RXR-3164: Pass observation-level weights to covariance matrix calcs, and re-activate FOCE-type vcov

[roninsightrx]

There were two issues with the covariance matrix computation:

• the computational approach was changed from finite differences using numDeriv, to FOCE-style analytic. While both are "correct", the latter one is guaranteed positive-definite, and faster. Slightly different definition, but now same definition as used in NONMEM. So we'll use that going forward.
• the new computation however didn't use the weights that were used in the fit, so wasn't computing the right cov matrix. These are applied now.

Note: PR to update the handful of failing f-tests. After merging this PR, immediately rerun f-tests there.

[jasmineirx]

lgtm, a couple suggestions for better tests

[Copilot]

Pull request overview

This PR fixes variance–covariance (vcov) computation for MAP estimation by ensuring per-observation weights are incorporated into the FOCE-style Hessian/vcov, and by re-preferencing the FOCE-derived vcov (aligned with NONMEM’s FOCE-I behavior) for speed and positive-definiteness.

Changes:

• Thread per-observation weights into FOCE Hessian/vcov computation (calc_cwres() via calc_residuals()), and prefer FOCE vcov in get_map_estimates().
• Correct numDeriv-based vcov scaling in mle_wrapper() for an objective on the -2*logLik scale.
• Add/adjust tests to validate FOCE vcov vs NONMEM reference values and to ensure weights affect vcov curvature.

Reviewed changes

Copilot reviewed 7 out of 10 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
tests/testthat/test-nonmem-vcov-comparison.R	Adds a hard-coded NONMEM cross-check for FOCE vcov under unit vs weighted observations.
tests/testthat/test-get_map_estimates.R	Updates MAP expectations and adds a FOCE vs numDeriv vcov comparison.
tests/testthat/test-calc_cwres.R	Adds a focused regression test that weights influence FOCE vcov but not CWRES.
R/mle_wrapper.R	Adjusts vcov scaling to match a -2*logLik objective.
R/get_map_estimates.R	Skips numDeriv Hessian when residuals/FOCE Jacobian is available; prefers FOCE vcov source.
R/calc_residuals.R	Forwards weights into calc_cwres() so FOCE Hessian curvature matches the weighted OFV.
R/calc_cwres.R	Applies per-observation weights in the FOCE Hessian likelihood term.
man/parse_residuals_from_predictions.Rd	Updates weights parameter documentation.
man/calc_residuals.Rd	Updates weights parameter documentation.
man/calc_cwres.Rd	Documents new weights argument in calc_cwres().

Files not reviewed (3)

• man/calc_cwres.Rd: Generated file
• man/calc_residuals.Rd: Generated file
• man/parse_residuals_from_predictions.Rd: Generated file

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jasmineirx]

lgtm. thanks for adding the NONMEM equivalence test