test(estimation): use an over-dispersed fixture for the NB fits

[arena-ai-coding-agent]

Why

After #66, the fits run. Run 36194648664 shows the remaining 19 failures and 6 errors. They all come from one fixture in test/unit/test_estimation.jl: two of its three features fail inside MASS::glm.nb (n_tested = 1, n_failed = 2, status partial).

The fixture is under-dispersed: its variance is below its mean in every group block (taxon 1, group A: mean 10.2, variance 3.1; taxon 2, group B: mean 30.3, variance 1.5). Under a negative binomial model, variance = μ + μ²/θ ≥ μ, so the maximum-likelihood θ for this data is infinite. A direct likelihood fit of the old table gives θ = 2e8 to 2e17 for every taxon under both offsets. theta.ml then stops at its iteration limit, and the estimator marks the fit failed. That is the right behaviour, and this PR does not change it.

What

• counts_effect is replaced by a table with the same written-in group means (10→40, 30→30, 50→5). Each block has a variance near μ + μ²/6, all counts are ≥ 1 (so RLE is defined), and the numbers are deterministic literals.
• A comment explains why the fixture has to be over-dispersed and records the outside-R check.
• CHANGELOG.md entry.

Evidence (outside R: there is no R/Julia in the authoring sandbox)

Two independent NB2 maximum-likelihood fits (a direct scipy likelihood and statsmodels NegativeBinomial) agree to 4 decimals, converge, and raise no warnings. The table uses the two offsets the tests use:

taxon	offset	slope	log2FC	θ	p
t_up	log lib size	1.4986	2.16	5.98	4e-14
t_flat	log lib size	0.1578	0.23	7.63	0.34
t_down	log lib size	−2.0470	−2.95	11.97	4e-29
t_up	RLE	1.5451	2.23	3.52	2e-10
t_flat	RLE	0.2378	0.34	14.89	0.07
t_down	RLE	−2.0510	−2.96	2.99	1e-13

Every assertion in the testsets holds against these numbers: estimate ≈ log 4 ± 0.5, log2FC ≈ 2 ± 0.8, |flat| < 0.6, down < 0, p(up) < 0.01. The parity with glm.nb itself is what this PR's CI run checks.

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Advanced

Run ID: f29a9adb-925d-4559-801b-002cd44c613d

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}