Knockoff Filter

The knockoff filter provides exact finite-sample FDR control under minimal assumptions, unlike heuristic thresholding methods. It guarantees that the expected proportion of false discoveries among all selections remains below a user-specified threshold (e.g., 10%). This is achieved by comparing the importance of each original feature against its synthetic knockoff counterpart, which is known to be null by construction.

Knockoffs are synthetic features engineered to satisfy two critical properties:

• Exchangeability: Swapping any original feature with its knockoff leaves the joint distribution unchanged.
• Nullity: Knockoffs are constructed independently of the response variable, ensuring they carry no true signal. For Gaussian designs, Model-X knockoffs generate these decoys by sampling from the conditional distribution of features given all others, preserving the exact covariance structure.

The filter relies on anti-symmetric statistics that measure the importance of each original feature relative to its knockoff. A common choice is the Lasso Signed Max (LSM) statistic:

• W_j = max(|β_j|, |β̃_j|) × sign(|β_j| - |β̃_j|) A large positive W_j indicates the original feature dominates its knockoff, providing evidence of true association. Features with W_j below a data-adaptive threshold are rejected as null.

The Model-X knockoff filter extends the methodology to high-dimensional settings where the number of features p can vastly exceed the number of observations n. It treats the feature distribution as known or estimable, generating valid knockoffs without requiring any model for the response. This makes it particularly powerful for genome-wide association studies (GWAS) and biomarker discovery where p may be in the millions.

To reduce the randomness inherent in a single knockoff draw, the multiple knockoff filter generates M independent knockoff copies for each original feature. This aggregation:

• Stabilizes the selection set across different random seeds.
• Increases statistical power by averaging out noise in the importance statistics.
• Allows the use of more flexible importance measures, including those from random forests and deep neural networks.

Unlike the Benjamini-Hochberg (BH) procedure, which operates on p-values from univariate tests, the knockoff filter:

• Accounts for multivariate correlation among features, avoiding redundant selections of correlated nulls.
• Does not require valid p-values, making it compatible with complex machine learning models.
• Controls FDR conditional on the selected set, not just in expectation over hypothetical replications. However, BH requires only p-values, while knockoffs demand accurate knowledge of the feature distribution.

The expected proportion of false positives among all rejected null hypotheses. The Knockoff Filter is explicitly designed to control the FDR at a target level (e.g., 20%).

• Benjamini-Hochberg: The foundational procedure for FDR control in multiple testing.
• q-value: The minimum FDR at which a test is called significant.
• In biomarker discovery, controlling FDR ensures that a known fraction of selected features are likely genuine.

The importance scores calculated for both original and knockoff variables to determine selection. The key is the antisymmetric property.

• Lasso Coefficient Difference: W = |β_original| - |β_knockoff|
• Lasso Signed Max: A variant using the lambda path.
• A large positive W value indicates a feature that is significantly more important than its synthetic decoy, providing evidence against the null hypothesis.

An extension that generates knockoff variables without assuming a specific model for the response Y, only requiring knowledge of the feature distribution P(X).

• Uses generative models or moment-matching to sample knockoffs.
• Guarantees exact FDR control in finite samples for any conditional distribution of Y|X.
• Essential for high-dimensional genomics where the true response model is unknown.

A linear regression method that performs embedded feature selection by shrinking coefficients to exactly zero. It is the most common feature statistic engine within the Knockoff Filter.

• Solves: min(||y - Xβ||² + λ||β||₁)
• The regularization path across λ values reveals the order of variable entry.
• Knockoffs compete with original features for entry into the LASSO model, creating a controlled selection environment.

A step-up method for controlling the false discovery rate in multiple hypothesis testing. The Knockoff Filter provides an alternative FDR control framework that accounts for feature correlation.

• Sorts p-values and compares to a linear threshold.
• Unlike BH, knockoffs do not require valid p-values, only symmetric feature statistics.
• Often used as a benchmark to validate the selection power of the knockoff approach.

A robust ensemble method that applies a selection algorithm to random subsamples of the data and selects features chosen with high frequency.

• Controls the expected number of false positives (Per-Family Error Rate).
• Complements the Knockoff Filter by adding a layer of selection stability.
• Combining both methods yields a highly conservative, reproducible biomarker set for clinical validation.