TL;DR
This paper introduces a novel hierarchical False Discovery Rate (FDR) control method for high-dimensional survival analysis, specifically designed to handle interactions between low-dimensional and high-dimensional covariates. This approach is crucial for accurately identifying significant genetic and environmental interactions in complex biological systems while respecting the "main effects first, then interactions" hierarchy, a common biological principle.
What Problem Does This Paper Address?
Analyzing survival data with a massive number of potential predictors (high-dimensional) is a common challenge in fields like genomics and epidemiology. A key aspect of this analysis is understanding how different factors interact. For instance, how does a specific gene (high-dimensional) interact with an environmental exposure (low-dimensional) to influence survival time? Existing methods for high-dimensional inference often struggle with these interaction models, particularly when trying to control for false discoveries (incorrectly identifying a non-significant factor as significant) while respecting a biological hierarchy where main effects are considered before interactions. This paper aims to bridge this gap by providing a statistically rigorous and practically useful method for this complex scenario.
Methods
The authors propose a two-pronged approach:
- Modeling: They use the Accelerated Failure Time (AFT) model for survival analysis, which offers simpler interpretation than the Cox model in high-dimensional settings.
- Estimation and Selection: A "weighted least squares + debiased Lasso" technique is employed for estimating model parameters and selecting important variables. The debiased Lasso is key to obtaining reliable estimates for inference.
- Hierarchical FDR Control: A novel hierarchical FDR control procedure is developed. This procedure first tests the high-dimensional main effects and then, only if a main effect is deemed significant, proceeds to test its associated interactions. This ensures that the "main effects, interactions" hierarchy is respected, preventing spurious interaction findings. The asymptotic properties of the debiased Lasso estimators are rigorously established, providing theoretical guarantees for the FDR control.
Datasets
The authors evaluate their proposed method using simulated data generated from model (1) under various scenarios (sample size, censoring rate, sparsity, dimension, correlation, signal magnitude). They also apply their method to a real-world breast cancer dataset from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC). This dataset includes gene expression data (high-dimensional) and clinical variables like Tumor Mutation Burden (TMB), age, and Estrogen Receptor (ER) status (low-dimensional).
Key Insights and Ideas
- Hierarchy Matters: The "main effects, interactions" hierarchy is not just a theoretical construct but a crucial principle for biologically meaningful discoveries in gene-environment interaction studies. Ignoring it can lead to misleading results.
- Debiased Lasso for Robust Inference: Standard Lasso can be biased, making inference unreliable. The debiased Lasso approach is essential for obtaining valid statistical tests and FDR control in high-dimensional settings.
- AFT Model Advantage: The AFT model provides a more interpretable framework for survival analysis in the presence of many covariates compared to the traditional Cox model.
- Practical Utility: The method demonstrates practical utility by identifying genes and interactions relevant to breast cancer survival, aligning with existing biological knowledge.
Key Results
- FDR Control: The proposed hierarchical FDR control method successfully controls the FDR at the pre-specified level across various simulation settings, outperforming several existing methods, especially under challenging conditions like high dimensionality or high censoring rates.
- Power: The method achieves competitive empirical power, meaning it effectively identifies true signals while maintaining FDR control.
- Breast Cancer Data Analysis: The analysis of the breast cancer dataset identified 10 genes and 3 interactions that are potentially important for breast cancer survival. These findings are supported by existing literature, highlighting the method's practical relevance.
- Performance Comparison: Simulations show that the proposed method is robust and performs well compared to alternatives like BH, BH-Hierarchy, Surv-FCD, VS-D-Lasso, VS-Lasso, and VS-MCP, particularly in controlling FDR under complex scenarios.
Limitations and Future Work
- Assumptions: The method relies on several statistical assumptions (e.g., sub-Gaussian error, restricted eigenvalue condition) which might not always hold in real-world data.
- Computational Cost: While competitive, the proposed method's computational cost is noted to be similar to other advanced methods like Surv-FCD and VS-D-Lasso, and slightly slower than VS-MCP.
- Generalizability: The authors suggest that the approach can be extended to other data types and interaction models, but significant theoretical and methodological development would be required. Further numerical studies are also recommended to better understand its finite-sample performance in broader contexts.
Summary updated