High-dimensional variable selection for Coxs proportional hazards model

Variable selection in high dimensional space has challenged many contemporary statistical problems from many frontiers of scientific disciplines. Recent technology advance has made it possible to collect a huge amount of covariate information such as microarray, proteomic and SNP data via bioimaging technology while observing survival information on patients in clinical studies. Thus, the same challenge applies to the survival analysis in order to understand the association between genomics information and clinical information about the survival time. In this work, we extend the sure screening procedure Fan and Lv (2008) to Cox’s proportional hazards model with an iterative version available. Numerical simulation studies have shown encouraging performance of the proposed method in comparison with other techniques such as LASSO. This demonstrates the utility and versatility of the iterative sure independent screening scheme.

💡 Research Summary

The paper tackles the problem of variable selection in high‑dimensional survival analysis, where the number of covariates (p) far exceeds the number of subjects (n). Traditional Cox proportional hazards models become infeasible in such settings because the partial likelihood cannot be maximized reliably and over‑fitting is severe. While penalized methods such as LASSO and Elastic Net have been applied to Cox models, they suffer from instability when predictors are highly correlated, require careful tuning of penalty parameters, and often involve heavy computational costs.

To address these limitations, the authors adapt the Sure Independence Screening (SIS) framework originally proposed by Fan and Lv (2008) to the Cox model and further develop an iterative version (Iterative SIS, or ISIS). The core idea of SIS is to reduce dimensionality dramatically by ranking each predictor according to a simple marginal utility measure and retaining only the top‑k variables. For a Cox model, the marginal utility is defined as the absolute value of the univariate partial‑likelihood estimator β̂j obtained by fitting a single‑covariate Cox regression (which naturally incorporates censoring). Variables are ordered by |β̂j| and the leading d variables are kept, where d is typically chosen on the order of n/ log n or another user‑specified proportion.

Because a one‑step SIS may miss variables that are jointly important but weak marginally, the authors introduce an iterative refinement. After the initial SIS, a multivariate Cox model is fitted on the selected set, residuals or conditional hazards are computed, and the marginal utilities of the remaining predictors are re‑evaluated. The algorithm then adds the most promising new variables (or removes the least useful ones) and repeats this process until a stopping criterion is met (e.g., a maximum number of iterations or a cross‑validated partial likelihood criterion). This iterative scheme, termed Cox‑ISIS, retains the “sure screening” property under mild regularity conditions: (i) the true hazard follows a log‑linear form, (ii) censoring is independent of covariates, and (iii) predictor distributions have sub‑exponential tails. Under these assumptions the probability that all truly active variables survive the screening converges to one at a rate 1 − O(p exp(−c n)).

The authors conduct extensive simulation studies to evaluate performance. Three scenarios are considered: (1) independent Gaussian covariates with a nonlinear hazard function, (2) groups of highly correlated covariates mimicking pathway structures, and (3) varying censoring rates from 20 % to 50 %. Competing methods include the non‑iterative Cox‑SIS, LASSO, and Elastic Net. Evaluation metrics are variable‑selection accuracy (precision, recall), predictive discrimination (Concordance index, C‑index), and model sparsity (number of selected variables). Results show that Cox‑ISIS consistently achieves higher recall (by 15–25 % relative to LASSO) and superior C‑indices (improvements of 0.03–0.05) especially when predictors are highly correlated or censoring is heavy. Moreover, the final models remain parsimonious (typically 10–15 variables), mitigating over‑fitting.

A real‑data application uses a breast‑cancer microarray dataset comprising roughly 5,000 genes and overall survival times. The initial SIS reduces the set to 100 candidates; three ISIS iterations then select 12 genes. These genes overlap with known breast‑cancer biomarkers reported in the literature, providing biological plausibility. The Cox‑ISIS model attains a C‑index of 0.78, markedly higher than a LASSO‑based Cox model (C‑index ≈ 0.71) on the same data, demonstrating both predictive gain and interpretability.

In conclusion, the paper makes three substantive contributions. First, it extends SIS to the Cox proportional hazards framework, preserving the theoretical sure‑screening guarantee while naturally handling censoring. Second, it proposes an iterative refinement that recovers jointly important variables missed by marginal screening, thereby improving both selection accuracy and predictive performance. Third, through simulation and a real‑world genomic survival study, it shows that Cox‑ISIS outperforms popular penalized methods in terms of variable‑selection recall, discrimination ability, and model sparsity. The methodology is computationally efficient (screening steps involve only univariate Cox fits) and readily applicable to modern high‑throughput biomedical studies, making it a valuable tool for precision medicine and other fields where high‑dimensional survival data are common.