Bayesian Hierarchical Models for Multi-Omic Effects

Why Use a Hierarchical Model?

A genetic variant may have an estimated effect on expression, protein abundance, or another molecular phenotype in several tissues or molecular layers. Those estimates may differ because the underlying effects differ, because sampling error is substantial, or both. Estimating every effect separately can leave noisy estimates extreme; forcing all effects to be identical can hide genuine differences.

A Bayesian hierarchical model represents both possibilities. It gives each layer or tissue its own effect, while describing those effects as related draws from a shared distribution. This permits partial pooling: estimates are informed by their own data and, through the hierarchy, by information about the other effects. The amount of pooling depends on the assumed or estimated between-layer variation.

Definition: A Bayesian hierarchical model specifies a likelihood for observed data and a probability model for parameters. In a multi-omic setting, a common hierarchy gives each layer or tissue a distinct genetic effect \(\theta_j\), with those effects distributed around a shared mean \(\mu\) and with between-layer standard deviation \(\tau\). Partial pooling means that estimates are neither treated as fully unrelated nor constrained to be identical.

A Normal Hierarchical Model

Suppose \(y_j\) is an estimated genetic effect for layer or tissue \(j\), and \(s_j\) is its standard error. For a clearly specified variant and aligned effect allele, a useful summary-data model is:

y_j \mid \theta_j \sim N(\theta_j,s_j^2), \qquad \theta_j \mid \mu,\tau \sim N(\mu,\tau^2), \qquad j=1,\ldots,J.

Here \(N(a,b)\) denotes a normal distribution with mean \(a\) and variance \(b\). The first line treats each estimate as a noisy measurement of its layer-specific effect. The second says that the effects vary around the shared mean \(\mu\); \(\tau\) describes the amount of variation across layers. A small \(\tau\) implies similar effects and stronger pooling, whereas a large \(\tau\) allows more heterogeneity.

A complete Bayesian model also assigns prior distributions to the shared parameters. For example, \(\mu \sim N(0,\sigma_\mu^2)\) can express prior knowledge about the plausible scale of the shared effect, while a half-normal or half-\(t\) prior on \(\tau\) restricts it to be nonnegative and allows uncertainty about heterogeneity. The prior scales must be stated in units of the effect. They should be scientifically defensible, not selected simply to obtain a desired posterior result.

The normal likelihood assumes that the reported estimates are approximately normally distributed around their true effects and that the standard errors are adequate. Conditional independence of the \(y_j\) values is also an assumption if the likelihood is written as a product across layers. When estimates share participants or otherwise have correlated errors, the likelihood should represent their covariance rather than treating them as independent. Effects must also be comparable: allele orientation and effect units need to be aligned before a shared distribution is meaningful.

What Partial Pooling Does

The pooling mechanism is easiest to see when \(\mu\) and \(\tau\) are temporarily treated as fixed. Combining the normal likelihood for \(y_j\) with the normal distribution for \(\theta_j\) gives a normal conditional posterior for that layer’s effect:

\theta_j \mid y_j,\mu,\tau \sim N(m_j,V_j), \qquad V_j=\left(\frac{1}{s_j^2}+\frac{1}{\tau^2}\right)^{-1}, \qquad m_j=V_j\left(\frac{y_j}{s_j^2}+\frac{\mu}{\tau^2}\right).

Equivalently, the posterior mean is a weighted average of the layer estimate and the shared mean:

m_j= \frac{\tau^2}{\tau^2+s_j^2}y_j + \frac{s_j^2}{\tau^2+s_j^2}\mu.

The weight on \(y_j\) is larger when its standard error is small relative to the between-layer standard deviation. When \(\tau\) is small, the effects are expected to be similar, and noisy estimates are pulled more strongly toward \(\mu\). When \(\tau\) is large, the hierarchy permits greater differences, so each estimate retains more of its own value.

Interpretation: Partial pooling is not a rule that all molecular effects must be equal. It is a model-based compromise: the strength of borrowing depends on sampling uncertainty and on the modeled variation among effects. If \(\mu\) and \(\tau\) are uncertain, their posterior uncertainty must also be included when summarizing each \(\theta_j\).

Worked Example: Borrowing Across Two Molecular Layers

Consider hypothetical estimates for a variant’s effect on two standardized molecular phenotypes, measured in standard-deviation units per additional effect allele. Suppose \(y_1=0.20\), \(y_2=0.05\), and both standard errors are \(0.10\). Use the model above with \(\tau=0.10\) and \(\mu \sim N(0,0.20^2)\). These fixed values are illustrative model assumptions, not estimates from a real study.

First find the posterior distribution for \(\mu\). Conditional on \(\mu\), each \(y_j\) has variance \(\tau^2+s_j^2=0.01+0.01=0.02\). The prior variance of \(\mu\) is \(0.04\). The posterior precision is therefore:

\frac{1}{0.04}+\frac{2}{0.02}=25+100=125.

Thus the posterior variance is \(1/125=0.008\), and the posterior mean is:

E(\mu\mid y_1,y_2) = \frac{y_1/0.02+y_2/0.02}{125} = \frac{0.20/0.02+0.05/0.02}{125} = \frac{12.5}{125} = 0.100.

Conditional on \(\mu\), the posterior variance of either \(\theta_j\) is \((1/0.01+1/0.01)^{-1}=0.005\), and its posterior mean is \((y_j+\mu)/2\). Averaging over the posterior for \(\mu\) gives posterior means \(0.5(0.20+0.10)=0.150\) and \(0.5(0.05+0.10)=0.075\). Each marginal posterior variance is \(0.005+0.25(0.008)=0.007\), so its posterior standard deviation is \(\sqrt{0.007}=0.0837\).

The first layer’s estimate moves down from \(0.20\), while the second moves up from \(0.05\); both move toward the shared mean. Under this model, approximate 95% posterior credible intervals are \(0.150\pm1.96(0.0837)=(-0.014,0.314)\) and \(0.075\pm1.96(0.0837)=(-0.089,0.239)\), respectively. Conditional on the model and the data, each interval contains 95% posterior probability for its effect. These are not statements that a fixed effect has a 95% frequentist probability of lying in the interval.

Weak and Conflicting Evidence

Hierarchical estimates are especially useful to examine when layer-specific data are noisy or disagree. The hierarchy does not erase such disagreement. Instead, it balances each estimate against the shared distribution. The next example keeps the same assumptions as the previous one so that the effect of conflicting evidence is visible.

Worked Example: Opposing Layer Estimates

Suppose hypothetical estimates are \(y_1=0.30\) and \(y_2=-0.10\), each with standard error \(0.10\), in the same standardized effect units. Use \(\tau=0.10\) and the prior \(\mu \sim N(0,0.20^2)\).

As before, the posterior variance for \(\mu\) is \(0.008\). Its posterior mean is:

E(\mu\mid y_1,y_2) = \frac{0.30/0.02+(-0.10)/0.02}{125} = \frac{15-5}{125} = 0.080.

The posterior means for the two layer-specific effects are \(0.5(0.30+0.08)=0.190\) and \(0.5(-0.10+0.08)=-0.010\). The marginal posterior standard deviation remains \(\sqrt{0.007}=0.0837\). Approximate 95% posterior credible intervals are therefore \(0.190\pm0.164=(0.026,0.354)\) and \(-0.010\pm0.164=(-0.174,0.154)\).

The estimates remain different in direction after pooling, but the second posterior is centered close to zero and is uncertain about its sign. This does not establish that the two true effects are equal or opposite. It shows what follows from the chosen likelihood, prior, and heterogeneity assumption; a different estimate of \(\tau\) could produce a different degree of shrinkage.

A weak estimate may also be sensitive to the prior scale. In a simple normal model, if \(y\mid\theta\sim N(\theta,s^2)\) and \(\theta\sim N(0,\sigma^2)\), the posterior mean and variance are:

E(\theta\mid y)= \frac{\sigma^2}{\sigma^2+s^2}y, \qquad \operatorname{Var}(\theta\mid y)= \left(\frac{1}{\sigma^2}+\frac{1}{s^2}\right)^{-1}.

This is the same shrinkage principle, here illustrating the role of a prior centered at zero. It is useful for understanding prior sensitivity, although a multi-layer analysis should use its full hierarchical structure rather than silently substituting this single-effect model.

Worked Example: Sensitivity to a Prior Scale With Weak Evidence

Suppose a hypothetical layer-specific effect estimate is \(y=0.12\) standard-deviation units per effect allele, with standard error \(s=0.15\). Compare two zero-centered normal priors, \(N(0,0.10^2)\) and \(N(0,0.50^2)\). This comparison isolates how prior scale can matter when the estimate is imprecise.

For the narrower prior, the posterior variance and mean are:

V=\left(\frac{1}{0.10^2}+\frac{1}{0.15^2}\right)^{-1} =\left(100+44.444\right)^{-1} =0.006923, \qquad m=V\frac{0.12}{0.15^2} =0.006923(5.333) =0.0369.

The posterior standard deviation is \(\sqrt{0.006923}=0.0832\), giving an approximate 95% posterior credible interval \(0.0369\pm1.96(0.0832)=(-0.126,0.200)\).

For the broader prior, the posterior variance is \((4+44.444)^{-1}=0.02064\), and the posterior mean is \(0.02064(5.333)=0.110\). The posterior standard deviation is \(0.1437\), giving an approximate interval \(0.110\pm1.96(0.1437)=(-0.172,0.392)\). The narrower prior pulls the estimate closer to zero and gives a narrower interval; the broader prior leaves the mean closer to the observed estimate but has greater posterior uncertainty. The data alone do not decide which prior is appropriate.

Choosing and Checking the Prior Structure

A prior should encode a defensible structure for effects, not just produce convenient shrinkage. The normal hierarchy is attractive when effects are plausibly exchangeable around a shared mean and moderate deviations are reasonable. In some applications, a shared mean of zero is scientifically appropriate; in others, the prior center should be estimated or informed by a defensible external scale. The prior for \(\tau\) is consequential: concentrating it near zero can force strong pooling, while allowing large values can leave effects close to their separate estimates.

If there is a scientific reason to expect different groups of layers to have different typical effects, the hierarchy can include group-specific means. If effects are expected to have different variances, those differences need their own model rather than being hidden in one common \(\tau\). More elaborate covariance structures can represent correlated effects, but they require enough information to estimate the additional parameters and a clear rationale for the assumed dependence.

Before interpreting posterior results, check that the model matches the analysis scale, that the effect alleles and signs are aligned, and that the likelihood reflects the uncertainty and dependence in the estimates. Examine sensitivity to reasonable prior choices and to the heterogeneity structure. Posterior estimates describe effects under the specified statistical model; they do not by themselves show that a genetic effect is causal, that one molecular layer mediates another, or that a particular biological mechanism has been established.

Common Mistakes and Why the Model Matters

Assuming partial pooling makes all effects equal. It does not. Each \(\theta_j\) remains layer-specific; \(\tau\) controls how much the model expects them to differ.
Treating the prior as irrelevant. With weak evidence, as the third example shows, changing a plausible prior scale can meaningfully change posterior estimates and intervals. Report and examine that sensitivity.
Pooling effects on incompatible scales. Effects with different allele orientations or units do not have a coherent shared mean until they are aligned or explicitly modeled on a comparable scale.
Calling a credible interval a confidence interval. A 95% Bayesian credible interval has 95% posterior probability under the model and data; it is not a claim that the fixed parameter has a 95% probability of lying in that particular interval.
Reading shrinkage as evidence of no effect. Movement toward the shared mean follows from the model and the evidence. It is not, on its own, proof that a layer-specific genetic effect is absent.

Key takeaway: A hierarchical model provides a principled way to share information across molecular layers while retaining layer-specific effects. Its conclusions depend on the likelihood, the exchangeability assumption, and the prior structure—especially the model for between-layer variation.

Check Your Understanding

Use the model definitions and calculations above to answer these questions.

In the normal hierarchy, what does \(\tau\) represent, and what happens to pooling as \(\tau\) becomes small relative to \(s_j\)?
Why does a layer with a smaller standard error generally retain more of its own estimate?
What assumption is violated if correlated layer estimates are entered into a likelihood that treats them as conditionally independent?
In the opposing-estimates example, why does the posterior mean for the negative estimate move toward zero without demonstrating that its true effect is exactly zero?
How should a 95% posterior credible interval be interpreted, and why is it not a 95% frequentist confidence statement about a fixed parameter?
Give one reason to examine sensitivity to the prior scale or to the prior for between-layer variation.