Efficient iLOCO CRT for Conditional Testing

• Interaction LOCO (iLOCO) is a method for conditional independence testing that systematically leaves out one covariate to compute valid p-values with near-zero variability.
• It overcomes limitations of knockoff methods by providing continuous p-values and reducing computational costs, especially in high-dimensional and multivariate Gaussian settings.
• The approach ensures rigorous error rate control and scalability, and it accelerates computations for L1-regularized estimators in sparse models.

Interaction LOCO (iLOCO) refers to the leave-one-covariate-out conditional randomization test, a computationally efficient methodology for conditional independence testing under the model-X framework. The model-X assumption requires knowledge of the joint distribution of covariates but makes no assumptions about the conditional distribution of the outcome given the covariates. Conditional independence testing is fundamental yet provably hard without assumptions; traditional approaches such as knockoffs suffer from high variability and only provide one-bit $p$-values per variable. The iLOCO CRT addresses these drawbacks by offering valid $p$-values with nearly zero algorithmic variability and computational efficiency, particularly with a closed-form solution in the multivariate Gaussian covariate setting (Katsevich et al., 2020).

1. The Model-X Framework and Conditional Independence Testing

The model-X conditional independence testing paradigm operates under the assumption that the joint distribution of a set of covariates is fully known, while the relationship of these covariates to the response variable remains unconstrained. This approach enables the principled evaluation of whether any given covariate is conditionally independent of the response, given all other covariates. Conditional randomization test (CRT) methodologies have emerged as “the right” solution under this framework but have been considered computationally inefficient in practice (Katsevich et al., 2020).

2. Limitations of Knockoff Methodologies

Knockoffs, a popular method associated with the model-X framework, enable variable selection with controlled error rates. However, two significant limitations impede its statistical and practical utility: (a) Knockoffs only provide “one-bit” $p$-values per variable, restricting the granularity of inference, and (b) the method is highly randomized, leading to substantial result variability across repeated runs. These issues motivate the search for alternative procedures that offer more stable and informative statistical output (Katsevich et al., 2020).

3. The Leave-One-Covariate-Out Conditional Randomization Test (LOCO CRT)

The LOCO CRT method is introduced as a response to the deficiencies of knockoffs and the computational bottlenecks of standard CRT. By systematically removing (leaving out) a single covariate, the LOCO CRT produces valid $p$-values for each variable that can be used for familywise error rate control. A key property is its near-zero algorithmic variability, which enhances reproducibility and reliability for downstream inference (Katsevich et al., 2020).

4. Acceleration with L1ME CRT for Regularized Estimators

The computational efficiency of LOCO CRT is further improved for L1 regularized M-estimators (e.g., lasso). The L1ME CRT variant reuses computation by exploiting the stability of the cross-validated lasso upon exclusion of inactive variables. This suggests that in sparse high-dimensional models, most L1 regularized solutions are unaffected by removing variables with zero coefficients, reducing the computational burden when reassessing models for leave-one-covariate-out testing (Katsevich et al., 2020).

5. Closed-Form LOCO CRT for Multivariate Gaussian Covariates

A notable advancement is presented for the special case where the covariates jointly follow a multivariate Gaussian distribution. In this scenario, the LOCO CRT $p$-value admits a closed-form solution, obviating the need for computationally intensive resampling. A plausible implication is that, for practitioners working with Gaussian predictors, conditional independence testing via iLOCO becomes both exact and scalable for large datasets (Katsevich et al., 2020).

6. Statistical Validity and Error Rate Control

LOCO CRT maintains the statistical guarantees central to the model-X framework. The $p$-values it generates can be used for rigorous error rate control at both per-variable and familywise levels. This enables practitioners to perform credible variable selection and hypothesis testing with full transparency about false-positive rates, resolving one of the primary criticisms of prior randomized and coarse-grained alternatives (Katsevich et al., 2020).