Lasso Cox Models For Omics Feature Selection

Why Add an L1 Penalty to a Cox Model?

In an omics study, the number of measured features can be large relative to the number of observed events. An ordinary Cox model may then be unstable or impossible to fit when the design matrix is rank-deficient. An L1 penalty offers one approach: it shrinks coefficient estimates and can set some exactly to zero. The resulting model is sparse, but sparsity is a modeling outcome, not proof that the retained features are biologically important or independently confirmed.

The earlier material on high-dimensional Cox regression establishes the need to address many candidate predictors relative to the event information. Here the focus is the specific L1-penalized fit, its tuning parameter, and the interpretation of its selected features.

Model, Objective, and Assumptions

For subject \(i\), let \(T_i\) be the observed follow-up time, \(\delta_i\) indicate whether the event was observed (\(\delta_i=1\)) or the observation was right-censored (\(\delta_i=0\)), and let \(x_{ij}\) be the value of feature \(j\). With \(p\) candidate features, write \(x_i=(x_{i1},\ldots,x_{ip})\) and \(\beta=(\beta_1,\ldots,\beta_p)\). The Cox proportional-hazards model specifies

where \(h_0(t)\) is an unspecified baseline hazard. A coefficient \(\beta_j\) is a conditional log-hazard association for a one-unit increase in feature \(j\), holding the other modeled features fixed. The proportional-hazards assumption means that, for two subjects with fixed covariates, their hazard ratio does not change over time.

For simplicity, first suppose event times are distinct. Let \(R(T_i)\) be the risk set just before time \(T_i\): the subjects still under observation and event-free at that time. The Cox partial log-likelihood is

The L1-penalized estimate minimizes the negative partial log-likelihood plus a penalty on the absolute coefficient sizes:

Here the loss is divided by \(n\), so the numerical scale of \(\lambda\) depends on that convention. Some software uses a different scaling; the selected fit should therefore be described along with its fitting and tuning convention. The penalty applies to the feature coefficients, not to a Cox intercept (there is no intercept in the partial likelihood) or to the unspecified baseline hazard.

The model requires appropriate risk sets and a suitable treatment of censoring. In particular, censoring should be independent of the event time conditional on the modeled covariates and any relevant design variables. The proportional-hazards form must be a reasonable working model for the effects being estimated. Subjects should also be independent unless the analysis explicitly accounts for clustering. If there are tied event times, the fitting procedure must use a specified tie method, such as Breslow or Efron; the examples below use distinct event times. A penalty does not make these assumptions unnecessary.

How the Penalty Produces Zero Coefficients

The L1 penalty has a corner at zero, which allows an optimum to occur exactly at zero. This can be seen from the first-order optimality condition. Let \(L(\beta)=-\ell(\beta)/n\). At a minimum, the derivative of the smooth loss and a subgradient of the penalty must balance. For a nonzero coefficient, this gives

At zero, the subgradient of \(|\beta_j|\) can be any value between \(-1\) and \(1\). Thus a coefficient can be zero at the optimum when

These conditions follow from the subgradient rule for the absolute value: its derivative is \(\operatorname{sign}(\beta_j)\) away from zero, and its subgradient at zero is the interval \([-1,1]\). Because the negative Cox partial log-likelihood is convex in \(\beta\), the penalized objective is convex too; a point satisfying the corresponding optimality conditions is a global minimizer. In a multivariable fit, however, each feature's derivative depends on the other coefficients, so selection is joint rather than a series of independent univariate tests.

Scaling and Coefficient Interpretation

The penalty acts on coefficient magnitudes, so measurement units matter. If one feature is recorded in very small units and another in large units, equal changes in their raw coefficients need not represent comparable shrinkage. A common approach is to center and scale each continuous feature using quantities estimated in the training data. The same training-set means and standard deviations are then applied to validation or test data. Any imputation, filtering, or other data-dependent preprocessing must likewise be learned within the training portion of the analysis.

After standardization, a coefficient is the conditional log-hazard ratio for a one-training-standard-deviation increase in that feature, with the other included features held fixed. Its hazard ratio is \(\exp(\hat{\beta}_j)\). A nonzero estimate describes the fitted penalized model; it is not automatically an unbiased estimate of an underlying biological effect. Shrinkage deliberately changes estimates, and selection makes ordinary post-fit uncertainty statements more complicated.

Tuning the Penalty

The data generally do not identify a single universally correct \(\lambda\). A typical strategy evaluates candidate values by cross-validation, fitting the model on each training fold and computing a prespecified survival loss on its held-out fold. For example, one may use held-out negative partial log-likelihood with risk sets defined consistently for that evaluation. The smallest mean cross-validated loss chooses the best-performing candidate under that criterion. A one-standard-error rule is another convention: choose the largest \(\lambda\) whose mean loss is within one estimated standard error of the minimum, often favoring a simpler model when performance is similar.

All data-dependent steps—including feature filtering, imputation, scaling, and the choice of \(\lambda\)—must be confined to the training data when estimating performance on held-out data. If the same data are used both to tune and to report final predictive performance, the reported performance can be optimistic. The earlier material on nested multi-omics evaluation addresses this separation of tuning from performance assessment.

Diagnostics, Comparisons, and Limits

Check more than the final list of nonzero coefficients. Inspect whether the optimization converged, whether the selected model and tuning choice are sensitive to reasonable analysis decisions, and whether the proportional-hazards assumption is plausible for the scientific question. Examine missingness, outliers, feature distributions, and the consequences of correlated measurements. Repeatedly selected features across resamples may be more stable than features appearing once, but stability alone does not prove a feature is associated in a new population.

The lasso is useful when a sparse model is desired. In contrast, an unpenalized Cox model does not shrink coefficients and may be unstable when there are many predictors or strong collinearity. Ridge-penalized Cox regression shrinks coefficients but generally does not set them exactly to zero, so it is aimed more at stable prediction than a short selected list. When omics features are strongly correlated, the lasso may select one feature from a group and omit others, with the chosen representative changing across samples. Such a pattern can reflect redundancy and instability rather than a uniquely important feature.

Selection is not confirmation. The lasso's coefficient estimates are shrunk, and the same data determine which features are retained and how their coefficients are estimated. Treating the selected features as if they had been specified in advance and then applying ordinary unadjusted tests or confidence intervals can overstate evidence. A selected feature is a candidate for validation; confirmatory claims require appropriate independent evidence and an inferential strategy designed for the selection process. The model estimates conditional associations under its assumptions, not causal effects.

• Common mistake: “The coefficient is nonzero, so this omics feature has been proved to affect survival.”
• Correction: A nonzero coefficient means the penalized fitting procedure retained the feature at the chosen tuning value. It is a conditional model association subject to shrinkage, selection variability, and the model assumptions—not proof of causation or a confirmed discovery.
• Common mistake: “A larger \(\lambda\) adds the same penalty amount to every coefficient.”
• Correction: Each feature contributes \(\lambda|\beta_j|\) to the objective. The penalty contribution depends both on \(\lambda\) and on the absolute coefficient, and the fitted coefficients change when \(\lambda\) changes.