Phylogenetic Inference and Congruence

Felsenstein, Joseph.  “Confidence limits on phylogenies: an approach using the bootstrap.”  Evolution 39(4):783–791.  1985.  A methods paper that brought the bootstrap into phylogenetics: the characters are resampled with replacement, the tree is rebuilt on each sample, and a group that appears in 95% or more of the samples counts as statistically significant.  The paper says plainly that the bootstrap gives an interval for the tree that repeated sampling of characters would estimate, not the true tree.  The article uses it to say what a bootstrap value measures. ↩

1. 1a.

   Page 786, right column:  “Bootstrapping provides us with a confidence interval within which is contained not the true phylogeny, but the phylogeny that would be estimated on repeated sampling of many characters from the underlying pool of characters.  As such it may be misleading if the method used to infer phylogenies is inconsistent.” ↩ ↩ ↩

Puigbò, Pere et al.  “Search for a 'Tree of Life' in the thicket of the phylogenetic forest.”  Journal of Biology 8:59.  2009.  A research article that examined a 'forest' of 6,901 gene trees from prokaryotes (bacteria and archaea) to see whether a central pattern survives horizontal gene transfer, the movement of genes between unrelated lineages.  The authors found that horizontal transfer is widespread, so very few gene trees agree fully, but 102 nearly universal trees still carry a consistent signal.  The article uses it to show both halves of the picture: transfer wrecks a single tree for prokaryotes, and a residual signal of inheritance remains. ↩

1. 2a.

   “the original tree of life concept is obsolete: it would not even be a 'tree of one percent'.  Nevertheless, there seems to be a discernible signal of consistency between the trees in the forest of life, down to the deepest branching levels.”  Printed as 59.12, which is page 12 of article 59.  The study used 6,901 maximum-likelihood gene trees from 100 prokaryote genomes, and the 102 nearly universal trees have inconsistency scores of 1.4% to 4.3% against a random-tree baseline near 80%.  Consistency is high at shallow depths.  Page 11:  “abruptly drops, almost down to the level of random trees”.  This happens at the deepest branchings. ↩

Wood, Todd Charles.  “Baraminological Analysis Places Homo habilis, Homo rudolfensis, and Australopithecus sediba in the Human Holobaramin.”  Answers Research Journal 3:71–90.  2010.  A creationist paper that reran eight published cladistic studies of fossil and living hominids with statistical baraminology, a method that groups organisms into created kinds by how alike their characters are.  It found that the hominins split into as many as four holobaramins, with the genus Homo including Australopithecus sediba, and it cites Wood's own 2009 survey, in which 28 of 73 earlier studies (38.4%) gave "uncertain" results.  The article cites it for the author's admission that the criteria for choosing characters are unknown. ↩ ↩

1. 3a.

   Wood raises the objection against his own position and does not answer it, then quotes Darwin at length without rebuttal. ↩ ↩

Measured for this article.  research/sim/rf-null.py, 20,000 replicates per row, seed 1.  Trees drawn uniformly over labelled unrooted binary topologies by sequential random-edge insertion.  The expected number of shared non-trivial splits between two random trees falls from 0.24 at 10 taxa to 0.13 at 100 and converges on 1/8.  Tree counts are (2N−5)!!, computed exactly. ↩

Kumar, Sudhir et al.  “Statistics and truth in phylogenomics.”  Molecular Biology and Evolution 29(2):457–472.  2012.  “For genome-scale data sets, the site-sampling variance reduces to virtually zero.”  “In contrast, the systematic error does not appear to be reduced (and may even increase) with the increasing size of data sets and the antiquity of the species included, so it becomes the central determinant of statistical support.”  A review of how statistical confidence behaves when trees are built from whole genomes.  With that much data the random error shrinks almost to nothing, so a branch can carry very high support while the result is wrong, because the systematic bias of the method is what remains.  The article uses it to explain why the bootstrap support value, a resampling test that gives each branch a percentage, fails on large data sets. ↩ ↩

Cunningham, C. W. et al.  “Best-fit maximum-likelihood models for phylogenetic inference: empirical tests with known phylogenies.”  Evolution 52(4):978–987.  1998.  A laboratory experiment with bacteriophage T7 lineages grown along a known tree, so the correct answer is known in advance.  When terminal branch lengths were very unequal, parsimony (choosing the tree that needs the fewest changes) supported the wrong tree in 96% of bootstrap replicates, while the best-fit model supported the correct tree with 86% bootstrap support.  The article uses it as its documented case of a method giving high support to a known-wrong tree. ↩ ↩

Collard, Mark, and Bernard Wood.  “How reliable are human phylogenetic hypotheses?”  PNAS 97(9):5003–5006.  2000.  A research paper in which two anthropologists ran the methods used on fossil human skulls and teeth on apes and monkeys whose DNA trees are well established.  The skull and tooth trees disagreed with the DNA trees, and the authors concluded that little confidence can be placed in phylogenies built from higher-primate skull and tooth evidence alone.  The article uses it to show that bootstrap support can be high for trees that are false. ↩ ↩

1. 7a.

   Page 5005:  “More problematically, the tests suggest that such analyses can strongly support phylogenetic hypotheses that are misleading.  For example, the bootstrap-based tests indicate that craniodental data can return impressive levels of statistical support (e.g., 97%) for patterns of phylogenetic relationship that are most likely incorrect.  In other words, cladistic analyses of higher primate craniodental morphology may yield not only 'false-positive' results, but false-positive results that pass, by a substantial margin, the statistical test favored by many researchers.”  The seven clades at or above 70% were 98, 97, 95, 92, 78, 73 and 73. ↩

Salichos, Leonidas, and Antonis Rokas.  “Inferring ancient divergences requires genes with strong phylogenetic signals.”  Nature 497:327–331.  2013.  A phylogenomic study of 1,070 orthologous genes (genes that descend from one ancestral gene) from 23 yeast genomes, testing how gene trees conflict with the tree built by joining all the genes into one data set, a method called concatenation.  The authors found that all 1,070 gene trees differ from the concatenated tree, that the conflict is worst at short branches deep in the tree, and that choosing genes or branches with strong support makes the result more robust, with similar results in vertebrate and animal datasets.  The article uses it to show that gene trees conflict a lot, and that the authors also report which corrections work. ↩ ↩

1. 8a.

   Figure 1b:  Average normalised Robinson-Foulds distance of the 1,070 yeast gene trees to the species tree 0.40; among the gene trees themselves 0.52; among 1,000 random trees 0.99.  This is the single most useful triple in the gene-conflict literature, and it sits in the same paper as the headline figure of 1,070 distinct topologies. ↩ ↩

2. 8b.

   The 18-taxon vertebrate dataset had 15 gene trees of 1,086 identical to the concatenation tree.  The famous result that no gene tree matches is partly a function of taxon number. ↩

3. 8c.

   Figure 2, page 329:  The authors' own audit of standard remedies on their own data.  Excluding gap-containing sites lowered internode certainty at 7 internodes and raised it at 0.  Using the 100 slowest-evolving genes lowered it at 9 and raised it at 2.  Selecting the 131 genes with highest average bootstrap improved gene support frequency at 15 internodes and worsened none. ↩

Measured for this article.  Wickett et al. 2014 never state a median.  Their Figs. 2 and 3 print each branch as bootstrap,conflict %, where conflict is the percentage of gene trees disagreeing at above 75% bootstrap.  Extracting those labels from the PDF text layer (pdftotext -layout, then the pattern (\*|\d{1,3}),(\d{1,2})) gives 41 labelled branches on Fig. 2 with median 9% and 50 on Fig. 3 with median 4% — the 91st label is 2,6 beside Oryza sativa, whose one-digit bootstrap a two-digit pattern misses;  pooled, 91 branches with median 6%, minimum 1% and maximum 16%.  Re-extracted by the orchestrator on 2026-08-22 with the same command (41 + 50 = 91, median 6).  The label order is inferred from the caption, which names both quantities and gives no other candidate. ↩ ↩

Measured for this article.  research/sim/compare-domains.py and research/sim/distro-truth-test.py, seeds 1-8.  Package data from Repology (17,911 projects × 278 repositories); documented genealogy from the GNU/Linux Distribution Timeline, 450 distributions with a curated parent.  Thirty-seven distributions have both.  Scored at 24 taxa on the 6,600-7,500 quartets the documented genealogy resolves; about 35% of quartets are unresolved because ten of the thirty-seven are direct children of Debian, and an earlier version of this test scored those as disagreements and understated the match.  Raw output in research/results/distro-truth-test.csv. ↩ ↩

Measured for this article.  The Computer-Assisted Stemmatology Challenge heinrichi tradition — Piispa Henrikin surmavirsi, hand-copied in pencil by 17 volunteers under controlled conditions, 37 manuscripts × 1,208 loci, with the true copying graph released.  Split-half QS 0.754 at 24 taxa and 0.884 using all 668 informative characters.  The comparison against the released stemma is research/sim/stemma_run_truth.py, written up in research/stemma-truth-test.md with raw output in research/results/stemma-truth-test.csv:  neighbour joining on Hamming distance over the 690 informative characters of the 34 witnesses that appear in the stemma scores QS +0.487 against it and parsimony +0.846, where the truth resolves 40,578 of the 46,376 quartets and a null of 2,000 random labelled trees has a 95th percentile of +0.066;  the 21 witnesses of the 0.8837 run that appear in the stemma score +0.78.  The dataset carries a licence restriction:  analysed in house, not redistributed.  Note that published accuracy figures on this corpus are close to meaningless without their floor — the no-information baseline of the challenge's own scoring rule is 56.6%, and the published neighbour-joining submission's "direction of copying" score of 58.467% is exactly what a graph with no edges scores, which is why nothing here is claimed about who copied whom. ↩

Nilsson, Maria A. et al.  “Tracking marsupial evolution using archaic genomic retroposon insertions.”  PLoS Biology 8(7):e1000436.  2010.  A genome study of marsupial mammals, such as kangaroos and opossums, which screened about 217,000 retroposon-containing loci (places where copies of jumping DNA sit) from opossum and kangaroo genomes and kept 53 informative markers.  The authors report that the four Australasian orders share a single origin, with Microbiotheria as their closest relative, and that the South American opossums branch off first.  The article uses the 53 markers as its example of rare genomic changes, a different kind of evidence from sequence similarity. ↩ ↩

Schultz, Darrin T. et al.  “Ancient gene linkages support ctenophores as sister to other animals.”  Nature 618(7963):110–117.  2023.  A genome study that compared which genes sit on the same chromosome across animals and their single-celled relatives, a pattern called synteny.  Chromosome fusions that could not be reversed placed comb jellies as the first branch off from all other animals, and a shuffling test on randomized genomes showed that chance does not produce the same pattern.  The article uses it as the clearest case of a branch that sequence data could not settle and chromosome structure did. ↩ ↩ ↩

Benton, Michael J., and Rebecca Hitchin.  “Congruence between phylogenetic and stratigraphic data on the history of life.”  Proceedings of the Royal Society B 264(1383):885–890.  1997.  A survey of 384 published family trees (cladograms) for echinoderms, fishes and tetrapods, comparing the branching order of each tree with the order in which its fossils appear in the rock.  It found good agreement on a variety of comparative metrics, which the authors read as a sign that most trees are broadly accurate and that the fossil record is a reasonably faithful sequence.  The article uses it as the main test of trees against fossil order. ↩ ↩

Measured for this article.  research/sim/convergence-curve.py, seed 20260815, 15 random half-splits per cell, taxon count held at 24 throughout, except the 19-sequence chloroplast alignment, so the only thing changing is how much evidence each half receives.  Raw output in research/results/convergence-curve.csv. ↩

Collard, Mark et al.  “Branching, blending, and the evolution of cultural similarities and differences among human populations.”  Evolution and Human Behavior 27(3):169–184.  2006.  A comparative study that fitted the branching tree model used in biology to 21 biological datasets and 21 cultural datasets, then compared how well the model fitted each group.  The cultural datasets were not significantly less tree-like than the biological ones, which the authors read as evidence against the claim that blending has always dominated cultural evolution.  The article uses it as the most-cited result that artifacts fit trees about as well as organisms do, and it examines where that result fails. ↩ ↩ ↩ ↩

Nunn, Charles L. et al.  “Simulating trait evolution for cross-cultural comparison.”  Philosophical Transactions of the Royal Society B 365:3807–3819.  2010.  Page 3817:  “while high values may indicate phylogenesis, low values are uninformative… these metrics provide very low power.”  A simulation study of cultural traits that can pass sideways between groups as well as from parent to child.  Across 5,000 simulations the authors tested whether the consistency index and retention index can detect that sideways passing, and found that faster change lowers these indexes more than borrowing does, so a low value does not mean borrowing.  The article uses it to show that the two indexes are weak tests, because borrowing does not show up in them. ↩ ↩

Gray, Russell D. et al.  “On the shape and fabric of human history.”  Philosophical Transactions of the Royal Society B 365:3923–3933.  2010.  A paper for a theme issue on cultural and linguistic diversity that sets out two debates, about how tree-like human history is (its shape) and how coherent its lineages are (its fabric), and then tests them with computational methods from biology.  It argues that these debates are better seen as positions on continuous dimensions, and it compares how tree-like vocabulary evolution is in different parts of the world.  The article uses it for the Polynesian canoe comparison, in which both the functional and the symbolic canoe traits lie close to the random distribution on the language tree. ↩ ↩

Measured for this article.  research/sim/ewert_repro.py against Ewert's own released supplementary archive (sources/data/ewert-2018c_dependency-graph/supplemental.zip, md5 b7741f8021252a5d470fa3eca71fb14e).  All fifteen datasets reproduce in direction — nine biological databases, five simulated-evolution controls and the JavaScript control;  the ratio of reproduced to published log Bayes factor runs 1.04-1.79 on the nine biological databases, 0.82-0.93 on the five simulated controls and 1.08 on the JavaScript one.04-1.79 on the nine biological databases and 0.82-0.93 on the five simulated controls.  Surprisal share — the fraction of the tree model's own error that the graph removes — is 0.94% (UniRef-50), 3.12% (OMA), 3.52% (Pfam), 5.35% (HomoloGene), 5.49% (TreeFam), 8.81% (EggNOG), 10.03% (OrthoDB), 13.76% (Ensembl) and 14.55% (Hogenom), with the JavaScript control at 5.64%.  Input matrices verified to agree cell-for-cell between his independently fitted tree and graph files.  Raw output in research/results/ewert-repro.csv. ↩ ↩ ↩ ↩ ↩

Ewert, Winston.  “The Dependency Graph of Life.”  BIO-Complexity 2018(3):1–27.  2018.  Page 21:  “In the case of the dependency graph model, there are many possible dependency graphs, and we need to assign a probability to each one.  This probability is only used in the case of the dependency graph model.”  A research article from the Biologic Institute by an author who writes as a defender of intelligent design, proposing that the nested pattern of life comes from shared software-style modules rather than from common descent.  It tests predictions of this model against biological and synthetic data, and the paper says that, as the first evaluation of the hypothesis, its tests are limited.  The article uses it as the one design model fitted to real genomes and compared head-to-head with a tree. ↩ ↩

1. 20a.

   Page 9:  The paper lists incomplete lineage sorting, gene flow, horizontal gene transfer, convergent evolution and gene resurrection as mechanisms that.  “occur in nature, but are not included in this model.”  Pages 5–6:  The concession reads.  “given the wide variety of mechanisms that might be invoked along with common descent, it would not be surprising that some combination of those could produce data which fit a dependency graph better than a tree.” ↩ ↩

Wood, Todd Charles.  “A baraminology tutorial with examples from the grasses (Poaceae).”  TJ (Journal of Creation) 16(1):15–25.  2002.  A short tutorial on baraminology, the creationist method for finding "created kinds" (baramins) by how alike organisms are, using the grass family Poaceae as its worked example.  The article cites it for how large a created kind is taken to be: one holobaramin of about 10,000 grass species, and about 150 fossil horse species in one monobaramin. ↩

1. 21a.

   On failing to find discontinuity, Wood wrote:  “should not [be considered] a failure.”  Paired with O'Micks 2016, who wrote: "We cannot demand that God created each baramin to have significant morphological differences compared to other baramins." This removes the model's falsifiability in its advocates' own words. ↩

2. 21b.

   The grass holobaramin of roughly 10,000 species is the worked example the tutorial is built around, and it is the clearest published statement of how large a created kind is taken to be. ↩ ↩

Reeves, Colin R.  “A Critical Evaluation of Statistical Baraminology: Part 1 - Statistical Principles.”  Answers Research Journal 14:261–269.  2021.  Page 263:  “commits a category error”.  “the use of a t-test is completely invalid”.  “we have no idea what the distribution of these values would be under either a null or an alternative hypothesis”.  “really should be abandoned”.  A critique of baraminic distance correlation, the statistical method that creationist researchers led by Wood use to test whether species fall into created kinds (baramins).  Reeves, a mathematician at Coventry University, argues that the method counts matching traits between every pair of organisms, runs a t-test on the results, and rests on a shaky understanding of statistical principles, so its use should be abandoned.  The article uses it as the most damaging critique of the method, and it comes from inside the creationist camp, in a creationist journal. ↩ ↩

1. 22a.

   The second part of the same critique, by the same author, on alternatives to the method and on its conceptual and practical problems, which the article cites for the alternatives it tests.  The citation is Reeves, C. R., A Critical Evaluation of Statistical Baraminology: Part 2-Alternatives and Conceptual and Practical Issues, Answers Research Journal 14:271-282, 2021, at https://assets.answersresearchjournal.org/doc/v14/statistical_baraminology_alternatives_practical_issues.pdf. ↩

Doolittle, W. Ford.  “Uprooting the Tree of Life.”  Scientific American 282(2):90–95.  2000.  A popular magazine feature in which Doolittle argues that gene sequences are reshaping the tree of life and that its early branches and root are in doubt.  The article uses it to show that a quotation in wide circulation is cut off from the sentence that follows it. ↩

Saitou, Naruya, and Masatoshi Nei.  “The Neighbor-joining Method: A New Method for Reconstructing Phylogenetic Trees.”  Molecular Biology and Evolution 4(4):406–425.  1987.  A methods paper that introduces neighbour-joining, a fast way to build a phylogenetic tree from a table of distances between organisms, by repeatedly joining the pair of taxa that gives the smallest total branch length.  The authors tested it by computer simulation against five other tree-making methods, and found that neighbour-joining and Sattath and Tversky's method were generally better than the rest at recovering the correct unrooted tree.  The article uses it as the standard distance method, the one its own software uses, and it sets its predecessor UPGMA against it. ↩

1. 24a.

   Table 6, page 420:  With rates varying among lineages, UPGMA recovered the true 8-OTU unrooted topology in 0 of 100 replicates at 600 nucleotides, against 75/100 and 72/100 for neighbour-joining on the same simulated trees. ↩

Felsenstein, Joseph.  “Evolutionary Trees from DNA Sequences: A Maximum Likelihood Approach.”  Journal of Molecular Evolution 17:368–376.  1981.  A methods paper that gave a computationally feasible way to find the maximum likelihood tree from DNA sequences under a simple model of DNA change, with a computer program to do it.  Unlike parsimony, which can give misleading results when rates of change differ between lineages, the method can test whether rates are constant and gives a rough indication of the error in the tree.  The article uses it to define maximum likelihood, the workhorse method. ↩

Mirarab, S. et al.  “ASTRAL: genome-scale coalescent-based species tree estimation.”  Bioinformatics 30(17):i541–i548.  2014.  A methods paper that introduces ASTRAL, a fast program that estimates a species tree from many gene trees when the conflict between them comes from incomplete lineage sorting, where ancestral gene versions have not yet sorted into their species.  The authors report that it is statistically consistent, runs on datasets of thousands of genes, and beats the other summary methods in every simulated condition they tried, though it is often more accurate than concatenation only when incomplete lineage sorting is high and there are enough gene trees.  The article cites it as an example of a species-tree method, which treats each gene tree as noisy evidence about the species tree rather than as a measurement of it.

1. 26a.

   ASTRAL is a species-tree method, so it treats each gene tree as noisy evidence about the species tree.  The authors describe it as statistically consistent and able to handle datasets with thousands of genes. ↩

Liu, Liang et al.  “A maximum pseudo-likelihood approach for estimating species trees under the coalescent model.”  BMC Evolutionary Biology 10:302.  2010.  A statistics paper that proposes MP-EST, a way to estimate a species tree from many gene trees under the multispecies coalescent model, which describes how gene lineages sort through ancestral populations.  The authors tested it on simulated data and on a placental mammal dataset, where its four major clades matched those of a Bayesian concatenation tree.  The article cites it with ASTRAL as a species-tree method, one that treats each gene tree as noisy evidence about the species tree.

1. 27a.

   The method this paper introduces, MP-EST, estimates the species tree directly from many gene trees.  The authors show it gets the right answer more and more reliably as the number of genes grows, even when the species tree is in the anomaly zone. ↩

Kubatko, Laura Salter, and James H. Degnan.  “Inconsistency of Phylogenetic Estimates from Concatenated Data under Coalescence.”  Systematic Biology 56(1):17–24.  2007.  A simulation study of concatenation, which joins all the genes into one long sequence before building a tree, when the coalescent (a model of how gene lineages sort in populations) makes individual gene trees conflict.  It found that concatenation becomes statistically inconsistent as genes are added, so more data drive it toward the wrong tree, and that the bootstrap can give moderate to strong support to that wrong tree.  The article uses it as the simulation evidence that gluing genes together can mislead. ↩

Roch, Sebastien, and Mike Steel.  “Likelihood-based tree reconstruction on a concatenation of aligned sequence data sets can be statistically inconsistent.”  Theoretical Population Biology 100:56–62.  2015.  A mathematical proof that building a tree by maximum likelihood from genes joined into one long sequence (concatenation) is statistically inconsistent, which means it converges on the wrong species tree as the data grow.  The authors work under the multispecies coalescent, the model of how gene trees arise inside a species tree, and say the proof gives a formal basis to earlier simulation results by Kubatko and Degnan.  The article uses it as the proof that concatenation can keep returning the wrong tree as more genes are added.  Mirror:  arXiv ↩

Felsenstein, Joseph.  “Cases in which Parsimony or Compatibility Methods will be Positively Misleading.”  Systematic Zoology 27(4):401–410.  1978.  A theoretical paper showing that parsimony (choosing the tree that needs the fewest changes) and compatibility methods can be inconsistent: when rates of change differ between lineages, they converge on the wrong tree as more data are added.  It works out the exact conditions for simple three- and four-species cases.  The article uses it for long-branch attraction, in which two lineages that have each changed a lot are pulled together on the tree. ↩

1. 30a.

   Page 407:  Equation (18), for unrooted Wagner parsimony on a four-taxon tree, consistency requires P² < Q(1−Q), where P is the probability of change on the two long non-sister branches and Q on the short branches. ↩

2. 30b.

   Page 408:  The sentence is aimed at defenders of parsimony, and it applies equally to anyone dismissing a phylogenetic result by naming an assumption without showing that violating it changes the answer. ↩

Wong, Karen M. et al.  “Alignment Uncertainty and Genomic Analysis.”  Science 319(5862):473–476.  2008.  A research report that tested how much the choice of sequence alignment (the step that lines up matching positions across genes) changes later results.  The authors aligned 1,502 sets of matching genes from seven yeast species with seven alignment programs, and 46.2% of the genes gave one or more different trees depending on the program used.  Trees from dissimilar alignments also had lower bootstrap support, a measure of confidence in a branch, and the article cites it for the point that alignment choice can change the tree. ↩

Abadi, Shiran et al.  “Model selection may not be a mandatory step for phylogeny reconstruction.”  Nature Communications 10:934.  2019.  A simulation and real-data study of the step where software picks the mathematical model of DNA change before it builds a tree.  The authors found that the standard selection criteria often pick different models yet give nearly the same tree, and that skipping the step and using the most parameter-rich model, GTR+I+G, gave similar trees.  The article uses it against the charge that researchers tune the model until the tree looks right. ↩

1. 32a.

   Section Results:  AIC and BIC selected different models in 62% of empirical datasets, yet the minimum pairwise agreement on inferred topology between any two selection criteria was 83% over 7,200 simulated datasets, and all distances to the true tree were statistically indistinguishable (Wilcoxon signed-rank, Bonferroni corrected). ↩

Shavit, Liat et al.  “The Problem of Rooting Rapid Radiations.”  Molecular Biology and Evolution 24(11):2400–2411.  2007.  A computer simulation study that generated DNA sequences on trees shaped like a rapid radiation, a burst of many species splitting in a short time, and tested which methods rebuilt the right tree.  It found that several standard methods were biased toward specific tree shapes, and that rooting with one outgroup (a distant species used to set the direction of the tree) often broke an otherwise correct tree.  The article uses it to show that rooting is the weakest step in tree building. ↩

Aberer, Andre J. et al.  “Pruning Rogue Taxa Improves Phylogenetic Accuracy: An Efficient Algorithm and Webservice.”  Systematic Biology 62(1):162–166.  2013.  A methods paper that presents a faster algorithm for finding rogue taxa, which are species whose place in a set of trees is unstable, and a free web service that runs it.  On simulated datasets whose true tree is known, removing the rogues made the bootstrap consensus trees and the maximum-likelihood trees closer to the true trees.  The article uses it to test the charge that dropping badly fitting taxa makes a tree look better than the data allow. ↩

1. 34a.

   Section Accuracy, page 165:  The study used 400 simulated datasets, described as.  “for which the 'true' tree is known.”  Pruning an equal number of taxa at random had a negative effect on congruence with the true tree, while informed pruning improved it, with support gain tracking truth gain linearly. ↩

2. 34b.

   Page 162:  “are mostly simply excluded (pruned) in current studies.”  This is the limitation that survives:  the published tree carries no answer for the taxa whose placement was contested. ↩

Working note on the tree of life: Pruning Rogue data.  n.d.  A section of this project's own working notes on the tree of life.  Its paraphrase of the paper it discusses is withdrawn.  The article cites it for the paraphrase of the paper that the rogue-taxa charge usually cites. ↩

Shen, Xing-Xing et al.  “Dissecting Incongruence between Concatenation- and Quartet-Based Approaches in Phylogenomic Data.”  Systematic Biology 70(5):997–1014.  2021.  A study that compared two ways of building a species tree from genome-wide data, using three published datasets from animals, fungi and plants.  It found that 30–36% of the genes in each dataset point the two methods to different answers, and that dropping those genes helped when gene-tree error and lineage sorting (genes whose history differs from the species history because of ancestral variation) were low.  At higher levels the disagreement disappeared, but the trees that remained were not always the true ones, and the article cites this for agreement without accuracy. ↩

Hillis, David M., and James J. Bull.  “An empirical test of bootstrapping as a method for assessing confidence in phylogenetic analysis.”  Systematic Biology 42(2):182–192.  1993.  A computer-simulation and laboratory study, using a laboratory-generated phylogeny, of whether bootstrap values from parsimony analyses measure what they are usually taken to measure.  It found bootstrap values unbiased but very imprecise as measures of repeatability, and found that a value of 70% or more usually meant at least a 95% chance that the clade is real only under equal rates of change, symmetric trees and at most 20% internodal change; with unequal rates, values above 50% overestimate accuracy.  The article uses it to show that the familiar 70% rule is conditional, and that its conditions are rarely checked. ↩

Erixon, Per et al.  “Reliability of Bayesian posterior probabilities and bootstrap frequencies in phylogenetics.”  Systematic Biology 52(5):665–673.  2003.  A simulation study that analysed the same simulated DNA data by maximum likelihood bootstrapping and by Bayesian analysis, using identical models, so the true tree is known.  Bayesian posterior probabilities were higher than bootstrap frequencies (how often a branch recurs when the data are resampled) for true clades, but they also led to more wrong conclusions, and both kinds of error grew sharply when the model was too simple.  The article uses it for the clearest comparison of how often well-supported branches turn out to be wrong. ↩

Lewis, Paul O. et al.  “Polytomies and Bayesian phylogenetic inference.”  Systematic Biology 54(2):241–253.  2005.  A simulation study of Bayesian tree-building, a method that gives each branch a posterior probability.  The authors simulated data from a star tree, which has no true internal branches, in 25 replicates of 1 site or 100,000 sites, and in replicate 11 one arbitrary resolution reached a posterior probability of 0.9892.  The article uses it for the case where a true internal branch has zero length, so that the three possible answers do not settle at one-third each. ↩

Yang, Ziheng.  “Fair-balance paradox, star-tree paradox, and Bayesian phylogenetics.”  Molecular Biology and Evolution 24(8):1639–1655.  2007.  A theory paper that approximates Bayesian posterior probabilities for three species with binary characters and checks the result by simulation.  It confirms the star-tree paradox: when the true tree is a star with no resolved internal branch, the probabilities for the three possible trees do not settle at one-third each as data grow, and they stay very sensitive to the prior.  The article cites it for why those probabilities stay polarised when a true branch has zero length. ↩

Yang, Ziheng, and Bruce Rannala.  “Branch-length prior influences Bayesian posterior probability of phylogeny.”  Systematic Biology 54(3):455–470.  2005.  A simulation and real-data study of how the prior (the starting assumption about how long each internal branch of the tree is) changes Bayesian posterior probabilities, which are scores for how likely a branch is to be true.  The authors found that priors assuming long internal branches gave high probabilities to trees, and that the bootstrap and the posterior probability are different measures of accuracy.  The article cites it for the effect of the branch-length prior on Bayesian probabilities. ↩

Phillips, Matthew J. et al.  “Genome-Scale Phylogeny and the Detection of Systematic Biases.”  Molecular Biology and Evolution 21(7):1455–1458.  2004.  A short letter that re-analysed a dataset of 106 genes from eight yeast species, which had given a tree with 100% bootstrap support on every branch.  The authors found that a different method, minimum evolution with a distance measure, gave a different tree, also at 100% bootstrap support, and traced the cause to a GC/AT bias, a skew in the mix of bases that can mislead tree-building.  Recoding the bases as purines and pyrimidines reduced the bias and reinforced the original tree.  The article uses it as a real-data case of two contradictory answers, both at maximum bootstrap support, from one dataset. ↩

Minh, Bui Quang et al.  “New methods to calculate concordance factors for phylogenomic datasets.”  Molecular Biology and Evolution 37(9):2727–2733.  2020.  A short methods paper that adds two agreement measures to the IQ-TREE software.  The gene concordance factor is the percentage of gene trees that contain a branch, and the site concordance factor is the percentage of informative sites that support it.  The authors tested both on ten empirical datasets and argue that they show what the bootstrap leaves out.  The article uses it to explain why the field moved on from the bootstrap to these measures. ↩

1. 43a.

   “resampled datasets may always return the same tree (i.e., 100% bootstrap support), even though incomplete lineage sorting or other processes that lead to genealogical discordance are at work (e.g., gCF and sCF values ≪100%).”  On their 3,220-locus lizard dataset the bootstrap is 100% on every branch but one and the smallest gene concordance factor is 4.5%.  Note that the site concordance factor's no-information baseline is 33%, not 0%, because a random site picks one of three resolutions. ↩

Hillis, D. M. et al.  “Experimental Phylogenetics: Generation of a Known Phylogeny.”  Science 255(5044):589–592.  1992.  A laboratory experiment that generated a phylogeny whose true answer was known in advance, so that methods could be scored against it.  The article uses it as its main example of methods recovering a known tree, and it carries the authors' own caveat about the tree shape they chose. ↩

1. 44a.

   Page 590:  “We chose to construct a symmetric phylogeny with equal distances among nodes (Fig. 1).  This topology (tree shape) has proven especially amenable to accurate reconstruction in theoretical studies and may thus be regarded as a 'null' model against which other topologies may be compared.”  Reporting the 1992 success without this sentence overstates it. ↩

Hillis, David M. et al.  “Application and Accuracy of Molecular Phylogenies.”  Science 264(5159):671–677.  1994.  Page 675:  “all methods estimated an incorrect phylogeny when using high-resolution restriction fragments, but they estimated the correct phylogeny when using restriction sites.”  “The sequence data consist of 1091 base pairs across four genes of T7 ... There are only 63 variable sites across the sequences.”  With those sequences.  “only parsimony and weighted parsimony estimate the correct tree”.  while maximum likelihood, neighbour-joining, Fitch-Margoliash and UPGMA.  “each estimate a single, incorrect tree”.  A Science article on how molecular phylogenies are tested, by computer simulation and by experiments on organisms whose history is known.  It argues that the two approaches are complementary, that simulations can be biased toward the method being tested, and that known phylogenies add a reality check.  The article uses it for the same known T7 tree, where the methods gave wrong answers from restriction-fragment data and four of six were wrong on 1,091 base pairs containing only 63 variable sites. ↩

Strait, D. S., and F. E. Grine.  “Inferring hominoid and early hominid phylogeny using craniodental characters: the role of fossil taxa.”  Journal of Human Evolution 47(6):399–452.  2004.  A cladistic study (a method that builds family trees from shared traits) of one hundred and ninety-eight craniodental characters (skull and teeth features) from apes and early hominids.  It tested whether adding fossil species changes the tree: trees from living apes alone were inconsistent with the molecular ape tree, while trees that included fossil hominids matched it, which the authors took to refute the claim that these characters are unreliable.  The article cites it as one side of a dispute with Collard and Wood, in which Strait and Grine replied that adding fossils fixes the problem. ↩

Collard, Mark, and Bernard Wood.  “Hominin homoiology: an assessment of the impact of phenotypic plasticity on phylogenetic analyses of humans and their fossil relatives.”  Journal of Human Evolution 52(5):573–584.  2007.  A study that tested the homoiology hypothesis, which says that shape changes from chewing strain in the hominin skull, not shared descent, mislead trees built from it.  Using hominoid (ape) skull measurements with a robust molecular tree, the authors found that the strain-affected measurements varied more, as predicted, but the trees built from them matched the molecular tree better, so homoiologies had no significant negative effect.  The article uses it in the back-and-forth with Strait and Grine over hominin trees. ↩

Sayyari, E., and S. Mirarab.  “Fast Coalescent-Based Computation of Local Branch Support from Quartet Frequencies.”  Molecular Biology and Evolution 33(7):1654–1668.  2016.  A methods paper that gives each branch of a species tree (the tree of species, as opposed to the trees of single genes) a score for how likely it is to be correct.  The authors tested the score on simulated trees with known answers and on real datasets, and found it very precise and better than multi-locus bootstrapping, a resampling method, at recovering true branches.  The article cites it as a well calibrated support score, unlike the bootstrap values most studies still print. ↩

Penny et al.  “Testing the theory of evolution by comparing phylogenetic trees constructed from five different protein sequences.”  Nature 297(5863):197–200.  1982.  A 1982 test of evolutionary theory that builds a phylogenetic tree, a family tree of species, from each of five protein sequences and compares the trees with one another.  The article uses it as the origin of the test it describes, in which agreement between trees counts as evidence only if disagreement was possible. ↩

Robinson, and Foulds.  “Comparison of phylogenetic trees.”  Mathematical Biosciences 53(1–2):131–147.  1981.  A 1981 paper on how to compare two phylogenetic trees.  The article uses it for the Robinson-Foulds distance, which counts the groupings that are present in one tree but not in the other. ↩

Smith, Martin R.  “Information theoretic generalized Robinson-Foulds metrics for comparing phylogenetic trees.”  Bioinformatics 36(20):5007–5013.  2020.  A methods paper that proposes new measures of how far apart two phylogenetic trees are, in bits of information, and tests them against the standard Robinson-Foulds measure.  In a test on 100,000 random 11-leaf trees compared with a pectinate (caterpillar-shaped) reference tree, 86% sat at the maximum Robinson-Foulds distance, and that measure reported fewer than eight distinct distance values.  The author concludes that mutual clustering information should be preferred to Robinson-Foulds, and the article uses the test to show that the standard measure is a poor gauge of how close two trees are. ↩

Darlu, Pierre, and Guillaume Lecointre.  “When does the incongruence length difference test fail?”  Molecular Biology and Evolution 19(4):432–437.  2002.  A simulation study of the incongruence length difference (ILD) test, a standard significance test for whether two sets of characters, such as DNA sequences, conflict with each other.  Under the same evolutionary conditions the test rejected agreement less often than its nominal 5% level, but it had little power to detect real conflict unless there were many informative sites and the rate of change was similar across sites.  The article uses it to show that the standard test for conflict is weaker than its use implies. ↩

Sanderson, Michael J., and Michael J. Donoghue.  “Patterns of variation in levels of homoplasy.”  Evolution 43(8):1781–1795.  1989.  A statistical study of 60 published cladistic analyses, which are trees built from several kinds of characters, such as anatomy or molecular data.  It found that the consistency index, a score of how well the data fit a tree, falls as more taxa are added, because homoplasy (the same trait evolving independently more than once) rises.  The article uses it for that effect of taxon number on the index. ↩

Steell, Elizabeth M. et al.  “Revealing patterns of homoplasy in discrete phylogenetic datasets with a cross-comparable index.”  Zoological Journal of the Linnean Society 204(1):zlaf024.  2025.  A methods paper that tests the consistency index (CI) and retention index (RI), two widely used measures of homoplasy (the same trait evolving more than once), on simulated and real morphological datasets, and proposes a new index, the relative homoplasy index (RHI).  It measures CI, RI and RHI against taxon number on the same datasets and finds that they move in opposite directions, whereas Sanderson and Donoghue measured only the consistency index.  The article cites it for the point that the indices used in the homoplasy literature move with taxon number in opposite directions. ↩

Jarvis, Erich D. et al.  “Whole-genome analyses resolve early branches in the tree of life of modern birds.”  Science 346(6215):1320–1331.  2014.  A genome study by a large international team that built a tree of modern birds from 48 avian genomes and then checked it against each locus separately.  The main tree was strongly supported, but no single intron, exon or ultraconserved element locus had a topology identical to it, and the least certain branches were short ones at deep divergences.  The article uses it to show where the conflict between genes sits in the bird tree. ↩

1. 55a.

   “a wall of many (25 of 45; ~55%) very short internal branches (0.0006 to 0.002 substitutions per site), almost all at deep divergences within Neoaves.”  All nine branches below 100% bootstrap were among the shortest. ↩

2. 55b.

   Per-branch statistics: 0-38% of 2,485 intron gene trees conflict at above 75% bootstrap; only 0-20% support the same branches at that level.  The remainder is uninformative rather than conflicting, and the authors state.  “the conflicting nodes largely reflected branches with low statistical support.” ↩

Wickett, Norman J. et al.  “Phylotranscriptomic analysis of the origin and early diversification of land plants.”  PNAS 111(45):E4859–E4868.  2014.  A transcriptome study that generated data for 92 streptophyte taxa (the group of green algae and land plants) and added 11 published plant genomes, using up to 852 nuclear genes.  The authors ran 69 analyses with different data and methods to test how robust the tree was, and found robust support for land plants sitting with one group of green algae, the Zygnematophyceae.  The article cites it for the conflict percentages reported for 91 branches of its trees. ↩

Smith, Stephen A. et al.  “Analysis of phylogenomic datasets reveals conflict, concordance, and gene duplications with examples from animals and plants.”  BMC Evolutionary Biology 15:150.  2015.  A methods article describing phyparts, a program that maps onto a species tree how many gene trees agree with each branch, conflict with it, or contain gene duplications.  The authors applied it to published transcriptome datasets from ants, bees and wasps (5,863 gene groups) and from flowering plants (Caryophyllales), and found significant conflict throughout both trees.  The article cites it for the point that conflicting gene trees are the more poorly supported ones. ↩

Song, Sen et al.  “Resolving conflict in eutherian mammal phylogeny using phylogenomics and the multispecies coalescent model.”  PNAS 109(37):14942–14947.  2012.  A genome-scale study of 447 nuclear genes from 33 placental (eutherian) mammal species and four outgroups, analyzed with coalescent methods, which model gene trees that can differ from the species tree.  The authors found that joining all genes into one data set gave strongly conflicting trees when they subsampled the data, while the coalescent methods gave a stable tree.  The article cites it for the finding that the coalescent model accounts for 77% of the variation in gene trees, a prediction of common descent rather than a failure of it. ↩

Reddy, S. et al.  “Why do phylogenomic data sets yield conflicting trees? Data type influences the avian tree of life more than taxon sampling.”  Systematic Biology 66(5):857–879.  2017.  A study of the early divergences in Neoaves, the largest group of living birds, which built a new matrix of 235 species that is 90% non-coding DNA to find out why two earlier avian phylogenomic studies disagree.  The authors found that data type mattered more than the number of taxa: the new trees agreed more with the earlier study built mostly on non-coding DNA, and they blame model fit, a problem that may be difficult to detect.  The article uses it to show that the disagreement tracks the kind of data more than the number of species, so more data does not settle it. ↩

Pisani, Davide et al.  “Genomic data do not support comb jellies as the sister group to all other animals.”  PNAS 112(50):15402–15407.  2015.  A reanalysis of three genome datasets that had been used to argue that comb jellies (ctenophores) are the closest living relatives of all other animals.  The authors say the earlier studies used simple models of how proteins change and distant outgroups to root the tree, and they corrected for both, along with a gene-loss bias.  With better-fitting models they found no support for comb jellies in that position, and they conclude the earlier support came from uncorrected bias.  The article uses it as one side of the dispute over the deepest branch of the animal tree, where the choice of model decides the answer. ↩

Li, Yuanning et al.  “Rooting the animal tree of life.”  Molecular Biology and Evolution 38(10):4322–4333.  2021.  A reanalysis of whether sponges or comb jellies (ctenophores) are the closest living relatives of all other animals.  The authors gathered 15 earlier phylogenomic studies and re-ran them under the same conditions with more models.  Comb jellies came out as the closest relatives across the whole range of conditions, while sponges appeared only in narrow settings, with most outgroups removed and a complex model with hundreds of extra categories, which fit no better than simpler models.  The article uses it as an example of the deepest branch of the animal tree flipping with the model. ↩

Pease, James B. et al.  “Phylogenomics reveals three sources of adaptive variation during a rapid radiation.”  PLoS Biology 14(2):e1002379.  2016.  A genome-wide study of all 13 species of wild tomato, using whole transcriptomes (the sets of genes switched on in a plant's tissues) to reconstruct how the group diversified.  The authors date the radiation to about 2.5 million years ago and find three sources of adaptive variation: hybridisation between species after they split, new genes evolved within lineages, and ancestral variation sorted differently across populations.  The article uses it for gene exchange between wild tomato lineages, which it traces to hybridisation rather than to the sorting of ancestral variation. ↩

Shen, Xing-Xing et al.  “Contentious relationships in phylogenomic studies can be driven by a handful of genes.”  Nature Ecology & Evolution 1:0126.  2017.  A reanalysis of published phylogenomic datasets from plants, animals and fungi, measuring how much each gene and site favours each branch of the tree.  The authors examined 17 contested branches and 6 well-established control branches, and found that removing the single most influential gene flipped the best-supported answer on 3 of the contested branches.  The article uses it to show that a single gene can flip a branch only where the branch was already contested, since removing random genes changed nothing and removing the top gene from a control branch did not switch its answer. ↩

Degnan, James H., and Noah A. Rosenberg.  “Discordance of species trees with their most likely gene trees.”  PLoS Genetics 2(5):e68.  2006.  A theoretical study that uses the coalescent model, a population-genetics model of how gene lineages merge backward in time, to prove a counterintuitive result.  For any species tree with five or more species there are branch lengths at which the most likely gene tree is not the species tree, so taking the most common gene tree as the species tree can converge on a wrong answer as more genes are added.  The article uses it for the proof that this anomaly zone exists. ↩

Linkem, Charles W. et al.  “Detecting the anomaly zone in species trees and evidence for a misleading signal in higher-level skink phylogeny.”  Systematic Biology 65(3):465–477.  2016.  A phylogenomic study of 16 lizard species in the skink family (Scincidae), using 429 new loci of DNA, that tests whether a hard-to-resolve part of the tree lies in the anomaly zone.  That is a set of conditions in which a wrong gene tree is more probable than the true species tree, and the authors checked for it by estimating ancestral population sizes and how long each species persisted.  They found at least three regions of the skink tree with the pattern the anomaly zone predicts, and say this helps reconcile conflicting earlier studies.  The article uses it as a real case showing that the anomaly zone exists in practice. ↩

Tehrani, Jamshid, and Mark Collard.  “Investigating cultural evolution through biological phylogenetic analyses of Turkmen textiles.”  Journal of Anthropological Archaeology 21:443–463.  2002.  A study that applied cladistic tree-building to decorative designs on woven Turkmen textiles, to test whether cultural change works more like branching descent or like borrowing between neighbouring groups.  Before Tsarist Russia's defeat of the Turkmen in 1881 the consistency index was 0.68 and branching accounted for about 70% of the resemblances; after 1881 it was 0.61, and the index sits in the same range as published hominid cladograms (family trees of the great apes and humans), which score 0.58 to 0.71.  The article cites it as one of the two artifact studies that point in opposite directions. ↩ ↩

O'Brien, Michael J. et al.  “Cladistics Is Useful for Reconstructing Archaeological Phylogenies: Palaeoindian Points from the Southeastern United States.”  Journal of Archaeological Science 28(10):1115–1136.  2001.  Figure 11:  “The shortest tree we found ... has a length of 22 and a CI of 0.59”.  Collard et al. 2006, Table 3, gives the retention index of 0.70 for the same 17 taxa and 8 characters.  A cladistics study, a method that sorts objects into a nested family tree by the features they inherited, applied to seventeen Palaeoindian stone projectile-point classes from the southeastern United States, coded for eight characters.  The authors built the shortest tree, which needs the fewest changes, and argue that such trees are testable hypotheses about how the point styles descended from one another.  The article uses it for the consistency index of 0.59 on this dataset, and sets against it the retention index of 0.70 that Collard and colleagues report for the same data. ↩

Rogers, Deborah S., and Paul R. Ehrlich.  “Natural selection and cultural rates of change.”  PNAS 105(9):3416–3420.  2008.  A study of design features on Polynesian canoes, split into functional features, which are tested against the environment, and symbolic features, which are not.  The authors found that natural selection seems to slow the change of functional designs, while symbolic designs differentiate more rapidly, which they take to show that cultural change can follow theoretically derived patterns.  The article uses the canoe traits as its test case, in which functional design fits the true history worse than randomised data does. ↩ ↩

Measured for this article.  research/sim/calibrate.py --plan rate --seed 20260815 for the ceiling:  process S1, Mk characters on a random tree with equal branch lengths at substitution rates 0.01 to 0.05 and 1,000 characters, reads exactly 1.0000 at 12 taxa on all three seeds and at 24 taxa at rate 0.02 on all three.  Give the same tree exponentially distributed branch lengths, so that some branches are too short to resolve, and it reads 0.75 to 0.98 instead — which is what the qualification in the sentence above is doing.  research/sim/calibrate.py --plan design --seed 20260815 for the floor:  process S5, characters drawn independently per object, gives a median of −0.003 over eighteen conditions (12, 24 and 48 taxa × two character budgets × three seeds), range −0.092 to +0.091.  Raw output in research/results/calibration-summary-rate.csv and research/results/calibration-summary-design.csv. ↩

Swamidass, S. J.  “Winston Ewert: The Dependency Graph of Life.”  Peaceful Science discussion thread.  Jul 21, 2018.  An online discussion thread, not a peer-reviewed paper, in which Swamidass replied to Ewert's paper; 23 posts survive, numbered up to 35, and Ewert took part for one day and then withdrew.  Post #25 makes the scale argument that a Bayes factor must be read against the size of the model's total surprisal, not in isolation, and Swamidass quoted the UniRef-50 row with its column labels transposed, though the arithmetic and the point stand.  The article cites it as the source of the scale argument it reframes, and this thread and its 2023 sequel are the entire published response to the paper. ↩

Lammers, Fritjof et al.  “Phylogenetic Conflict in Bears Identified by Automated Discovery of Transposable Element Insertions in Low-Coverage Genomes.”  Genome Biology and Evolution 9(10):2862–2878.  2017.  A genome study of seven bear species that built a pipeline, TeddyPi, to find transposable elements (pieces of DNA that copy and move themselves) in low-coverage genomes, and used them to reconstruct the bear family tree.  The authors got a well-resolved species tree, and they trace the conflict around the Asiatic black, sun and sloth bears to incomplete lineage sorting, where ancestral gene versions have not yet sorted out between species.  The article uses it as a counterweight: this marker class is treated as nearly free of repeated change, and it still gives conflicting counts. ↩

O'Connor, Anne, and Matthew A. Wills.  “Measuring Stratigraphic Congruence Across Trees, Higher Taxa, and Time.”  Systematic Biology 65(5):792–811.  2016.  A study of how well the order of branching in a tree matches the order in which fossils first appear, using 647 published animal and plant cladograms (family trees) spanning the entire Phanerozoic eon, the span of the animal and plant fossil record.  Eleven biasing variables account for 2% of the deviance (the share of variation in the scores that they explain) in the gap excess ratio (GER), against 26% or more for the other indices, so GER is the one to use.  The article cites it for the warning that published comparisons using the stratigraphic consistency index (SCI) and the modified Manhattan stratigraphic measure (MSM*) should be distrusted. ↩

Springer, Mark S. et al.  “The Adequacy of Morphology for Reconstructing the Early History of Placental Mammals.”  Systematic Biology 56(4):673–684.  2007.  A Points of View article that tested whether anatomy alone can rebuild the early history of placental mammals.  The authors ran pseudoextinction analyses, deleting the DNA of one living order at a time and re-running the tree on morphology alone.  The article cites it as the first of three studies on how well morphology recovers the molecular tree. ↩

Springer, Mark S. et al.  “Evolutionary Models for the Diversification of Placental Mammals Across the KPg Boundary.”  Frontiers in Genetics 10:1241.  2019.  A paper that re-ran the pseudoextinction experiment on O'Leary et al. 2013's 4,541-character matrix, which is twenty-four times larger than the 2007 dataset.  The majority of orders still moved, which kills the sample-size defence of morphology.  The article cites it for the re-run on the larger matrix. ↩

1. 74a.

   “all three orders with highly specialized myrmecophages (Xenarthra, Pholidota, Tubulidentata) cluster with one of the other myrmecophagous orders when treated as pseudoextinct.”  The failure is driven by diet, not by noise, which is why more characters do not fix it. ↩

Brady, Peggy L., and Mark S. Springer.  “The effects of fossil taxa, hypothetical predicted ancestors, and a molecular scaffold on pseudoextinction analyses of extant placental orders.”  PLoS ONE 16(9):e0257338.  2021.  “only 8 of 19 pseudoextinct placental orders (42%) retained the same interordinal placement as on the molecular scaffold.”  A study in which each placental order was treated as extinct in turn and placed by morphology, with fossils, reconstructed ancestors and a molecular scaffold (a fixed frame taken from DNA trees).  The authors' figure of 42% is the share of orders that kept the same placement with all three; without fossils or a scaffold, the figure was 21%.  The article cites it for how often morphology still gets the placement wrong, even with fossils and reconstructed ancestors. ↩ ↩

Wood, Todd Charles.  “Taxon Sample Size in Hominin Baraminology: A Response to O'Micks.”  Answers Research Journal 9:369–372.  2016.  A reply to O'Micks, who had concluded that Homo naledi was probably not part of the human kind, which re-runs O'Micks's own 87-character matrix at two taxon samples.  With the larger sample every pair within Homo correlated significantly, and with six species only one pair did, including Homo sapiens against Homo neanderthalensis, which lost significance.  The article cites it to show that the answer depends on which species are included. ↩

DeWitt, David A. et al.  “Baraminological Analysis Places Homo habilis, Homo rudolfensis, and Australopithecus sediba in the Human Holobaramin: Discussion.”  Answers Research Journal 3:153–158.  2010.  “it would appear that the statistical baraminology results placing Homo habilis, Homo rudolfensis, and Australopithecus sediba within the human holobaramin is a false positive result.”  The nine hominin analyses behind the article's sentence that creationist analysts disagree.  Au. sediba human: Wood 2010 (ARJ 3:71-90), Wood 2011 (JCTS B 1:6-14), Wood 2016 (JCTS B 6:1-8) and Sinclair & Wood 2021 (ARJ 14:451-462).  Called a false positive: DeWitt, Habermehl & Menton 2010 (ARJ 3:153-158), a reply paper that ran no baraminic distance analysis of its own.  H. naledi human: O'Micks 2016 (JCTS B 6), Wood 2016 (JCTS B 6:14-30), Sinclair & Wood 2021.  Not human: O'Micks 2016 (ARJ 9:263-272) and O'Micks 2017 (ARJ 10:103-113).  The factor-of-ten downweighting is in O'Micks's 2016 reply on taxon sample size.  A reply to Todd Wood's 2010 paper on the same fossils, which calls the result placing Homo habilis, Homo rudolfensis and Australopithecus sediba in the human kind a false positive, without running a baraminic distance analysis of its own.  The article cites it for the flat disagreement among creationist analysts on the same data. ↩

O'Micks, Jean.  “Reply to 'Taxon Sample in Hominin Baraminology: A Response to O'Micks'.”  Answers Research Journal 9:373–375.  2016.  Page 373:  “We cannot demand that God created each baramin to have significant morphological differences compared to other baramins.”  The same page reports the factor-of-ten downweighting of the craniodental characters and concedes that.  “the weighted BDIST program does not calculate bootstrap values.”  The earlier citation gave the title as.  “Reply to 'Taxon Sample Size in Hominin Baraminology: A Response to O'Micks'”.  A reply by Jean O'Micks, an independent scholar, to a critic of his baraminology study of Homo naledi, a fossil species.  Baraminology is a creationist method that sorts species into groups thought to descend from one created kind, and O'Micks argues that a weighted version of the BDIST method places H. naledi with the australopiths.  The article uses it to show a baraminologist conceding that these groups need not differ in form. ↩

Penny, David, and Michael Hendy.  “Estimating the reliability of evolutionary trees.”  Molecular Biology and Evolution 3(5):403–417.  1986.  A 1986 paper that tests how reliable an evolutionary tree is, using six real protein sequences from the same 11 mammals rather than computer simulations.  The authors measure how trees built from larger random subsets of nucleotide positions become more alike, and report that the rodent-primate-lagomorph trichotomy is the least certain part of the final tree, followed by the position of the dog.  The article takes the design and headline figures of the 1982 test from this follow-up.

Theobald, D.  “29+ Evidences for Macroevolution, Part 1: the unique universal phylogenetic tree.”  TalkOrigins Archive.  2012.  A web article in the TalkOrigins Archive, a long-running online encyclopedia on evolution and creation, that sets out predictions of common descent and includes a section on how many trees are possible for a set of species.  The article cites it for the author's counts of possible trees, which it checks against his own formula.

1. 80a.

   “nearly 14 trillion different possible phylogenies”.  The prose figure for 12 organisms, which is wrong: his own formula gives 13,749,310,575 rooted trees, about 13.7 billion.  Table 1.3.1:  Table 1.3.1 gives 34,459,425 unrooted trees for 11 taxa, which is correct.