Comparative And Ancestral Study

Comparative regression and ancestral reconstruction can share a tree and trait table without sharing an estimand. Admit their common population once, then preserve a separate model, result, and acceptance decision for each analysis.

The regression asks whether named predictors are associated with a response after accounting for the declared phylogenetic covariance. The ancestral analysis asks which states are supported at identified internal nodes under a declared transition model and root treatment. Similar inputs do not make these questions interchangeable.

flowchart TB
    tree["Tree identity"] --> join["Taxon reconciliation"]
    traits["Trait table<br/>units and coding"] --> join
    join --> regression["PGLS<br/>association estimand"]
    join --> ancestral["Discrete reconstruction<br/>node-state estimand"]
    regression --> review["Separate diagnostics<br/>and claims"]
    ancestral --> review

Join Lineage Without Joining Conclusions

The shared admission record supports a precise lineage statement: both analyses began from identified source objects and their derived populations are known. It does not create a combined scientific result.

Shared fact	Regression still owns	Ancestral analysis still owns
source tree and trait-table identity	response, predictors, formula, design columns	reconstructed trait, state space, node keys
taxon reconciliation ledger	rows eligible after response/predictor missingness	tips eligible after state ambiguity/missingness
branch-length and rooting record	covariance geometry and coefficient interpretation	direction, node identity, and transition interpretation
common software/run context	fit, residual, coefficient, and sensitivity diagnostics	model, node-probability, transition, and tree sensitivity

A cross-analysis sentence is defensible only when it cites both result identities and states the logical connection being made. Association between a predictor and response does not demonstrate a particular ancestral state; reconstruction of a state does not demonstrate that the state caused the regression pattern.

Admit The Shared Population

Before fitting either model, record the tree revision, rooting and branch length meaning; trait-table checksum; taxon column; units; factor reference levels; missing-value policy; and every taxon excluded from the join. Row order must not become an implicit taxon identifier.

Create an admission table before either fit:

Taxon	Tree state	Response/predictor state	Ancestral-trait state	Disposition
stable identifier	present or absent	observed, missing, transformed, or excluded	observed, missing, recoded, or excluded	admitted to regression, ancestral analysis, both, or neither

The table makes derived populations visible. If regression drops a taxon for a missing predictor while ancestral reconstruction retains it, the analyses no longer have one effective denominator even though they began from the same files.

The curated Python workflows expose both analyses:

from pathlib import Path

from bijux_phylogenetics.api import (
    run_ancestral_reconstruction_workflow,
    run_comparative_model_workflow,
)

tree = Path("study/tree.nwk")
traits = Path("study/traits.tsv")

regression = run_comparative_model_workflow(
    tree,
    traits,
    response="longevity",
    predictors=["body_mass", "activity_period"],
    taxon_column="taxon",
    lambda_value="estimate",
)
regression.write_json(Path("artifacts/comparative/pgls.json"))
regression.write_tsv(Path("artifacts/comparative/pgls-coefficients.tsv"))

ancestral = run_ancestral_reconstruction_workflow(
    tree,
    traits,
    trait="activity_period",
    taxon_column="taxon",
    model="equal-rates",
    root_prior_mode="equal",
)
ancestral.write_json(Path("artifacts/ancestral/activity-period.json"))
ancestral.write_tsv(Path("artifacts/ancestral/activity-period-nodes.tsv"))

Use either response plus predictors or the documented formula route for PGLS, and retain the exact choice. Changing a factor reference, interaction, transform, λ policy, ancestral model, state ordering, transition constraint, or root prior creates a different scientific result.

Read The PGLS Result

The coefficient table records estimate, standard error, test statistic, p-value, confidence interval, λ, log likelihood, AIC, and admitted taxon count. Interpret coefficients only with the design matrix, units, factor coding, covariance model, residual diagnostics, and reconciled population. A small p-value does not validate those upstream choices.

Estimated λ is part of the fitted model. A fixed lambda_value=1.0 asks a different question and must not be presented as a retry of estimation.

Read The Ancestral Result

Node estimates require stable node identity and the exact tree. A state label without node/clade identity, ambiguity, model, and root treatment is not a portable ancestral observation. Ordered states and allowed transition pairs encode scientific assumptions; preserve them even when the preferred state at a named node is unchanged.

Keep Claims Separate

PGLS estimates association conditional on the admitted covariance and design. Ancestral reconstruction estimates internal states conditional on the tree and transition model. Neither result validates the other. A shared input checksum establishes lineage, not independent evidence.

For publication, retain a common admission record plus separate JSON and TSV results, configurations, diagnostics, exclusions, and claim identifiers. If the two analyses use different taxon populations after filtering, record two derived dataset identities rather than implying one common denominator.

Use separate acceptance ledgers as well. The regression ledger should name the formula, coding, units, lambda policy, covariance assumptions, residual checks, and coefficient-level claim. The ancestral ledger should name the trait coding, transition model, state ordering, root prior, node or clade keys, and uncertainty. A favorable regression diagnostic cannot close an ancestral-model sensitivity, and a stable ancestral state cannot validate a coefficient.