feat(conformal): cross-sectional conformal inference (vartype = "conformal")

[xuyiqing]

Summary

Adds vartype = "conformal": a cross-sectional conformal prediction interval for the average treatment effect on the treated. It ranks the treated unit's average post-treatment prediction error against the leave-one-control-out prediction errors of the donor pool, giving a distribution-free interval with no bootstrap draws. The interval is closed-form and never empty.

This implements the sc-conformal paper's Algorithm 1 as an estimator-agnostic fect feature.

Design: Family A (level statistics)

Every interval is a level statistic S_i(τ) = |m_i − τ| / scale_i, inverted in closed form to m_tr ± scale_tr · Q, where Q is the conformal (1−α) quantile of the control scores. Never empty; unbounded only below the resolution floor α < 1/(Nco+1). (An earlier path/constant-effect score family was removed because it could return an empty acceptance set, which is confusing to report.)

Options (orthogonal knobs)

arg	values	default
conformal.scale	sd · none · rmspe · mad · diff (model-se reserved)	sd
conformal.center	mean · median	mean
conformal.weight	cell · unit · precision	cell
conformal.band	pointwise · simultaneous	pointwise
conformal.cutoff	per-period · pooled	per-period

The simultaneous (sup-t uniform) band is also always stored in fit$est.att.sim.

Default chosen by simulation

scale = "sd" (the studentized-mean) was selected by a coverage study (tests/coverage-study/conformal_default_study.R, 150 reps, true effect 0): it has the highest worst-case scalar coverage (0.920) and the smallest mean width (3.16) across iid / AR(1) / heteroskedastic / nonstationary DGPs. It is decisively tighter than none under heteroskedasticity (width 2.82 vs 5.43 at equal coverage). This default is provisional and easy to change — the maintainer's call.

Scope and guards

• Requires a separated (controls-only) fit and method not in c("mc", "both").
• The bootstrap F / equivalence diagnostic (diagtest) is skipped for conformal (no draws), as it is for se = FALSE.
• Group / reversal / weighted (W) / balanced / placebo / carryover options error clearly and direct the user to bootstrap/jackknife.
• Existing vartypes (bootstrap/jackknife/parametric) are untouched — all changes are gated to the conformal branch / conformal.R.

Known open items (documented, not blocking)

• Multi-treated calibration ranks an Ntr-aggregate against single controls; the simultaneous band undercovers nominal in small N (~0.77 vs 0.90), and cell/unit weights undercover under treated-unit heterogeneity (precision recovers it). A placebo-average calibration is the future refinement.
• conformal.scale = "model-se" is reserved (needs estimator prediction-SE plumbing); it stops with a clear message.

Verification

• tests/testthat/test-conformal.R: 55 tests pass (core engine, closed-form interval, all scales/centers/weights, pointwise + simultaneous bands, multi-treated weights, end-to-end fect() integration, coverage guards).
• print() and plot(type = "gap"|"counterfactual") work on conformal fits.
• Coverage study results: tests/coverage-study/results/conformal_default_study.md.
• Full testthat suite (all files, serial): 0 failures, 0 warnings, 0 skips.
• Inferential coverage study (run_para_error_coverage.R): T19/T20/T21 all PASS (the parametric variance path is unaffected; conformal changes are additive and gated).

Update — additional buildout

• Staggered adoption is now supported (per-cohort donor pools, union-window calibration, event-time-aggregated bands). Block behavior is preserved exactly. Coverage on a two-cohort staggered DGP (100 reps, effect 0, α=0.10): scalar 0.90 (iid) / 0.94 (ar1); per-period band 0.91 / 0.91; never empty/unbounded.
• Plotting: plot() renders the event-study and counterfactual for block and staggered conformal fits. The simultaneous band is plottable via conformal.band = "simultaneous" (it fills est.att) or the est.att.sim slot.
• True inner band: est.att90 is now a genuine 1 − 2α conformal band (was a placeholder mirroring the main band), computed at no extra fitting cost.
• Print: an Inference: conformal (scale, band, N.calib) line.
• Joint-band coverage characterized (300 reps): the simultaneous band's joint coverage is ~0.84–0.87 (mildly anti-conservative, in the jackknife+ range), far better than the pointwise joint ~0.66. Full conformal on the max statistic would close the gap and is left as a documented follow-up; the scalar interval is already exact.

Design + session record: statsclaw-workspace/fect/runs/2026-06-08-conformal-inference.md and runs/REQ-conformal/spec.md.

🤖 Generated with Claude Code

[xuyiqing]

Overnight additions (2 commits, 9b79579 + 00483a3):

• Fix: vartype="conformal" + covariates crashed in output labelling (no est.beta under conformal); null-guarded, staggered+covariate regression test added.
• conformal.fit made user-facing and two-sided (it was unreachable from fect()): custom separated learners now score both the held-out donors and each treated unit (own pre-window). A quadprog simplex-SC learner reproduces the hand-rolled Proposition 99 conformal interval exactly (ATT -19.5136, CI [-53.6576, 14.6304], p .0769).
• time.component.from defaults to NULL = auto: resolves to "nevertreated" under conformal + fe/ife/cfe (strict separation, matches the paper's Algorithm 1; message emitted); explicit "notyettreated" warns.

Merge gate still open: the full testthat suite re-run was interrupted twice by session ends (zero failures at both kill points; the conformal file is fully green standalone). Run NOT_CRAN=true TESTTHAT_CPUS=10 Rscript -e 'library(fect); testthat::test_dir("tests/testthat", reporter="summary")' to completion before merging.