[ck37]

This branch carries Pratik Khadse's commit from #71 (id argument for cluster-robust standard errors) unchanged, with three follow-up commits on top. It is a draft so that CI runs on all three platforms; #71 failed on macOS.

What the follow-up commits change

• Formatting: restores the original indentation in the nine files Add cluster-robust standard errors #71 re-indented, so the diff against master is about 1,600 lines rather than 7,000, and restores RoxygenNote: 7.3.2.
• Clustered variance over all folds: one cluster-robust variance over every fold's influence curve, re-centered at the overall estimate, replaces the average of per-fold variances. This follows from the cross-validated TMLE's influence curve with clusters as the independent units (Zheng & van der Laan 2010; Hubbard, Kherad-Pajouh & van der Laan 2016). Folds holding a single cluster no longer make variables disappear, and clusters that span folds count once.
• Small-sample correction: jackknife-style (CR3) variance; equals sandwich::vcovCL(type = "HC3", cadjust = FALSE) for one-, two- and three-way clustering, and the usual G/(G-1) version when clusters are equal in size.
• Per-variable degrees of freedom: from the clusters holding each variable's low and high level, shown in a new df column of results_all when clustering. This count is a conservative heuristic motivated by Bell & McCaffrey (2002) and Imbens & Kolesár (2016), not a published rule.
• Fold assignment: every fold gets at least two clusters when there are enough, so a cluster holding about 1/V of the data no longer fills a fold by itself.
• Crossed clustering: when crossed variables cannot all stay within folds, the warning now names the variable whose clusters span folds and explains how to choose folds_by.
• Smaller fixes: macOS test failure (floating-point summation order), id = NULL warning only for data$column typos, factors whose levels each sit in one cluster are set aside with a message, SuperLearner's grouped folds stay stratified for g and g.Delta, print() notes clustering, README and ?varimpact examples and references, Pratik added to Authors@R.

Validation

• Full test suite passes with 100% line coverage; R CMD check passes locally (Linux); CI green on macOS, Windows and Ubuntu.
• Tests for continuous and missing outcomes with id, and for the variance against sandwich::vcovCL() and the cluster jackknife.
• An opt-in simulation study in tests/testthat/test-cluster-simulation.R (skipped unless VARIMPACT_RUN_SIMULATIONS=true):
  • Where every clustering variable stays within folds, null variables were flagged 0-9% of the time at the 5% level with id, against 12-73% without.
  • With site crossed with school and the default folds on school, a site-level null factor was flagged 13% of the time (36.5% without id): training and validation folds share sites, which biases it. With folds_by = "site" it was 3.0%.
  • The variance formula alone kept 92-95% interval coverage across crossed, one-way and dominant-cluster designs.
  • Reference results are at the end of the file.
• Without id, results are bit-identical to master.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UcpuhTnMKP8W19CLUUyU2T

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (dcc3a6d) to head (4be483d).
⚠️ Report is 5 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##            master       #72    +/-   ##
==========================================
  Coverage   100.00%   100.00%            
==========================================
  Files           31        32     +1     
  Lines         2301      2618   +317     
==========================================
+ Hits          2301      2618   +317     

☔ 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.