Status complete for cr and dd, superseded by E45 (conditioned, against the exact conditioned MLE, 25-800 tips) when it reports; the ned arm stopped (3 usable fits of 23, no merged table) and is historical: it led to E7
Package build 64edf63 and a87de1c (the two builds in the 2026-10-02 build audit, which found them identical to a clean build except for the ned thinning random stream; S's cr and dd arms do not use that proposal). Package commit 9fdcdbe (2026-10-01 23:51 CEST) says builds from a clean tree were not affected, while the build audit committed two hours later finds 64edf63 and 4f42b27 differing from a clean build for ned on thinning and a87de1c on both ned proposals; the record does not reconcile the two
Driver / analysis sim/S-samplesize.R, sim/S-analyse.R; figures sim/figures-study.R
Data results/S/cr/fits.csv (360 rows), results/S/dd/fits.csv (125 rows), summary manuscript/tables/S-summary.csv; results/S/ned/jobs/ (23 files)

Question

E1 checks MCEM against the exact MLE only up to 100 tips. S holds the generating parameters fixed and changes only the amount of data, 25 to 800 tips, so that the curve is a statement about sample size: how close MCEM gets to the exact estimate of the same tree, how close both get to the generating value, and at what size each reaches a usable accuracy. Its ned arm, which failed at 200 tips and above, is what led to the feasibility map E7; its cr fits ending on the clock at 800 tips led to E9.

Design

parameter cr dd meaning
tips 25, 50, 100, 200, 400, 800 25, 50, 100, 200, 400 target size
trees per size 30 25
tree draw exactly n tips, TreeSim::sim.bd.taxa crown age calibrated per size (6.53, 6.53, 7.97, 8.38, 9.99) so the median lands on the target (realised medians 31, 51, 109, 191, 414)
rates \(\lambda\) = 0.5, \(\mu\) = 0.2 (turnover 0.4) \(\lambda\)0 = 0.8, \(\mu\)0 = 0.2, \(\gamma\)N = 0, K = 2n so \(\beta\)N = -0.3/n K tied to n so every tree sits at the same half of its capacity
proposal surrogate surrogate the package's recommended setting
draws grown by the mc_error rule from 200 to a cap of 1600 ("adaptive"); a second arm at 200 fixed grown, 200 to 1600
stopping rule mc_error mc_error
exact reference DDD::bd_ML per tree DDD::dd_ML per tree (three parameters, \(\mu\) constant); one 50-tip reference returned the failure code and is dropped
link, sampling linear, \(\rho\) = 1 linear, \(\rho\) = 1

The two dd estimators do not maximise the same function (dd_ML holds \(\mu\) constant, emphasis fits \(\gamma\)N as well), so each is compared with the generating value rather than with the other.

Pre-set rule

S has no design entry in 00-design.md and no rule fixed before its data; its design is the driver header. The supplement and report read it against a working box: within 10% bias and 25% RMSE of the generating value, with bias and RMSE in % of the generating speciation rate. That box is used below, as a reading and not as a registered rule.

Results

cr, MCEM against the exact estimate. The central 80% range of (MCEM - exact) is 0.08-0.10 of the range of (exact - truth) for speciation at every size, and 0.26, 0.10, 0.10, 0.12, 0.10, 0.10 for extinction at 25-800 tips. MCEM sits slightly above the exact estimate: median +0.6, +0.9, +3.5, +3.2, +3.7, +3.2% of the speciation rate on speciation, +1.2, +1.5, +5.8, +5.4, +6.0, +5.1% on extinction.

MCEM minus the exact MLE against the exact MLE minus the truth, constant rate, 25 to 800 tips
Figure 1. Grey bars: the central 80% of the exact MLE's own error at each size; dark bars: the central 80% of MCEM minus the exact MLE on the same trees. The dark bar is a tenth or less of the grey one from 50 tips on, and offset to the right by 1-6% of the speciation rate.

cr, against the generating value (draws grown; % of the generating speciation rate):

tips speciation bias / RMSE MCEM exact extinction bias / RMSE MCEM exact fits on the clock median s per fit
25 -10.4 / 25.0 -12.3 / 25.4 -27.8 / 33.1 -31.2 / 35.0 0 of 30 225
50 -10.4 / 21.9 -11.5 / 23.4 -18.7 / 33.1 -20.8 / 35.4 2 500
100 -1.8 / 13.7 -5.0 / 15.1 -3.4 / 19.9 -8.8 / 22.5 22 1875
200 -1.5 / 9.8 -4.9 / 11.0 -5.8 / 16.2 -11.6 / 19.9 24 3739
400 +2.9 / 11.3 -0.6 / 11.2 +1.6 / 12.6 -4.1 / 13.9 29 5672
800 +1.7 / 7.0 -1.7 / 7.2 +2.3 / 10.9 -3.1 / 11.7 30 5681

The cells from 100 tips on are clock-limited (the clock ended 73-100% of fits). At 400 and 800 tips the draws reached 1013 and 675, not the cap of 1600. The exact estimate returns extinction at zero on 63% and 57% of trees at 25 and 50 tips, MCEM on 13% and 3%. With 200 draws held fixed the estimates drift upward from 100 tips: speciation bias +3.3, +4.3, +9.3, +8.4% and extinction +4.4, +3.3, +11.6, +12.6% at 100-800 tips, at 255-2371 s a fit.

dd (draws grown; median of fitted over generating, MCEM / exact):

tips intercept slope extinction intercept bias / RMSE, MCEM vs exact stop: converged / max_iter / clock
25 1.11 / 0.63 1.37 / 0.27 0.00 / 0.00 +27.7 / 48.8 vs -33.1 / 36.5 11 / 14 / 0
50 1.11 / 0.72 1.11 / 0.45 0.00 / 0.00 +15.8 / 25.5 vs -28.5 / 57.5 7 / 17 / 0
100 0.97 / 0.74 0.99 / 0.68 0.00 / 0.00 -0.5 / 11.7 vs -11.9 / 41.7 3 / 22 / 0
200 1.13 / 0.79 1.22 / 0.53 0.32 / 0.00 +13.9 / 17.2 vs -6.2 / 34.2 0 / 23 / 2
400 1.23 / 0.74 1.58 / 0.44 1.67 / 0.03 +22.4 / 23.7 vs -17.5 / 26.9 2 / 18 / 5 (clock-limited)

MCEM returns extinction at exactly zero on 92, 76, 48, 12 and 0% of trees at 25-400 tips; the exact dd_ML on nearly every tree at every size. Extinction bias / RMSE is -24.9 / 24.9 (MCEM) against -24.0 / 24.2 (exact) at 25 tips, -4.9 / 20.7 against -12.8 / 23.2 at 200, +15.5 / 19.7 against -16.1 / 20.6 at 400. Most dd fits ended on the 100-iteration cap, so the 200- and 400-tip numbers are those of unfinished fits.

dd: relative error of the speciation intercept and of the diversity dependence, MCEM and exact MLE, by tree size
Figure 2. Each large dot is the median at one tree size. MCEM (dark, four parameters) errs high on both the intercept and the diversity dependence, the exact three-parameter MLE (grey) errs low; MCEM's median passes through the generating value at 100 tips.

ned arm. 23 job files at 100, 200 and 400 tips: 20 failed in their first E-step (e_step_failure, 0 iterations) and 3 ended on their clock after 31, 39 and 1 iterations, with \(\hat{\beta}\)_ED at 0.80, 0.74 and 0.73 of the generating value on effective samples of 10.6, 10.5 and 19.2. Not a recovery measurement.

Conclusion

Read against the 10% / 25% box (not a registered rule): MCEM arrives at 100 tips for both cr rates and for the dd intercept and slope, and at 200 tips for the dd extinction rate; the exact dd_ML does not reach the box on the intercept at any size tested. MCEM is not worse than the exact estimate on any cell and has the smaller bias on most. Under cr MCEM lands on the exact estimate to a tenth of that estimate's own spread, offset upward by 1-6% of the speciation rate. What S does not show: the cr cells from 100 tips and the dd cell at 400 are clock-limited and the dd fits from 50 tips mostly ended on the iteration cap, so those are readings of unfinished fits; the dd intercept leaves the box again at 200 and 400 tips (+13.9 and +22.4% bias) on such fits. The fixed-draw arm shows the draw count must grow with the tree. S covers turnover 0.4 (cr) and 0.25 (dd) at complete sampling only.