WLR for stratified design

[LittleBeannie]

The following codes give an error message.

library(gsDesign2)
library(dplyr)

enroll_rate <- tibble(stratum = rep(c("B+", "B-"), 3), 
                      duration = c(3, 3, 3, 3, 12, 12),
                      rate = c(3, 2, 6, 4, 12, 8))

fail_rate <- tibble(stratum = c(rep("B+", 2), rep("B-", 2)),
                    duration = c(4, 100, 6, 100),
                    fail_rate = log(2) / 18,
                    hr = c(1, .5, 1.2, .7),
                    dropout_rate = -log(.99) / 12)

gs_design_wlr(
  enroll_rate = enroll_rate,
  fail_rate = fail_rate,
  alpha = 0.02,
  beta = 0.15,
  info_frac = NULL,
  analysis_time = c(16, 24, 36),
  ratio = 1,
  binding = FALSE,
  upper = gs_spending_bound,
  upar = list(sf = gsDesign::sfLDOF, total_spend = 0.02, param = NULL, timing = NULL),
  lower = gs_b,
  lpar = c(qnorm(0.1), -Inf, -Inf))

Besides, the stratified max combo is not workable... considering max combo is the max of a series of WLR.

[elong0527]

Would you feel we need to implement stratified design in next half?

[LittleBeannie]

Would you feel we need to implement stratified design in next half?

I discussed it with Keaven yesterday, and he suggested it is nice to have but not of high priority.

[elong0527]

I saw the way we combine stratum in ahr is to take sum of the information.

https://github.com/Merck/gsDesign2/blob/main/R/ahr.R#L120

I will implement the approach, could you confirm with @keaven if this is what we want in WLR?

[LittleBeannie]

I saw the way we combine stratum in ahr is to take sum of the information.

https://github.com/Merck/gsDesign2/blob/main/R/ahr.R#L120

I will implement the approach, could you confirm with @keaven if this is what we want in WLR?

Hi @elong0527 , thanks for checking! Based on the conversation I had with Keaven yesterday, the answer is yes, but there is more to consider regarding how we combine different strata. Devan's 5-star approach suggests that we could use weights based on information or sample size, or a combination of both.

In the initial version of the stratified gs_design_wlr/gs_power_wlr, let's use information-based weights, which is similar to line 120 of the ahr.R file. However, in the future, we can explore the possibility of offering flexibility in choosing different weighting strategies.