Dating the branches of an evolutionary tree needs outside anchors. DNA differences only record rate multiplied by time, so the rate and the time cannot be separated without them. Biologists therefore pin several internal nodes to fossil ages, sample collection dates, or an assumed mutation rate, and run a Bayesian sampler such as BEAST 2. When these anchors disagree, the sampler quietly settles on a compromise timeline, often with deceptively tight uncertainty. It gives no account of which anchor pulled the answer where. We propose to treat each calibration as a "player" in a cooperative game and compute its Shapley value, the fair share of the change in estimated node ages that it causes. We also compute pairwise interaction terms that flag pairs of calibrations working against each other. The computation borrows directly from explanation tools for hyperparameter optimization.
The obstacle is cost. An exact Shapley analysis over $K$ calibrations needs the posterior under every subset, which means $2^K$ sampler runs that each take days. Our key technical observation is that calibrations enter the posterior only as multiplicative prior factors on node ages, while the expensive sequence likelihood is shared across all subsets. Every subset's posterior can therefore be recovered by reweighting samples from a small number of reference runs. When reweighting fails, the failure is itself a quantitative signature of conflict. The result is a conflict audit that costs roughly two to four MCMC runs instead of $2^K$.
Core Claim
Game-theoretic attribution with pairwise interaction terms identifies distorting and mutually conflicting calibrations more reliably than the one-calibration-at-a-time checks biologists use now. When computed by importance reweighting, it matches brute-force enumeration at a small fraction of the cost. The novelty is not "Shapley values for phylogenetics" in the abstract. It is a specific explanation game whose players are calibration priors, together with an estimator that exploits the fact that removing a calibration changes only the prior and never the likelihood. Nothing in the gathered literature does either.
Background and Positioning
The method comes from the optimization-transparency line. HyperSHAP defines several "explanation games" over hyperparameters: ablation, sensitivity, tunability, and optimizer bias. It shows that low-order Shapley interactions capture landscape dependencies that single-parameter measures miss. Earlier, surrogate fANOVA showed that varying one parameter at a time fails when parameters interact. We reuse HyperSHAP's ablation-game construction and its interaction indices. We differ in two ways. Our players are prior factors in a Bayesian posterior rather than configuration knobs. We also replace HyperSHAP's learned surrogate with an exact identity: reweighting is unbiased given enough overlap, and it carries no surrogate-fidelity risk. That fidelity risk is the limitation surrogate ablation inherits from its random-forest performance model.
Three papers motivate the problem. Performance of A Priori and A Posteriori Calibration Strategies simulates trees with independent rate speedups and slowdowns. It finds that calibrations must enter the joint Bayesian analysis to recover rates and times, and that without calibrations every method fails. That result makes calibrations the decisive inputs, yet the paper offers no way to apportion their influence. PrioriTree documents that prior choices dominate phylodynamic conclusions in BEAST, but it provides no attribution. Calibrating the Human Mutation Rate via Ancestral Recombination Density is a real-world case of conflicting anchors. Fossil-calibrated rates of about $2.0$–$2.5\times10^{-8}$ disagree with pedigree rates of about $1.0$–$1.2\times10^{-8}$, and the authors resolve this with an internal clock giving $1.61\pm0.13\times10^{-8}$ and a human–chimpanzee split at $6.8\pm0.6$ million years ago.
One paper constrains the design. The same a priori/a posteriori study implies we cannot compute attributions on a post-hoc rescaled tree, in the style of RelTime, because that two-stage route itself mis-dates the tree under rate shifts. Attribution must operate on the joint posterior.
Research Program
Let $\mathcal{C}=\{c_1,\dots,c_K\}$ be the calibrations, each a density $p_k(t_{n_k})$ on the age of node $n_k$. Let $\pi_0(\tau)$ be a base prior over the tree $\tau$ (topology, times, and branch rates) that contains no calibrations. For a subset $S\subseteq\mathcal{C}$, the posterior is $\pi_S(\tau)\propto L(D\mid\tau)\,\pi_0(\tau)\prod_{k\in S}p_k(t_{n_k})$, where $L$ is the sequence likelihood. Given samples $\tau^{(m)}$ from a reference posterior $\pi_R$ with calibration set $R$, the self-normalized importance weight for subset $S$ is
$$w_S^{(m)} \propto \frac{\prod_{k\in S} p_k\big(t^{(m)}_{n_k}\big)}{\prod_{k\in R} p_k\big(t^{(m)}_{n_k}\big)}.$$
This weight involves only calibration densities, so evaluating all $2^K$ subsets costs seconds even for $K$ around 15.
Uncalibrated posteriors are nearly unidentifiable, and a single reference run will not cover every subset. We therefore use a small mixture of reference runs, combined by multiple importance sampling. The mixture includes a run with all calibrations, a run with every calibration widened by a tempering factor, and, where diagnostics call for it, one run seeded on the subset with the worst coverage. Any subset whose reweighted effective sample size falls below a threshold gets a dedicated run.
The value of a subset, $v(S)$, is a divergence between the node-age posterior of interest under $\pi_S$ and under the uncalibrated reference. We use the 1-Wasserstein distance on node age, and also the posterior mean age for readability. Shapley values $\phi_k$ and pairwise Shapley interactions $\phi_{jk}$ follow HyperSHAP's estimators. Strongly negative $\phi_{jk}$, meaning two calibrations that each shift the age but partly cancel when combined, is our operational definition of conflict. A calibration whose Shapley value is large at nodes far from its own position is flagged as distorting.
The collapse of reweighted effective sample size between $R$ and $S$ is itself informative. It measures how far adding a calibration moves the posterior out of its previous support. We report it alongside the Shapley values as a cheap second conflict signal.
Baselines
The naive reference is the current default: report the full-calibration posterior and compare each calibration's marginal prior against its marginal posterior. This is the "sample from the prior" check in BEAST 2 (BEAST 2.5). It runs on identical data and costs one extra prior-only run. It detects a calibration that is overridden, but it cannot say which other calibration overrode it.
The strongest practical competitor is leave-one-out calibration analysis, meaning $K+1$ full MCMC runs, each dropping one calibration. It is what careful dating studies already do, it needs no new machinery, and on small $K$ it may be enough. It gets the same data, the same chain length per run, and the same convergence criteria, so its compute budget is $K+1$ times ours. We also give it a matched-compute variant: shorter chains totalling the same CPU hours as our method.
The accuracy ceiling is brute-force exact Shapley, which runs MCMC under all $2^K$ subsets. It is feasible only for $K\le 6$. We use it as ground truth for estimator error, not as a practical method.
A surrogate variant of HyperSHAP (HyperSHAP) fits a regression model mapping subset indicators to $v(S)$ from a sampled set of subset runs, at a budget matched to ours. It tests whether our reweighting identity matters or whether a generic surrogate suffices. No reported result was found on Lacuna for any attribution method applied to calibration priors. HyperSHAP reports only qualitative gains over fANOVA, and no numeric convergence figures are given in the file.
Two ablations isolate the key ingredients. The first drops the interaction terms and keeps only first-order Shapley values. The second replaces the tempered-mixture reference with a single full-calibration reference.
Beating a baseline means one of two things on simulations with planted conflicts. The first is higher recall of the planted distorting calibration, and of the planted conflicting pair, at equal or lower compute. The second is a node-age estimate, after removing the flagged calibration, that lies closer to the true age than the estimate after removing the calibration flagged by the baseline.
Why not just…?
Why not just run leave-one-out over calibrations? This is the real threat. When calibrations act roughly additively, leave-one-out and Shapley values rank players identically, and leave-one-out is simpler to explain. Leave-one-out fails in two structured cases. One is redundancy: two consistent fossils mask each other, so dropping either changes little. The other is conflict: two wrong-but-opposed fossils cancel. These are exactly the interaction failures that fANOVA and HyperSHAP document for one-at-a-time analysis in configuration spaces. Our simulations plant both cases and also include purely additive controls. If leave-one-out matches us on the redundant and cancelling scenarios, we will report that the added machinery is unnecessary and fall back to presenting the reweighting estimator as a cheap way to compute leave-one-out itself.
Why not just compare each calibration's prior to its posterior? That comparison is local. It shows that node A's fossil was overridden, but not by which other anchor, and it cannot express conflict between calibrations at different nodes. The metric where this should show up is recall of the conflicting pair, which prior-versus-posterior checks cannot produce at all.
Why not just brute-force all subsets, or fit a surrogate over subsets? Brute force is exact but exponential, and it is the target our estimator must match. A learned surrogate reintroduces the fidelity problem that bounds surrogate-based transparency tools (Efficient Parameter Importance Analysis via Ablation with Surrogates). Reweighting avoids that problem because it needs no model of $v(S)$ at all. The matched-budget surrogate baseline tests whether this advantage is real.
Why not just avoid priors with a two-stage method such as RelTime? The a priori/a posteriori study found that post-hoc calibration consistently fails to recover the magnitude and direction of rate changes. Attributions computed on such timelines would explain the wrong answer.
Proposed Experiments
The first testbed extends the forward rate-shift simulations of the calibration-strategy study, with 100 replicates per scenario as in that paper. We vary $K\in\{3,5,8,12\}$ and plant one of four conditions: a single biased fossil (age shifted by 20–50%), a redundant consistent pair, an opposed conflicting pair, or no error. The metrics are:
- recall at rank one of the planted calibration or pair;
- rank correlation and absolute error of estimated Shapley values against brute-force values at $K\le6$;
- node-age error against the true age after removing the flagged calibration;
- total CPU hours, with reweighted effective sample size logged throughout.
Uncorrelated lognormal and autocorrelated clocks are both run, because the study shows they fail differently under rate shifts.
The second testbed is empirical. The first case is the multi-fossil Juglandaceae plastome analysis (Plastome evolution of Engelhardia), which has several fossil anchors on internal nodes. The second is a primate dating analysis in which a mutation-rate prior is itself a player. That player is set in turn to the fossil-implied, pedigree, and recombination-derived rates reported in the human mutation-rate paper. For this case the decisive check is external. The attribution should assign the largest share of the shift in the human–chimpanzee split away from roughly 6.8 million years ago to the rate player under the pedigree setting, and it should flag a negative interaction between that rate prior and deep fossil calibrations. If it instead implicates fossils at shallow nodes, the method is miscalibrated.
The decisive empirical check overall is the planted redundant and conflicting scenarios at $K=8$. Our estimator must match brute-force Shapley rankings (Spearman above 0.9 on the scenarios where brute force is feasible, extrapolated with spot checks) and beat leave-one-out on pair recall, while using no more than about four MCMC runs. If reweighted effective sample size collapses so often that most subsets need dedicated runs, the cost advantage disappears. We will report that as the method's boundary, quantified as a function of calibration tightness.
References
- HyperSHAP: Shapley Values and Interactions for Explaining Hyperparameter Optimization
- An Efficient Approach for Assessing Hyperparameter Importance
- Efficient Parameter Importance Analysis via Ablation with Surrogates
- Performance of A Priori and A Posteriori Calibration Strategies in Divergence Time Estimation
- PrioriTree: a utility for improving phylodynamic analyses in BEAST
- Calibrating the Human Mutation Rate via Ancestral Recombination Density in Diploid Genomes
- BEAST 2.5: An advanced software platform for Bayesian evolutionary analysis
- Plastome evolution of Engelhardia facilitates phylogeny of Juglandaceae