2 7
## 3 5
##
## CI Statistics:
## Overall coverage and coverage for rejections:
##
## overall rejection
## -------- ----------
## 1 1
##
## P(theta_test is in the confidence interval)
##
##
## coverage selectedCount rejectedCount
## --------- -------------- --------------
## 1 22 0
## 1 28 28
## NULLLai, Tze Leung, Philip W. Lavori, and Olivia Yueh-Wen Liao. 2014. “Adaptive Choice of Patient Subgroup for Comparing Two Treatments.” Contemporary Clinical Trials 39 (2): 191–200. doi:10.1016/j.cct.2014.09.001.
Lai, Tze Leung, Philip W. Lavori, and Ka Wai Tsang. 2015. “Adaptive Design of Confirmatory Trials: Advances and Challenges.” Contemporary Clinical Trials 45, Part A: 93–102. doi:10.1016/j.cct.2015.06.007.