Abstract
Feature screening in ultrahigh-dimensional settings is a critical step for identifying relevant predictors when the number of variables significantly exceeds the number of observations. However, traditional screening methods often struggle with censored data, non-linear relationships, and outliers. This paper introduces Spearman Rank Correlation Screening for Censored Data ($SRCS_{cen}$), a model-free approach that leverages rank-based correlation and a local Kaplan-Meier imputation technique. Our results demonstrate that $SRCS_{cen}$ maintains high coverage probability and rank consistency even under extreme censoring rates (up to 65%) and heavy-tailed distributions, significantly outperforming existing model-dependent and model-free alternatives.
1. Introduction
In fields such as genomics and financial risk analysis, researchers frequently encounter "ultrahigh-dimensional" datasets where the number of potential predictors ($p$) is vastly larger than the sample size ($n$). Feature screening serves as a preliminary filter to reduce the dimensionality to a manageable scale. This task is complicated by censoring, which occurs when the exact value of a response variable (e.g., survival time) is only partially observed.
Existing screening methods typically face three primary challenges:
- Model Dependency: Many methods assume a specific functional form (e.g., linear or Cox models), leading to failure if the true relationship is non-linear.
- Sensitivity to Outliers: Standard correlation measures can be skewed by extreme values or heavy-tailed distributions.
- Censoring Fragility: Performance often degrades sharply as the censoring rate (CR) increases.
2. Methodology
The proposed $SRCS_{cen}$ method addresses these limitations through two core innovations:
2.1 Spearman Rank Correlation
By utilizing Spearman rank correlation rather than Pearson correlation, the method becomes "model-free." It is invariant to monotonic transformations of the data, meaning it can identify relationships regardless of whether they are linear or non-linear (e.g., logarithmic). Because it relies on the relative order of data points, it is inherently robust against outliers.
2.2 Imputation for Censoring
To handle censored observations, the authors employ a local Kaplan-Meier estimator to "impute" the conditional expectation of the response variable. This allows the calculation of a robust correlation index ($\omega_k$) for every variable by incorporating information from censored data that other methods might discard. The method establishes "Sure-Screening" properties, ensuring the probability of including all true active variables approaches 1 as the sample size grows.
3. Experimental Results
3.1 Simulation Performance
The authors evaluated $SRCS_{cen}$ against several competitors, including $CQS_{cen}$ and $RCS_{cen}$, across various scenarios. As shown in Figure 1, $SRCS_{cen}$ exhibits superior performance, particularly at high censoring rates. In a linear model simulation with a 65% censoring rate, $SRCS_{cen}$ achieved a Coverage Probability (CP) of 86.5%, while the next best competitor ($CQS_{cen}$) reached only 60%, and others fell as low as 11%.
3.2 Robustness Across Distributions
The robustness of the method was further tested across different error distributions, including Normal, Logistic, and Weibull. As detailed in Figure 2, $SRCS_{cen}$ consistently identified the true active variables (indicated by the $P_{all}$ and $m_{10}$ metrics) more reliably than competing methods across all tested distributions and censoring levels.
3.3 Real-World Application: Lymphoma Data
When applied to Mantle Cell Lymphoma microarray data (92 patients, 6,312 genes), $SRCS_{cen}$ identified 20 key genes. Notably, it successfully identified the cell division cycle 2 gene (Gene 28990), a well-documented marker in cancer progression, demonstrating its practical utility in biological discovery.
4. Discussion and Conclusion
$SRCS_{cen}$ represents a significant advancement in ultrahigh-dimensional feature screening. Its primary strength lies in its robustness; by relying on ranks rather than absolute values, it remains immune to outliers and heavy-tailed noise. Furthermore, the imputation strategy effectively extracts signal from highly censored datasets where traditional methods fail.
Future work may address the computational complexity of the local Kaplan-Meier estimation for datasets with millions of variables and explore the relaxation of the independent censoring assumption. Nevertheless, $SRCS_{cen}$ provides a powerful, model-free tool for high-dimensional survival analysis and related fields.
Summary updated