Statistical inference after variable selection in Cox models: A simulation study

Technical Summary: Statistical Inference After Variable Selection in Cox Models

1. Problem Statement

In biomedical research, high-dimensional survival data often require variable selection (e.g., using Lasso) to identify relevant predictors from a large pool of covariates. However, a fundamental conflict exists between variable selection and classical frequentist inference:

• Selection Bias: Classical inference assumes the model is fixed a priori. When a model is chosen based on the data, the resulting p-values and confidence intervals (CIs) for the selected coefficients are typically biased, anticonservative, and misleading.
• Censoring Complexity: In Cox proportional hazards models, right-censoring introduces additional uncertainty, complicating the asymptotic arguments used to correct these biases.
• The Gap: While post-selection inference (PSI) is well-studied in linear regression, its application and performance in the context of penalized Cox models remain underexplored.

2. Methodology

The authors systematically evaluate several inference procedures applied after variable selection (using both standard and adaptive Lasso) in the Cox model framework. They categorize these methods based on their conditioning strategies and inferential targets:

• Inference Procedures:
  • Sample Splitting: Divides the data into two parts: one for model selection and one for unpenalized estimation/inference. It targets submodel-specific coefficients.
  • Exact Conditional Post-Selection Inference (Exact PSI): Conditions explicitly on the selection event to provide finite-sample valid intervals. It targets submodel-specific coefficients.
  • Debiased Lasso: Applies a one-step correction to the penalized estimator to restore asymptotic normality. Unlike the others, it targets the full-model coefficients and does not condition on the selection event.
• Baseline/Naive Comparators: The study includes a "Full Model" (all covariates), an "Oracle Model" (only true predictors), and "Naive Refitting" (standard Wald CIs applied to the selected model without adjustment).
• Simulation Design: A comprehensive Monte Carlo study ($n_{sim}=1000$) using two settings:
  • Toy Simulations: Controlled environments with varying dimensions ($p$), correlation structures ($\rho$), and baseline hazards (Exponential and Weibull).
  • Realistic Simulation (METABRIC): A calibrated design using clinical covariate distributions and effect sizes derived from a real breast cancer cohort.
• Tuning Strategies: The study tests how the choice of the regularization parameter $\lambda$ (e.g., CV-based, AIC, BIC, or fixed) interacts with inferential validity.

3. Key Contributions

• Extension of Frameworks: Adapts and extends existing Gaussian-based simulation designs (Kammer et al.) to the right-censored survival setting.
• Comparative Evaluation: Provides a systematic assessment of the trade-offs between coverage (validity), width (precision), and power (sensitivity) across multiple methods.
• Tuning Interaction Analysis: Identifies how data-driven tuning (like Cross-Validation) can inadvertently invalidate certain exact inference methods.
• Practical Illustration: Demonstrates the utility of these methods using real-world clinical data (METABRIC) to distinguish stable prognostic signals from unstable ones.

4. Key Results

• Inferential Validity:
  • Debiased Lasso and Sample Splitting are the most robust, achieving selective coverage close to the nominal 90% level across most scenarios.
  • Exact PSI suffers from significant undercoverage (it is too optimistic) when the penalty $\lambda$ is chosen via data-driven methods like Cross-Validation, as it violates the required "fixed $\lambda$" assumption.
• Efficiency (Interval Width):
  • The Debiased Lasso provides the most efficient (narrowest) selective confidence intervals (SCIs).
  • Exact PSI produces the widest intervals due to the high uncertainty of conditioning on complex selection events.
  • Sample Splitting yields wider intervals than the debiased Lasso because it effectively reduces the sample size available for inference.
• Type I Error and Power:
  • Methods accounting for selection (Debiased Lasso, Sample Splitting) control Type I error much better than naive refitting.
  • There is a clear trade-off: prediction-oriented tuning (e.g., $\lambda_{min}$) increases power but leads to higher Type I error variability, whereas selection-oriented tuning (e.g., BIC) favors sparsity and stability.
• Real Data Findings: In the METABRIC analysis, the authors found that the choice of method affects interval stability depending on covariate type: the Debiased Lasso is more precise for continuous variables, while Exact PSI provides more stable intervals for binary/ordinal predictors.

5. Significance

This paper provides critical guidance for biostatisticians and data scientists working with high-dimensional survival data. It demonstrates that predictive accuracy (low IBS) does not guarantee inferential validity.

The study highlights that for practitioners seeking a balance between precision and validity, the Debiased Lasso is a strong candidate, whereas Exact PSI requires caution if the regularization parameter is tuned using the same data. Ultimately, the work emphasizes that post-selection inference is essential for moving from "predictive modeling" to "reliable scientific inference" in clinical survival analysis.