E44 — on complete trees, is the likelihood the simulator's, for every model?
| Status | complete; the raw-ED failure diagnosed as estimability and selection, not a mismatch (see Conclusion) |
| Package build | before: emphasis d3dee92; after: 4e0b9dd (the likelihood holds D constant between events) |
| Driver / analysis | sim/E44-complete.R, sim/E44-analysis.R |
| Data | results/E44/before/E44.csv, results/E44/after/E44.csv, results/E44/analysis.txt, results/E44/design.rds (the 720 simulated L-tables) |
Question
Every estimation experiment so far mixed two things: whether the complete-data likelihood is that of the process the simulator runs, and how well Monte Carlo EM integrates over the hidden part. On a complete tree, every lineage kept, extinct ones included, nothing is hidden, so maximising the complete-data likelihood directly tests the first alone. Step 0.2 of the estimability plan. The D model was known to be defined two ways: the simulator held \(D\) constant between events, the likelihood held the clade mean \(M\) at its end-of-segment value so \(D\) grew within a segment (Open issues).
Design
| models | cr, dd, d, nd, ed, edc, ned, nedc, k, nk, exponential link |
| trees | complete (extinct lineages kept), \(\rho = 1\), crown age 15, kept if both crown lineages survive and 60-160 extant tips |
| turnover | 0.10 at the crown |
| effect | the model's lineage-level coefficient at \(-0.35\) and \(+0.35\) per scale unit; single-covariate models in E24's unit (log(1 + effect) / sd of the covariate on cr clades, intercept calibrated), \(N\) models in E27's (\(\lambda_0 = 0.5\), \(\beta_N = \log 0.1 / K\), \(K\) calibrated) |
| trees per cell | 40, the same trees before and after |
| fit | the complete-data log-likelihood maximised by Subplex from the generating values and from a perturbed start, each restarted until a restart gains less than \(10^{-6}\); the two starts reached the same maximum on every tree |
Pre-set rule
For every model and cell: the median ratio of each speciation coefficient to its generating value lies in [0.9, 1.1], and each slope has the right sign on at least 95% of trees. A model that fails has a likelihood that is not its simulator's and is corrected before the estimability study.
Results
Median ratio of the lineage-level speciation coefficient, and the share of trees with its right sign:
| model, cell | before | after | rule before / after |
|---|---|---|---|
cr, dd |
- | - | pass / pass |
d -0.35 |
0.70, 1.00 | 0.95, 1.00 | fail / pass |
d +0.35 |
1.46, 1.00 | 1.16, 1.00 | fail / fail |
nd -0.35 |
0.77, 0.975 | 1.01, 0.975 | fail / pass |
nd +0.35 |
1.20, 1.00 | 0.80, 0.95 | fail / fail |
ed -0.35 |
0.81 (\(\beta_0\) 1.36) | same | fail / fail |
ed +0.35 |
1.22 | same | fail / fail |
edc -0.35 / +0.35 |
1.04 / 1.07 | same | pass / pass |
ned -0.35 / +0.35 |
0.82, 0.90 / 1.00, 0.875 | same | fail / fail |
nedc -0.35 / +0.35 |
1.03 / 1.01 | same | pass / pass |
k -0.35 / +0.35 |
1.005 (\(\beta_0\) 0.23) / 0.97 | same | fail / pass |
nk -0.35 / +0.35 |
1.14, \(\beta_N\) 1.93, sign 0.60 / 1.00 | same | fail / pass |
Bootstrap 95% intervals of the medians (a reading beside the rule, not a change to it): before the D change the d medians exclude 1 (0.70 [0.58, 0.78]; 1.46 [1.24, 1.65]) and so does nd -0.35 (0.77 [0.66, 0.91]); after it all four d/nd intervals contain 1 (d [0.86, 1.05] and [0.94, 1.40]; nd [0.91, 1.17] and [0.63, 1.25]). Raw ed excludes 1 in both cells (0.81 [0.74, 0.93]; 1.22 [1.02, 1.32]), both consistent with an additive +0.04 in \(\beta_{ED}\). The two D conventions differ by 0.40-1.20 nats per complete tree.
Conclusion
- The D change was needed and brings
dandndto their generating values (emphasis4e0b9dd); thedandnd+0.35 cells still miss the [0.9, 1.1] band at 40 trees, with intervals that contain 1. Every D experiment before4e0b9dd(E19, E24'sdarm, E30, E31, E37, \(D\) on the path in E20, E22, E33, and E40'sndfits) maximised the old convention's likelihood; the persistent \(\beta_D\) shortfall of those experiments is consistent with it. - Raw ED: no mismatch, a ridge and a window. An independent R computation of the ED likelihood from the definition equals
eval_logfto \(3 \cdot 10^{-13}\), and an independent R simulator of the raw-ED process agrees withsimulate_tree()(KS p 0.14-1 on 17 of 18 comparisons over 400 clades a cell). Pooled over 261 and 348 trees conditioned on survival only, the MLE returns \(\beta_{ED}\) at 1.005 and 1.002 of its value. What fails is the per-tree estimate: raw ED grows with time like the clade mean, so \(\beta_0\) and \(\beta_{ED}\) lie on a ridge and a single 100-tip tree places them along it (\(\beta_{ED}\) 0.93 and 1.17, \(\beta_0\) 1.20 and 1.07, with no selection on size). In the -0.35 cell E44's 60-160 tip window adds a selection bias: the generator's survivors have a median of 145 tips, so the window keeps the smaller quarter (pooled MLE on those trees \(\beta_0\) 1.25, \(\beta_{ED}\) 0.89). Centring removes the ridge, which is whyedcandnedcpass. A regression test now holds the ED likelihood to the definition (emphasise1c2e82). The diagnosis is inresults/E44/ed-diagnosis/. nedfails on sign share at a small effect in E27's unit (\(\beta\) -0.095 / +0.066): a power limit, its medians' intervals contain 1.k-0.35 fails on the intercept ratio because the generating intercept is -0.12; \(\beta_K\) recovers at 1.005.nk-0.35 fails because its generator stops at a median 73 tips, where \(\beta_N \approx -0.001\) is not identified.