Bayesian Shrinkage For Multi-Omics Survival Models

Why Shrinkage Needs an Uncertainty Account

A multi-omics Cox model can contain far more candidate predictors than observed events. Unregularized estimates can then be unstable or even fail to be finite, while correlated measurements can make it difficult to attribute an association to one feature rather than another. Shrinkage priors regularize the coefficients by expressing, before observing the outcomes, a preference for effects near zero. Bayesian inference carries that preference into a posterior distribution, so the result includes both regularized estimates and uncertainty about them.

The aim is not to declare a feature important merely because its posterior mean is nonzero. A coefficient may be pulled substantially toward zero, and correlated coefficients may remain uncertain even when their shared direction is better estimated. Interpretation therefore requires attention to the prior, the scale of each predictor, and the joint posterior.

Model and Prior Assumptions

For a partial-likelihood analysis, let \(T_i\) be a subject’s event time, let \(C_i\) be the censoring time, and observe the follow-up time and event indicator. The Cox partial likelihood compares each event subject with the subjects at risk at that event time. Its use for Bayesian coefficient inference treats that partial likelihood as the likelihood contribution for \(\beta\), then combines it with a prior. This is a partial-likelihood-based posterior for the coefficients; it does not, by itself, specify a posterior distribution for the baseline hazard.

The interpretation relies on correctly constructed risk sets and the proportional-hazards model being adequate for the chosen covariates and time period. It also relies on censoring being non-informative conditional on the variables in the model, and on independent subjects unless dependence is explicitly modelled. The earlier material on proportional-hazards diagnostics is relevant: a shrinkage prior does not repair a systematically time-varying effect. If a full posterior for survival probabilities or the baseline hazard is needed, a model for the baseline hazard must also be specified.

Before specifying coefficient priors, put continuous predictors on interpretable, comparable scales, commonly by centering and scaling each feature. Then a coefficient can be read as a log hazard ratio per one standard-deviation increase, rather than per arbitrary assay unit. Categorical predictors need an explicit coding convention. Priors should reflect that scale: a prior variance that is sensible for standardized predictors may be implausibly tight or wide for unscaled measurements.

A Gaussian prior gives smooth, continuous shrinkage and is computationally convenient. The Bayesian lasso has more prior mass near zero and heavier tails than a Gaussian with the same variance. Global-local priors can combine strong shrinkage of many small coefficients with less shrinkage of a few large ones. None of these priors automatically identifies a unique causal feature, and an independent prior does not make correlated predictors independently identifiable from the data.

How Gaussian Shrinkage Changes an Estimate

A useful derivation comes from a normal likelihood approximation to the Cox partial likelihood. Suppose the likelihood for one coefficient is approximately \(N(\hat\beta,V)\), and the prior is \(N(0,\tau^2)\). Multiplying the likelihood kernel and prior kernel gives a normal posterior. The quadratic terms in \(\beta\) are \( -\frac{1}{2}\{(\beta-\hat\beta)^2/V+\beta^2/\tau^2\}\). Completing the square yields the posterior variance and mean:

The factor \(\tau^2/(\tau^2+V)\) lies between zero and one. Thus the posterior mean is the likelihood estimate multiplied by a shrinkage factor: smaller prior variance means stronger movement toward zero, while greater likelihood uncertainty \(V\) also makes the prior more influential. The posterior variance is smaller than either variance alone. In an actual high-dimensional Cox model the likelihood is generally not exactly normal, so these calculations are exact for the stated normal likelihood and otherwise serve as a local approximation or illustration.

Correlated Features and Joint Posterior Uncertainty

With correlated omics features, interpreting coefficients one at a time can conceal what the data identify. The partial likelihood may constrain a shared direction of effects well but provide little information about a contrast between nearly redundant measurements. The posterior then reflects both the likelihood’s geometry and the prior’s behavior in those directions. Even independent priors on the original coefficients do not ensure independent posterior coefficients.

A transparent way to see this is to rotate two coefficients into orthogonal coordinates. Define \(z_+=(\beta_1+\beta_2)/\sqrt{2}\), the common direction, and \(z_-=(\beta_1-\beta_2)/\sqrt{2}\), the contrast. If the likelihood is approximately normal in these coordinates, independent normal priors and likelihoods give independent normal posteriors for \(z_+\) and \(z_-\). Transforming back can produce correlated posteriors for \(\beta_1\) and \(\beta_2\).

Global-Local Shrinkage in Practice

Global-local priors are attractive when many coefficients are expected to be small but a few may be substantial. The global scale \(\tau\) controls overall shrinkage; each \(\lambda_j\) adjusts shrinkage locally. Posterior summaries must account for uncertainty in these scales rather than treating a fitted shrinkage factor as known. Priors on \(\tau\) matter especially in high dimensions, where the data may not contain enough events to determine the overall scale precisely.

Diagnostics, Comparisons, and Limitations

A shrinkage prior is part of the model, not a substitute for checking it. Examine convergence and mixing of posterior computation, including effective sample sizes and trace behavior for coefficients and scale parameters. Inspect whether results change materially under scientifically defensible alternatives for the global-scale prior or feature scaling. Prior predictive checks can reveal whether the implied hazard ratios are implausibly extreme or implausibly close to one before seeing the outcomes. For correlated predictors, inspect joint posterior dependence and avoid interpreting marginal intervals as if every coefficient were estimated independently.

Also examine the survival-model assumptions: risk-set construction, influential observations, and proportional-hazards adequacy. A concentrated posterior can still be misleading if the outcome model is misspecified. In very high dimensions, limited event information can leave the global shrinkage level weakly identified. A proper prior can regularize unstable likelihood estimates, but it cannot create information absent from the event data, correct informative censoring, or establish causal effects.

Bayesian shrinkage differs from common alternatives in what it reports. Penalized Cox methods such as ridge, lasso, and elastic net typically provide a penalized point estimate selected through a tuning procedure; Bayesian priors express regularization probabilistically and yield posterior distributions, conditional on the model and prior. A Gaussian prior corresponds to ridge-like shrinkage in its penalty form, but the posterior is not just the penalized estimate. Global-local priors allow more coefficient-specific adaptation than a shared Gaussian penalty. Feature selection procedures instead aim to produce a sparse selected set; continuous shrinkage generally leaves coefficients nonzero and should not be described as definitive selection.