Fix loo_compare() for subsampling results

[florence-bockting]

Fixes #390

Description

loo_compare for subsampling results provided a difference based on reference model - comparison model, while loo_compare for other loo objects returned the reverse difference leading to a flip in the sign of elpd_diff.

This PR fixes the sign flip, adds a test for the comparison direction, and updates the related outputs of the tutorial vignettes/loo2-large-data.Rmd.

Tests

test_loo_subsampling.R: [ FAIL 0 | WARN 0 | SKIP 0 | PASS 408 ]

Documentation

Updated the three printed tables in loo2-large-data.Rmd.

TODO

• devtools::check()
• Fill the PR number into NEWS.md

AI assistance

I used AI assistance while working on this PR. I have reviewed all changes.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.86%. Comparing base (a59519c) to head (8635a88).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #391   +/-   ##
=======================================
  Coverage   92.86%   92.86%           
=======================================
  Files          31       31           
  Lines        3041     3041           
=======================================
  Hits         2824     2824           
  Misses        217      217           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 8635a88 is merged into master:

• ✔️loo_function: 1.51s -> 1.5s [-1.61%, +0.23%]
• ✔️loo_matrix: 1.5s -> 1.49s [-1.79%, +0.23%]
  Further explanation regarding interpretation and methodology can be found in the documentation.

[jgabry]

Looks good, thanks. I made one testing suggestion and I noticed one other small bug that doesn't have to be fixed here, but is one line so maybe we can just include it (see review comment)?

[jgabry]

Unrelated to this PR, but I think this line should be

names(x) <- rownames(comp)[order(ord)] 

that is, inside the brackets it should be order(ord) instead of just ord.

I noticed this while reviewing this PR and I thought about opening a separate PR, but then I decided to just suggest including it here.

[jgabry]

Not strictly necessary, but maybe we should also add a test like this, since I don't think we have anything like it? It tests both the sign and the magnitude at the same time:

# subsampling should agree with non-subsampling when nothing
# is actually subsampled (pins the direction *and* the magnitude)
l1ss <- loo:::as.psis_loo_ss.psis_loo(l1)
l2ss <- loo:::as.psis_loo_ss.psis_loo(l2)
l3ss <- loo:::as.psis_loo_ss.psis_loo(l3)
expect_equal(
  unname(loo_compare(l1ss, l2ss, l3ss)[, "elpd_diff"]),
  loo_compare(l1, l2, l3)$elpd_diff
)

The first argument to expect_equal() could then be simplified with the data.frame conversion in the other PR.