Lasso Cox Regression

The model adds an absolute value penalty term (λ Σ|β|) to the partial likelihood. This geometric constraint forces many coefficients to exactly zero, effectively performing embedded feature selection. Unlike ridge regression (L2), which only shrinks coefficients, the L1 penalty produces a sparse solution where irrelevant biomarkers are completely removed from the model.

In genomic studies where p >> n (e.g., 20,000 genes for 300 patients), standard Cox regression fails due to non-identifiability. Lasso Cox regression thrives in this setting by:

• Selecting a small subset of prognostic genes
• Preventing overfitting through regularization
• Producing interpretable models that map directly to biological pathways

The regularization strength λ controls the degree of penalization. Optimal λ is typically chosen via cross-validation using the partial likelihood deviance or the concordance index (C-index). Common practice selects λ at the minimum cross-validated error (λ_min) or the most regularized model within one standard error (λ_1se) to favor sparsity.

Lasso Cox regression inherits the proportional hazards assumption from the standard Cox model—the hazard ratio between any two subjects is constant over time. Violations can be diagnosed using Schoenfeld residuals on the selected features. If the assumption fails, time-varying coefficients or stratified models may be required.

When predictors are highly correlated (e.g., co-expressed genes), Lasso tends to arbitrarily select one representative. Elastic net combines L1 and L2 penalties to encourage grouped selection, retaining correlated biomarkers together. This is critical in pathway-based analyses where multiple genes in the same biological cascade should be preserved.

Efficient implementations use cyclical coordinate descent algorithms, available in R packages like glmnet and Python's scikit-survival. These tools handle right-censored data natively and support cross-validation, coefficient extraction, and survival function estimation. For ultra-high dimensional data, screening rules like strong rules accelerate convergence.

The foundational semiparametric model that Lasso Cox extends. It quantifies covariate effects on the hazard rate without specifying the baseline hazard.

• Assumes proportional hazards over time
• Cannot natively handle p > n scenarios
• Lasso Cox adds an L1 penalty to its partial likelihood for automatic feature selection

A direct extension combining L1 (Lasso) and L2 (Ridge) penalties. Useful when Lasso alone is unstable.

• The L2 component handles groups of correlated predictors better
• Prevents Lasso from arbitrarily selecting one gene from a co-expressed cluster
• Controlled by the α mixing parameter (α=1 is pure Lasso)

Lasso Cox is a prime example of embedded feature selection for p >> n problems common in genomics.

• Contrast with filter methods (univariate screening) and wrapper methods (recursive feature elimination)
• The regularization path shows how coefficients enter the model as λ decreases
• Cross-validation selects the optimal λ penalty balancing bias and variance

A non-parametric ensemble alternative to penalized Cox models. No proportional hazards or linearity assumptions required.

• Naturally models high-order interactions and non-linear effects
• Provides variable importance measures as an alternative to coefficient magnitude
• Often used as a benchmark to validate Lasso Cox's linearity assumption

The primary diagnostic tool for validating the proportional hazards assumption underlying Lasso Cox.

• Plot residuals against time to detect time-varying effects
• A non-zero slope indicates a violation requiring time-varying covariates or stratification
• Essential before interpreting Lasso-selected coefficients as constant hazard ratios

The standard discrimination metric for evaluating Lasso Cox predictions. Measures ranking ability on censored data.

• Ranges from 0.5 (random) to 1.0 (perfect)
• Harrell's C-index compares permissible pairs; Uno's C-index weights by censoring distribution
• Often used with cross-validation to tune the Lasso penalty λ