Status complete; merged into E28: the table of record is the merged one (results/E26merged/), E28's rows for the 29 families inside the calibrated range and E26's for the 18 outside it, and E28's page carries it. E48 refitted the nd rung of the 18 under emphasis 4e0b9dd: on the 17 read here the parsimonious choice changes on 1 of 9 readable families (Procellariidae, cr to nd), above the 10% threshold, so this set's choice table is E48's, and on the merged table on 2 of 33 usable, so that table stands with the nd row updated; the observed-tree paths are rerun within E47 (Estimability)
Package build 3e596dd (the lineage-table cap)
Driver / analysis sim/E25-real.R --box auto, sim/E26-analysis.R, figures sim/figures-families.R
Data results/E26/families.csv (47 rows), fits.csv, paths.csv; merged table results/E26merged/families.csv (47 rows); report/e26-numbers.tex, report/e26-table.tex

Question

E25 ran the ladder and the path on three bird families. The biology paper's closing analysis needs every family-level tree the study's size range covers: per family the chosen model, its margin and the coefficients, and across families the distribution of signs and effects.

Design

Trees. Every Condamine et al. (2019) family tree in the phylo-species-atlas with 50 to 200 tips and \(\rho \ge 0.5\) (sim/data/atlas-candidates.csv): 47 families, 33 bird, 10 mammal, 3 squamate (Chamaeleonidae, Phrynosomatidae, Varanidae), 1 amphibian (Eleutherodactylidae).

quantity range median
tips 50-194 88
crown age (Ma) 11.2-101.4 34.7
\(\rho\) 0.651-1.000 0.918

Tip edges extended to one present where the depths spread by at most 15% of the crown age (E5's rule); a tree beyond it "is reported and not fitted".

Fits.

setting value
ladder estimate_rates(model = "all"), exponential link, each family's \(\rho\): cr → dd → {nd, ned, nk}
box auto_bounds()
draws 300
iterations 100
clock 30 min per fit (E23's setting at 100 tips)
path covariate_path() over \(N\), \(D\), ED, \(K\) on the observed tree and on 3000 surrogate draws at the ladder's dd estimate

Definitions (sim/E26-analysis.R): the winner is the lowest AIC; a family's verdict is "lineage-level" when the winner is nd, ned or nk and leads dd by more than 2, "diversity- dependent" when dd wins by more than 2, "constant rate" likewise for cr, otherwise "undecided"; a family is usable when its winner's fit has an effective sample of at least 10. Coefficients are given as \(\exp(\hat\beta_X\cdot\mathrm{sd}_X)\), the speciation-rate factor per standard deviation of the covariate over the family's tips.

Pre-set rule

None: a comparative description read against the floors of E21-E24. A family whose chosen fit has an effective sample below 10 is reported and left out of every across-family statement. The design of E27, registered after E26 ran, adds that the count of supported families is restated beside the false-positive rate of the nearest E27 cell, and that no family is added or removed on E27's account.

Results

Two tables carry E26. The run itself (README row, results/E26/families.csv) and the merged table the biology paper reads (results/E26merged/: E28's rows for the 29 families at \(\rho \ge 0.9\), E26's for the 18 below, marked †).

quantity E26 run (47) merged table (47)
usable (winner ESS ≥ 10) 37 37
dd leads cr by > 2 32 33
lineage-level model supported, of the usable 18 (nk 11, ned 7) 21 (nk 12, ned 9)
dd / cr / undecided, of the usable not recorded 11 / 0 / 5
\(\beta_K > 0\), nk fits with ESS ≥ 10 31 of 33 30 of 31 (median factor 1.40 per sd, 0.78-2.74)
\(\beta_{ED} < 0\), ned fits with ESS ≥ 10 32 of 33 30 of 32 (median factor 0.68, 0.33-1.08)
\(\beta_D < 0\), nd fits with ESS ≥ 10 not recorded 19 of 25 (median factor 0.86)
lineage-level fits meeting the stopping rule 19 of 141 24 of 141

Further readings from the merged table (report/e26-numbers.tex):

E26 verdict
Figure 1. Per family, the AIC lead of dd over cr (left) and of the best lineage-level model over dd (right), merged table with the nd rung refitted in E48; dark: lead above 2, open: fewer than 10 effective draws. 20 families have a supported lineage-level model: K in 12, ED in 6 and D in 2 (Meliphagidae and Procellariidae).
E26 shape
Figure 2. The K factor per standard deviation against the tree's Colless imbalance. The four most imbalanced trees (shaded band) all have K factors near 2 or above, among them the two largest (Leporidae 2.74, Pachycephalidae 2.69); below the band the rank correlation drops from 0.54 to 0.32. (The subtitle's "NA" is a missing label in the figure script, not a value.)

Conclusion

No rule. For the 29 families inside the calibrated range the parsimonious choice is E28's: cr 4, dd 5, nd 0, ned 8, nk 8 over 25 usable families. Beside it, as a sensitivity reading, the merged table's minimum-AIC winner makes a lineage-level model the supported choice in 21 of the 37 usable families, K in 12 and ED in 9; the record does not compute the parsimonious split over all 47. The parsimonious choice comes first because it is the rule registered before the data, and MCEM log-likelihoods carry Monte Carlo error and a downward bias, so AIC differences below 2 are within the estimator's noise. On the merged table K raises speciation in 30 of 31 readable nk fits and ED lowers it in 30 of 32 ned fits. What it does not show: the coefficients come from fits of which 19 (E26) or 24 (merged) of 141 met the stopping rule, and E27's cells put the fitted K coefficient at 0.65-0.76 of its generating value on fits stopped at 100 iterations, so the factors are read as attenuated. The extinction side is not read (41 of 47 at fitted turnover below 0.05). The signs and counts describe what the ladder chooses on these trees; E39 later found that on unbalanced real trees K and ED share much of their ranking (ED-K median -0.44), which bears on how far a K choice and an ED choice are distinct statements.

Recorded inconsistencies: the README's E26 row still carries the run's own counts (18, nk 11, ned 7), while e26-numbers.tex and the biology paper read the merged ones (21, 12, 9). Macropodidae (tip spread 19%) was fitted and appears in the merged table (dd, lead over cr 32.6), although the design says a tree beyond 15% is reported and not fitted; read_family in sim/E26-analysis.R has no refusal. report/report.tex calls the set "47 bird families"; the candidate file has 33 bird, 10 mammal, 3 squamate and 1 amphibian.

The nd rung and the paths on the observed trees maximised and read the earlier D likelihood and lineage table. Emphasis 4e0b9dd changed the D likelihood; before it the likelihood held the clade mean pendant age at its segment-end value, so \(D\) grew within a segment while the simulator held it constant; from it \(D\) is constant between events in both. On complete trees E44 measured d's \(\beta_D\) at 0.70 and 1.46 of its value under the earlier likelihood and 0.95 and 1.16 under the current one (nd at -0.35: 0.77 and 1.01). E48 refitted the nd rung of these families under the current likelihood: nd becomes the parsimonious choice on Procellariidae (readable) and on Dasyuridae and Strigidae (sampler-limited), the readable count of the 17 falls from 11 to 9, and over the 7 read fits with \(\beta_D\) off its edge the D factor is below 1 on all 7 (median 0.66, against 4 of 5 and 0.80 before); the paths are rerun within E47.