Optimize return object preparation from gs_power_npe()

[jdblischak]

As discussed in group meeting, this PR optimizes the preparation of the return object by gs_power_npe().

Sorry @yihui. I had prepared this yesterday afternoon, but didn't have time to submit the PR. Could we review/merge this targeted PR first, and then rebase your more comprehensive PR on top?

xref: #623

[jdblischak]

Explanation of changes:

• tibble() -> data.frame() - no advantage to using tibble here, and we know from experience it is noticeably slower
• mutate() - unnecessary because info0 and info1 can be defined in the data.frame() call above
• filter() - removed this dead code. It was already commented out 4 years ago when the function was originally added in 18337df (upload #26)
• arrange() - removed. This sorting is unnecessary. By construction of the columns analysis and bound, the order will always be upper followed by lower, and the analyses in ascending order (1:n_analysis)

I updated the tests by running as.data.frame() on the output of the legacy implementation gs_power_npe_(). I felt this was the most transparent option. Another option would have been to ignore the attributes on the objects, but I felt this was more opaque (and more verbose), eg:

  local_edition(3)
  expect_equal(x1_c, x2, ignore_attr = TRUE)

And lastly, we could continue to return a tibble via as_tibble() as Yihui did in #623. Just let me know if you prefer one of these alternative solutions.

[jdblischak]

To benchmark the improvement, I put my changes in a separate gs_power_npe2() so that I could directly compare to the original. It reduced the median runtime from ~7 ms to ~4 ms. The most noticeable difference is in the worst case maximum time. This was reduced from ~41 ms to ~12 ms.

microbenchmark::microbenchmark(
  before = gs_power_npe(
    theta = c(.1, .2, .3),
    info = (1:3) * 40,
    upper = gs_b,
    upar = c(Inf, 3, 2),
    lower = gs_b,
    lpar = c(qnorm(.1), -Inf, -Inf)
  ),
  after = gs_power_npe2(
    theta = c(.1, .2, .3),
    info = (1:3) * 40,
    upper = gs_b,
    upar = c(Inf, 3, 2),
    lower = gs_b,
    lpar = c(qnorm(.1), -Inf, -Inf)
  ),
  check = "equivalent"
)
## Unit: milliseconds
##    expr    min      lq     mean  median     uq     max neval cld
##  before 6.5527 6.81630 7.635802 7.14665 7.4633 41.6138   100  a 
##   after 4.0639 4.16815 4.518994 4.27110 4.5039 12.0563   100   b

[yihui]

No worries. I'm all for using data.frame() here! Thanks!

I simply asked Claude to optimize gs_design_ahr() without specific instructions yesterday, and it naturally chose base R. I hope it didn't do so just to accommodate my personal taste.

[jdblischak]

it naturally chose base R. I hope it didn't do so just to accommodate my personal taste.

No, I doubt it. Base R is clearly the right choice in this context. The data is too small to justify using something like {data.table}. For a simple gathering of results into a table, data.frame() is the way to go.

[jdblischak]

@LittleBeannie After you approve, please squash and merge this PR. I don't think it justifies a NEWS bullet, but please let me know if you feel otherwise.

[LittleBeannie]

Thanks for the fix!