Bayesian Borrowing Across Species Without Naive Pooling

Why Borrow Across Species?

Animal studies can inform human PK/PD predictions, but treating all species as if they had the same parameter is usually too strong. Conversely, fitting every species entirely independently can discard useful information, especially when data are sparse. Bayesian hierarchical models offer a middle ground: they allow species to differ while estimating how much those differences tend to vary.

The essential distinction is between a systematic prediction for a species and its remaining deviation from that prediction. For a positive parameter such as clearance, it is often convenient to model the logarithm of the parameter. Let \(y_s\) denote an estimated log parameter for animal species \(s\), and let \(m_s\) be the log-scale prediction from a specified scaling or mechanistic model. The residual deviation is then defined relative to \(m_s\), not relative to zero by default. This separation lets the model represent expected differences, such as those associated with body size, while borrowing information about deviations that may be shared across species.

A Hierarchical Model for Species Differences

A basic model for one transformed PK or PD parameter is

Here \(y_s\) is the observed estimate on the log scale, \(v_s\) is its sampling variance, and \(\theta_s\) is the latent species-specific log parameter. The quantity \(m_s\) is a specified log-scale prediction, \(\mu\) is the average residual across the modeled species, and \(\tau\) is the between-species residual standard deviation. The observation variance \(v_s\) represents uncertainty in estimating a species parameter; \(\tau^2\) represents variation in the underlying parameters across species. They are distinct sources of uncertainty.

A Bayesian analysis assigns distributions to unknown shared quantities, for example \(\mu\sim N(\mu_0,A^2)\) and a prior distribution for \(\tau\) on nonnegative values. The prior for \(\tau\) matters: when data are sparse, it can materially affect how much species-specific estimates are drawn toward the shared distribution. A model with \(\tau=0\) forces all residuals to equal \(\mu\); a model with very large \(\tau\) allows substantial differences and consequently transfers little information between species.

The basic model assumes that, after accounting for \(m_s\), species residuals are exchangeable: their labels do not provide further information about their residual distributions. It also assumes the stated observation model is adequate and that the standard errors or sampling variances \(v_s\) are reasonably characterized. Exchangeability is a substantive assumption, not a mathematical convenience. If a mechanism or known species feature predicts residual differences, it should be represented in the systematic component rather than ignored. For example, a covariate \(x_s\) could enter as \(\delta_s=x_s\beta+u_s\), with \(u_s\) drawn from a shared residual distribution.

How One Species Is Updated

Suppose the shared mean \(\mu\) and between-species standard deviation \(\tau\) are fixed for the moment. Define the residual-adjusted observation \(r_s=y_s-m_s\). From the observation model, \(r_s\mid\delta_s\sim N(\delta_s,v_s)\). Combining that likelihood with the normal distribution for \(\delta_s\) gives a normal posterior:

The posterior mean is a precision-weighted compromise between the observed residual \(r_s\) and the shared mean \(\mu\). If \(v_s\) is small, the species estimate is precise and receives greater weight. If \(\tau^2\) is small, the model regards species residuals as similar and gives greater weight to \(\mu\). The expression uses \(r_s=y_s-m_s\); using \(y_s\) in its place would be wrong unless \(m_s=0\) or the observation had already been residual-adjusted.

Learning the Shared Residual Distribution

For each species, the residual-adjusted observation satisfies

This follows because \(r_s=\delta_s+\epsilon_s\), where \(\delta_s\sim N(\mu,\tau^2)\) and the sampling error \(\epsilon_s\sim N(0,v_s)\) is independent of \(\delta_s\). Thus, the variance of the residual-adjusted observation includes both between-species variation and within-species sampling error. Omitting \(v_s\) would treat noisy species estimates as if they were exact and can exaggerate the inferred between-species variability.

A useful conjugate calculation is available when \(\tau\) is fixed and the prior for \(\mu\) is normal. If \(\mu\sim N(\mu_0,A^2)\), then

Each species contributes according to its precision \(1/(\tau^2+v_s)\), not simply according to its sample size or its raw estimate. In a full model, \(\tau\) is also unknown; its posterior is generally not given by this fixed-\(\tau\) formula and can be computed using numerical Bayesian methods. The conjugate result nevertheless makes the source of borrowing transparent.

Predicting a Human Parameter

For a human target with systematic prediction \(m_H\), introduce a human residual \(\delta_H\) from the same population distribution, conditional on the model’s exchangeability assumption. The latent human parameter is \(\theta_H=m_H+\delta_H\). Given \(\mu\) and \(\tau\), the human residual is \(N(\mu,\tau^2)\); integrating over their posterior distributions gives the human posterior predictive distribution. This predictive distribution is for a new species-level parameter, so it includes between-species variation as well as uncertainty in the shared quantities. It is not merely the posterior distribution of \(\mu\).

In the special case where \(\tau\) is fixed and the posterior for \(\mu\) is \(N(a,B)\), the predictive distribution has a simple form. Since \(\delta_H=\mu+u_H\), with \(u_H\sim N(0,\tau^2)\) independent of \(\mu\),

For a log-transformed positive parameter, exponentiating a credible interval on the log scale gives an interval on the original scale. The exponential of the posterior mean on the log scale is the predictive median on the original scale under this normal model; it is not the predictive mean.

Prior Assumptions, Diagnostics, and Alternatives

A hierarchical model does not remove assumptions; it makes them explicit. The choice of prior for \(\tau\) controls how readily the model permits large residual differences. With few species, the data may contain little information about \(\tau\), so human predictions can depend appreciably on that prior. The prior for \(\mu\) can also matter when the observed residuals are sparse or imprecise. These are reasons to examine whether plausible prior choices lead to meaningfully different human predictions, rather than reasons to select a prior solely because it produces a narrow interval.

Several checks help identify trouble. Compare species residuals with their sampling uncertainty: large systematic departures from the assumed residual distribution may indicate an inadequate scaling prediction or non-exchangeability. Check whether posterior predictions can reproduce the observed spread of residual-adjusted estimates, including their uncertainty. Examine posterior distributions for \(\mu\) and \(\tau\); a broad or strongly prior-sensitive distribution for \(\tau\) signals that the data do not determine the degree of borrowing precisely. A model can also be numerically well behaved yet scientifically unsuitable if an important species difference is omitted from \(m_s\).

The main alternatives make different assumptions. No borrowing estimates each species separately; it avoids an exchangeability assumption but can leave estimates highly uncertain when a species has limited data. Complete pooling forces all species residuals to be identical, effectively setting \(\tau=0\); it can give deceptively precise predictions when genuine differences exist. Hierarchical partial pooling lies between these extremes, with the amount of borrowing determined by the data and prior. A model with species-level covariates can represent structured differences, but does not make unmeasured differences disappear. These approaches are not interchangeable: the appropriate choice depends on biological knowledge, data quantity, and the purpose of the prediction.

Bayesian borrowing is most valuable when it reflects a defensible scientific claim: after known systematic differences are accounted for, the remaining species deviations are related enough to inform one another. It cannot establish that claim from computation alone. The human prediction should therefore be read as conditional on the chosen transformation, scaling structure, exchangeability assumption, observation model, and prior distributions. Making those conditions visible is part of the result, not an optional qualification.