Warnings from fixed_design_rd() caused by mismatched column names

[jdblischak]

I noticed the following two warnings

devtools::test(filter = "fixed_design")
ℹ Testing gsDesign2
✔ | F W  S  OK | Context
✔ |   2      7 | independent-fixed_design [4.1s]

Warning (test-independent-fixed_design.R:130:3): RD
Unknown or uninitialised column: `theta`.
Backtrace:
    ▆
 1. └─gsDesign2::fixed_design_rd(...) at test-independent-fixed_design.R:130:3
 2.   └─gsDesign2::gs_design_rd(...) at gsDesign2/R/fixed_design_rd.R:96:5
 3.     ├─x$theta at gsDesign2/R/gs_design_rd.R:207:5
 4.     └─tibble:::`$.tbl_df`(x, theta) at gsDesign2/R/gs_design_rd.R:207:5

Warning (test-independent-fixed_design.R:130:3): RD
Unknown or uninitialised column: `info`.
Backtrace:
    ▆
 1. └─gsDesign2::fixed_design_rd(...) at test-independent-fixed_design.R:130:3
 2.   └─gsDesign2::gs_design_rd(...) at gsDesign2/R/fixed_design_rd.R:96:5
 3.     ├─x$info at gsDesign2/R/gs_design_rd.R:208:5
 4.     └─tibble:::`$.tbl_df`(x, info) at gsDesign2/R/gs_design_rd.R:208:5

══ Results 
Duration: 4.6 s

[ FAIL 0 | WARN 2 | SKIP 0 | PASS 7 ]

The two warnings are generated from a single call to gs_design_rd() that defines the argument power:

gsDesign2/tests/testthat/test-independent-fixed_design.R

Lines 130 to 133 in 50067c4

	x <- fixed_design_rd(
	alpha = 0.025, power = 0.9,
	p_c = .15, p_e = .1, rd0 = 0, ratio = ratio
	)

Interestingly, the subsequent call to gs_design_rd() that defines n instead of power does not generate the warnings:

gsDesign2/tests/testthat/test-independent-fixed_design.R

Lines 135 to 138 in 50067c4

	y <- fixed_design_rd(
	alpha = 0.025, n = x$analysis$n,
	p_c = .15, p_e = .1, rd0 = 0, ratio = ratio
	)

When power is defined, fixed_design_rd() calls gs_design_rd():

gsDesign2/R/fixed_design_rd.R

Lines 85 to 104 in 50067c4

	if (is.null(power)) {
	d <- gs_power_rd(
	p_c = tibble::tibble(stratum = "All", rate = p_c),
	p_e = tibble::tibble(stratum = "All", rate = p_e),
	ratio = ratio,
	upper = gs_b, upar = qnorm(1 - alpha),
	lower = gs_b, lpar = -Inf,
	n = tibble::tibble(stratum = "All", n = n, analysis = 1),
	rd0 = rd0, weight = "unstratified"
	)
	} else {
	d <- gs_design_rd(
	p_c = tibble::tibble(stratum = "All", rate = p_c),
	p_e = tibble::tibble(stratum = "All", rate = p_e),
	alpha = alpha, beta = 1 - power, ratio = ratio,
	upper = gs_b, upar = qnorm(1 - alpha),
	lower = gs_b, lpar = -Inf,
	rd0 = rd0, weight = "unstratified"
	)
	}

gs_design_rd() then calls gs_info_rd(), and expects the returned table to contain the columns theta and info:

gsDesign2/R/gs_design_rd.R

Lines 206 to 212 in 50067c4

	if (h1_spending) {
	theta1 <- x$theta
	info1 <- x$info
	} else {
	theta1 <- 0
	info1 <- x$info0
	}

However, the result returned by gs_info_rd() does not contains these columns:

gsDesign2/R/gs_info_rd.R

Line 288 in 50067c4

select(analysis, n, rd, rd0, theta1, theta0, info1, info0)

Surprisingly the tests continue to pass even though theta1 and info1 are assigned NULL. I thought that maybe the autocomplete feature of $ still returned the correct values, but it can't distinguish between eg theta1 and theta0

x <- data.frame(theta1 = 1, theta0 = 0, info1 = 1, info0 = 0, long_name = 3)
x$theta
## NULL
x$info
## NULL
x$long
## [1] 3

[jdblischak]

I used git bisect to identify the first commit that generated the warnings. It was commit 3c71671 from @LittleBeannie, in which the default value of h1_spending was changed from FALSE to TRUE.

Thus I think my proposed solution in #293 to set h1_spending = FALSE when calling gs_design_rd() would restore the previous behavior.

[nanxstats]

Keaven suggested that Yujie should weigh in, and using both h0 and h1 might be the best.

So let's revisit this after the holidays.

[LittleBeannie]

I used git bisect to identify the first commit that generated the warnings. It was commit 3c71671 from @LittleBeannie, in which the default value of h1_spending was changed from FALSE to TRUE.

Thus I think my proposed solution in #293 to set h1_spending = FALSE when calling gs_design_rd() would restore the previous behavior.

Hi @jdblischak , thanks for tracing back! Since gsDesign2 uses h1_spending = TRUE consistently for consistency with most of the clinical trial practice, I would suggest not to change it back to FALSE.

[LittleBeannie]

This issue is fixed in the PR of #303. Please take a look and review, thanks!