Ensemble Conditional Independence Test (E-CIT)

• E-CIT is a flexible meta-framework that accelerates base conditional independence tests by partitioning data into disjoint subsets and aggregating their p-values.
• It reduces computational complexity by transforming a full-sample test into many efficient subset-level tests, achieving linear scalability when subset size is fixed.
• The method leverages the closure properties of stable distributions to ensure theoretical validity and improved power in constraint-based causal discovery.

Ensemble Conditional Independence Test (E-CIT) is a general, plug-and-play framework for conditional independence testing that accelerates a chosen base CIT by partitioning the sample into disjoint subsets, applying the base test independently on each subset, and aggregating the resulting $p$-values through a stable-distribution transform. Introduced for constraint-based causal discovery, especially in settings such as PC and FCI, it targets the per-test sample-size bottleneck of expensive conditional independence procedures and aims to reduce the effective sample-size complexity of a wrapped test to linear in $n$ when subset size is fixed (Guan et al., 25 Sep 2025).

1. Motivation and conceptual scope

Constraint-based causal discovery repeatedly tests hypotheses of the form

$$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$

In algorithms such as PC and FCI, the computational bottleneck is not only the number of such tests, but also the fact that many strong nonparametric CITs have high time complexity with respect to sample size. Kernel-based procedures such as KCIT are a canonical example. E-CIT addresses this by reducing the per-test computational cost with respect to $n$, rather than by changing the logical structure of causal discovery itself (Guan et al., 25 Sep 2025).

The framework is deliberately generic. It does not define a new base conditional independence statistic for every problem class; instead, it wraps an existing CIT $\mathcal H$ and replaces a single full-sample call with many subset-level calls. This makes E-CIT a meta-procedure rather than a domain-specific test. A plausible implication is that its main value lies in the interface between statistical testing and systems-level scalability: a strong but expensive CIT can be reused with different subset sizes and aggregation parameters without redesigning the underlying test.

The term “ensemble” in E-CIT refers to aggregation across subset-level subtests, not to an ensemble of heterogeneous predictive models. This distinguishes it from methods whose repeated component is over permutations, generators, or classifiers. For example, COINP is a predictive permutation CIT with repeated refitting over permutations, but not an ensemble over multiple learners (Inácio et al., 2019). E-CIT instead ensembles multiple $p$-values produced by the same base CIT on disjoint data partitions (Guan et al., 25 Sep 2025).

2. Divide-and-aggregate formulation

Suppose $n$ i.i.d. samples are available from the joint distribution of $(X,Y,Z)$, and let $\mathcal H$ denote a base CIT that returns a $p$-value. E-CIT assumes an equal partition

$n$0

where $n$1 is the number of disjoint subsets and $n$2 is the subset size. The data are split into

$n$3

each of size $n$4. For each subset $n$5, the same conditional independence hypothesis is tested with the same base procedure: $n$6 The ensemble test is denoted

$n$7

where $n$8 specify a stable distribution used for aggregation (Guan et al., 25 Sep 2025).

The aggregation step is the technical center of the method. Let $n$9 be the CDF of a stable random variable

$$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$0

E-CIT transforms each subset $$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$1-value by the inverse stable CDF and averages: $$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$2 If

$$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$3

then the global $$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$4-value is

$$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$5

The closure of stable distributions under averaging is what makes this construction analytically tractable under ideal null $$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$6-values (Guan et al., 25 Sep 2025).

The stable family includes a familiar special case. When $$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$7, the stable law is Gaussian, and the method reduces to a Stouffer-style normal-score combination rule. The paper therefore presents E-CIT not as an arbitrary meta-analysis heuristic, but as a flexible family of transformations indexed mainly by the stability parameter $$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$8, which controls tail behavior and changes the way small subset $$H_0: X \perp\!\!\!\perp Y \mid Z \qquad \text{vs.} \qquad H_1: X \not\!\!\perp\!\!\!\perp Y \mid Z.$$9-values are aggregated (Guan et al., 25 Sep 2025).

3. Statistical guarantees and assumptions

The paper’s main theorem states that for well-defined subtest $n$0-values, the ensemble test satisfies three properties: validity under $n$1, admissibility, and unbiasedness when the subtests are unbiased. The validity argument is direct. Under $n$2, if each $n$3 is uniform on $n$4, then

$n$5

the averaged statistic $n$6 is again stable with scale $n$7, and therefore

$n$8

This yields exact Type I validity in the idealized regime of independent, identically distributed, null-uniform subset $n$9-values (Guan et al., 25 Sep 2025).

The paper also gives a consistency statement in terms of the number of subsets. If $\mathcal H$0 are i.i.d. subtest $\mathcal H$1-values under the alternative and $\mathcal H$2 is integrable, then

$\mathcal H$3

provided

$\mathcal H$4

where $\mathcal H$5 is ensemble power at level $\mathcal H$6. A more interpretable sufficient condition is also given: if

$\mathcal H$7

if the density $\mathcal H$8 of $\mathcal H$9 satisfies

$p$0

and if

$p$1

then the power tends to one as $p$2 (Guan et al., 25 Sep 2025).

These guarantees are conditional on strong but explicit assumptions. Exact validity requires subset $p$3-values that are null-uniform and independent across subsets. Independence is natural for disjoint partitions under i.i.d. sampling, but exact uniformity can fail in practice because many CITs are only asymptotically calibrated. The paper therefore distinguishes exact theory from practical approximation. In particular, if a base CIT returns super-uniform permutation $p$4-values rather than exactly uniform ones, then the stable-transform argument is no longer exact. The paper notes that E-CIT is then only approximately valid, although the approximation improves when the number of permutations is large (Guan et al., 25 Sep 2025).

A common misconception is that the stable aggregation by itself repairs a miscalibrated base CIT. The paper does not claim this. If the subset-level null $p$5-values are already distorted, E-CIT may inherit or even amplify that distortion. This aligns with later calibration work such as ECCIT, which explicitly treats miscalibration as a separate problem and corrects a chosen base CIT through an adversarial calibration map rather than through subset aggregation (Pan et al., 24 Feb 2026).

4. Computational implications and practical tuning

If the base CIT has computational cost $p$6 on $p$7 samples, ordinary use costs $p$8 per test. Under E-CIT, the main cost becomes

$p$9

since aggregation of scalar $n$0-values is negligible. Using $n$1, this can be written as

$n$2

If $n$3 is held fixed while $n$4 grows, then $n$5 is constant, so the total complexity becomes $n$6. For a superlinear base test such as $n$7, the corresponding E-CIT cost is

$n$8

which is linear in $n$9 for fixed $(X,Y,Z)$0 (Guan et al., 25 Sep 2025).

This computational gain is traded against statistical degradation in the subtests. If $(X,Y,Z)$1 is too small, each subset-level CIT may have weak power or poor null calibration. If $(X,Y,Z)$2 is too large, the acceleration becomes modest. The paper repeatedly emphasizes this subset-size tradeoff and recommends choosing $(X,Y,Z)$3 large enough that the base CIT behaves reasonably on one subset. In the reported experiments, the default setting is

$(X,Y,Z)$4

motivated by empirical performance and validated by ablation (Guan et al., 25 Sep 2025).

The practically most important aggregation hyperparameter is the stable parameter $(X,Y,Z)$5. The paper recommends fixing

$(X,Y,Z)$6

and tuning only $(X,Y,Z)$7. In the main experiments, $(X,Y,Z)$8, with $(X,Y,Z)$9 corresponding to the Gaussian/Stouffer case. The paper reports that larger $\mathcal H$0 often increases power, although the better choice depends on the base CIT and the data-generating mechanism (Guan et al., 25 Sep 2025).

The partition strategy is also structurally important. The theory assumes disjoint subsets. The paper explicitly warns that bootstrap or overlapping splits would induce correlated $\mathcal H$1-values and require different analysis. For permutation-based CITs, another practical detail arises: exact subset-level $\mathcal H$2-values may equal 0 or 1. The paper recommends adding a small random perturbation in such cases to avoid numerical instability when applying $\mathcal H$3 (Guan et al., 25 Sep 2025).

5. Empirical behavior in testing and causal discovery

The empirical study spans both stand-alone CI testing and integration into causal discovery. In synthetic post-nonlinear experiments, the paper compares E-KCIT with KCIT, RCIT, and FastKCIT over 1000 trials. The data-generating setup includes nonlinearities sampled from

$\mathcal H$4

with $\mathcal H$5 drawn from either a standard normal or standard Laplace distribution, and noise distributions including Student-$\mathcal H$6, Laplace, and Cauchy. The significance level is 0.05. In this setting, E-KCIT yields substantial runtime reductions while maintaining competitive power, and under heavier-tailed noise it often behaves more consistently than the alternatives (Guan et al., 25 Sep 2025).

The framework is also applied to several other base CITs: RCIT, LPCIT, CMIknn, CCIT, and FisherZ. The reported pattern is not uniform across methods. E-CIT often improves power for RCIT, LPCIT, and FisherZ while keeping Type I error acceptable. For CMIknn, the gains are smaller and sometimes absent. For CCIT, whose original Type I behavior is poor in the reported simulations, E-CIT substantially reduces Type I error, with some power loss. This suggests that subset aggregation can act as a stabilizer for some unstable base tests, although the paper does not claim universal improvement (Guan et al., 25 Sep 2025).

On the Sachs flow-cytometry benchmark, the ensemble version improves the F1 score of several methods. The reported changes are:

• KCIT: from $\mathcal H$7 to $\mathcal H$8,
• RCIT: from $\mathcal H$9 to $p$0,
• LPCIT: from $p$1 to $p$2,
• FisherZ: from $p$3 to $p$4,
• CCIT: from $p$5 to $p$6,
• CMIknn: from $p$7 to $p$8.

These results are consistent with the paper’s claim that E-CIT is especially helpful on complex real-world data and for methods whose raw behavior is less robust. They also show that the framework is not uniformly dominant: CMIknn declines slightly on this benchmark (Guan et al., 25 Sep 2025).

The framework is further embedded into the PC algorithm. Replacing the base CI oracle with E-KCIT yields lower runtime than KCIT and often better graph-recovery quality than both KCIT and RCIT on nonlinear graphs with additive $p$9, Cauchy, and Laplace noise. The reported evaluation criteria are skeleton F1-score, structural Hamming distance, and runtime. A plausible implication is that E-CIT is most attractive when the downstream pipeline is dominated by many expensive CI calls, because its gains compound across the entire search procedure (Guan et al., 25 Sep 2025).

6. Relation to adjacent CI frameworks and common confusions

E-CIT belongs to a wider family of conditional independence methods that decompose the testing problem into smaller randomized, localized, or model-based subproblems, but its ensemble axis is specific. It is most closely characterized as subset-level $n$00-value aggregation for a single base CIT.

Method family	Repeated component	Relation to E-CIT
COINP (Inácio et al., 2019)	Permutations with refitting	Predictive permutation CIT, not subset aggregation
Wasserstein CI (Warren, 2021)	Local tests over bins of $n$01	Also divide-and-aggregate, but over conditioning regions
ECCIT (Pan et al., 24 Feb 2026)	Adversarial calibration datasets	Calibration wrapper, not an ensemble test
E-CIT (Guan et al., 25 Sep 2025)	Disjoint data subsets	Base-CIT acceleration by stable $n$02-value aggregation

A first misconception is to equate E-CIT with model ensembling. The framework does not average across different learners, loss functions, or feature representations. It wraps one chosen base CIT at a time. In this respect it differs sharply from predictive CI procedures that are “model-agnostic” because many learners can be plugged in, yet still run one learner per test instance, such as COINP (Inácio et al., 2019).

A second misconception is to conflate E-CIT with empirical calibration. ECCIT is a test-agnostic framework that measures worst-case miscalibration of a chosen base CIT and then adjusts its $n$03-values through a monotone calibration map. E-CIT does not solve that problem; it assumes that the subset-level $n$04-values are already sufficiently valid for stable aggregation to make sense (Pan et al., 24 Feb 2026).

A third misconception is that E-CIT is the only divide-and-aggregate formulation in the CI literature. It is not. Wasserstein Conditional Independence Testing also partitions the problem into multiple local tests, but it does so by binning the support of $n$05 and aggregating local Wasserstein discrepancies rather than by splitting the sample into disjoint subsets and combining $n$06-values (Warren, 2021). This suggests that “ensemble” in CI testing can refer to several aggregation axes: over subsets, over conditioning regions, or over model classes. E-CIT is specifically the subset-aggregation variant.

Historically, this places E-CIT alongside other modular CI designs rather than in opposition to them. A plausible implication is that the framework can serve as an outer systems layer around many existing CITs, while leaving open separate questions of calibration, local adaptivity, and model choice. That division of labor is central to its identity: E-CIT is primarily a scalable wrapper for causal-discovery workloads, not a replacement for the diverse statistical mechanisms used by modern conditional independence tests (Guan et al., 25 Sep 2025).