Local Invariant Causal Prediction

• Local invariant causal prediction is a framework for identifying the true causal parents of a target variable by testing whether its conditional mechanism remains stable under different environments and perturbations.
• It utilizes methods such as target-specific testing, sequential segmentation, and environment-specific local linear models to robustly capture invariant features across diverse data settings.
• The approach offers conservative statistical guarantees and has practical applications in fields like time series analysis, psychology, and vision-language models using both constructed and synthetic environments.

Local invariant causal prediction denotes target-centered causal discovery based on the claim that, for the true parent set $S$ of a response $Y$, the conditional mechanism for $Y$ remains stable when the data are viewed across environments, interventions, or other admissible perturbations. In the classical invariant causal prediction formulation, one assumes that there exists a subset $S \subseteq \{1,\dots,D\}$ such that

$$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$

for all environments $e,f$ and all $x$; accepted invariant subsets are then intersected to estimate the causal parents of $Y$ (Madaleno et al., 16 May 2025). The approach is “local” because it is posed for a fixed target variable rather than for the entire graph, but the literature uses the locality idea in several ways: target-specific parent recovery, invariance over local subsets of environments, local linear models, sequentially constructed environments, and local perturbations such as reparametrizations or augmentations (Peters et al., 2015, Mey et al., 2024, Jørgensen et al., 2020).

1. Foundational formulation

The classical reference point is invariant causal prediction in the sense of Peters, Bühlmann, and Meinshausen. In its linear formulation, there exists a vector $\gamma^*$ with support

$S^* := \{k : \gamma_k^* \neq 0\}$

such that, for all environments $Y$0,

$Y$1

with the same error distribution $Y$2 in all environments; the marginal distribution of $Y$3 may vary arbitrarily with $Y$4 (Peters et al., 2015). In a structural equation model without interventions on $Y$5, the direct causal parents of $Y$6 satisfy this invariance, so the task reduces to testing candidate subsets $Y$7 for invariant prediction and defining

$Y$8

If each test has size at most $Y$9, then

$Y$0

which gives family-wise control against false causal inclusions (Peters et al., 2015).

In the more recent BHIP restatement of ICP, the basic assumption is written directly as conditional distribution invariance across environments: $Y$1 for all $Y$2 and all $Y$3, with $Y$4 invariant with respect to $Y$5 under an SCM in which the mechanism $Y$6 and the noise distribution are shared across environments while the marginal $Y$7 may change (Madaleno et al., 16 May 2025). The common target across these formulations is therefore the same: identify a parent set of one response variable by exploiting the invariance of the response mechanism under heterogeneous conditions.

2. What makes the prediction local

The locality of ICP is first of all target-wise. The procedure is designed to recover the parents of a single node $Y$8, not the whole DAG, and can therefore be applied separately to different targets (Peters et al., 2015). This target-wise character remains unchanged in later variants even when the environment structure becomes more complex.

A second sense of locality appears in sequential data. In sequential ICP, environments are not given a priori; instead they are constructed from the order structure of the data. For a sequence $Y$9, one tests

$S \subseteq \{1,\dots,D\}$0

and builds local environments from a grid $S \subseteq \{1,\dots,D\}$1 through block collections

$S \subseteq \{1,\dots,D\}$2

The resulting method is explicitly described as a form of local invariant causal prediction along the time axis: the regression mechanism for $S \subseteq \{1,\dots,D\}$3 given $S \subseteq \{1,\dots,D\}$4 must remain stable across many local time segments rather than across externally labeled regimes (Pfister et al., 2017).

A third sense of locality concerns the model class used within each environment. In L-ICP, the structural support is invariant but the coefficients may vary across environments: $S \subseteq \{1,\dots,D\}$5 with a common support $S \subseteq \{1,\dots,D\}$6 and a common noise distribution across $S \subseteq \{1,\dots,D\}$7. This replaces the global linear restriction of classical ICP by environment-specific local linear models and defines the global parent set as the support shared by the family $S \subseteq \{1,\dots,D\}$8 (Mey et al., 2024). A plausible implication is that “local” in this line of work refers not only to a single target but also to a willingness to model the mechanism around that target with environment-specific local approximations.

3. Generalized invariance criteria

Later work broadens the invariance object itself. Rather than always demanding exact equality of $S \subseteq \{1,\dots,D\}$9 across environments, several methods replace it by graded pooling, moment invariance, or structural noise invariance.

Framework	Invariance object	Setting
BHIP	strong pooling via $$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$0 and non-zero global/local effects	multiple environments
Pearson-risk invariance	$$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$1	single-environment GLM/GAM
IMP-based local discovery	invariant matching of environment-specific linear predictors	intervened target $$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$2
Structural-restriction local discovery	$$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$3 and fixed noise law	single observational environment

In BHIP, the ICP hypothesis is encoded in a hierarchical prior

$$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$4

so that invariant predictors are characterized by effects that are credibly non-zero and strongly pooled across environments. Pooling strength is summarized by

$$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$5

with $$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$6, and $$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$7 interpreted as high invariance (Madaleno et al., 16 May 2025). This produces a soft, graded notion of local invariance across environments rather than the hard accept/reject logic of subset-wise ICP.

For generalized linear and additive models, Pearson-risk invariance provides a different criterion. If $$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$8 follows an exponential dispersion family, the population Pearson risk

$$Y^e \mid (X_S^e = x) \overset{d}{=} Y^f \mid (X_S^f = x)$$9

equals $e,f$0 under the true causal model and is invariant under observational or interventional changes in the distribution of $e,f$1. For Poisson and logistic regression, $e,f$2, so the causal model can be identified from a single environment by combining this invariant Pearson-risk property with likelihood optimality (Polinelli et al., 2024). This is a local invariant prediction principle for one target $e,f$3, but it no longer relies on multiple environments in the observed data.

When $e,f$4 itself is intervened, classical ICP breaks because $e,f$5 need not remain fixed even for $e,f$6. The intervened-target literature replaces conditional invariance by the invariant matching property (IMP): $e,f$7 satisfies IMP if, for every environment $e,f$8,

$e,f$9

for constants $x$0 and $x$1 that do not depend on $x$2. Under the model in which only the coefficients of $x$3 vary across environments, this yields identifiability results for $x$4 through intersections or voting over the $x$5-sets appearing in valid IMPs (Du et al., 2023).

A further single-environment analogue replaces environment invariance by structural noise invariance. For a class $x$6, a subset $x$7 is $x$8-plausible if there exists $x$9 such that

$Y$0

satisfies

$Y$1

and one defines

$Y$2

For sufficiently non-separable additive or location-scale mechanisms, no proper subset of the true parent set is $Y$3-plausible, so $Y$4 (Bodik et al., 2023). This suggests that local ICP can be reinterpreted as a broader search for a target-specific mechanism whose residual law is invariant in the conditioning variables rather than across observed environments.

4. Constructed and synthetic environments

A major theme in the recent literature is that environments need not be externally given. They may be constructed from observational data, from allowable reparametrizations, or from designed perturbations that isolate stable mechanisms.

One direct construction starts from observational data without environment labels. A decision tree is fit for each covariate $Y$5 against $Y$6, and the terminal leaves are treated as environments $Y$7 for that variable. The intended properties are

$Y$8

across environments, while for causal parents one seeks an invariant conditional relation between $Y$9 and $\gamma^*$0. Standard ICP is then run on the generated environments, with an improved version limiting subset size and using a voting rule in higher-dimensional settings (Santillan, 2023). The paper presents this explicitly as a practical way to deploy invariant learning on observational data.

Reparametrization-based pseudo-environments go further by changing the marginals themselves. The core principle is: $\gamma^*$1 for bijections $\gamma^*$2, so the causal direction should be invariant under marginal reparametrizations. The MQV framework operationalizes this by sampling many random bijections, transforming $\gamma^*$3, and recalculating the causal score. The authors describe this as a kind of local invariant causal prediction with respect to marginal reparametrizations: a global invariance principle is implemented through many local transformations of the marginals (Jørgensen et al., 2020).

In vision-LLMs, augmentations are treated as interventions on variant latent factors $\gamma^*$4 while invariant factors $\gamma^*$5 are held fixed. CLIP-ICM constructs an invariant subspace by learning a projection $\gamma^*$6 that approximately preserves the original CLIP embedding while enforcing

$\gamma^*$7

for each augmentation $\gamma^*$8. The paper explicitly interprets this as close in spirit to invariant causal prediction and as a local form of invariance because the environments are the augmentation-induced perturbations rather than arbitrary real-world shifts (Song et al., 2024). A plausible implication is that local ICP has become a general design pattern: create perturbations that alter nuisance structure while preserving the target mechanism, then search for predictors stable under those perturbations.

5. Statistical guarantees, uncertainty, and limits

The inferential appeal of local ICP lies in its conservative guarantees. In classical ICP, testing all subsets and intersecting the non-rejected ones yields

$\gamma^*$9

so accepted variables can be interpreted as a confidence set for direct causes of the target (Peters et al., 2015). L-ICP retains the same logic: if the test of the true invariant set has size at most $S^* := \{k : \gamma_k^* \neq 0\}$0, then the intersection of accepted sets satisfies

$S^* := \{k : \gamma_k^* \neq 0\}$1

and, under Gaussian noise, the practical test is based on the ratio of minimum and maximum residual sums of squares across environments (Mey et al., 2024).

Sequential ICP provides a different type of guarantee. After fitting the global linear model for a candidate $S^* := \{k : \gamma_k^* \neq 0\}$2, it computes scaled residuals and uses exact conditional resampling under the null

$S^* := \{k : \gamma_k^* \neq 0\}$3

which yields asymptotically valid tests for arbitrary measurable statistics of the residual process. In a change-point asymptotic regime, the decoupled test attains rate consistency for coefficient or variance changes whenever

$S^* := \{k : \gamma_k^* \neq 0\}$4

and with a logarithmic grid the detectable rates are close to $S^* := \{k : \gamma_k^* \neq 0\}$5 (Pfister et al., 2017).

Uncertainty quantification has also become more explicit. BHIP returns posterior distributions over global and environment-specific effects, HDIs with ROPE thresholds, posterior inclusion probabilities under spike-and-slab priors, and the pooling factor $S^* := \{k : \gamma_k^* \neq 0\}$6 as a direct invariance score (Madaleno et al., 16 May 2025). MQV produces

$S^* := \{k : \gamma_k^* \neq 0\}$7

and summarizes uncertainty by

$S^* := \{k : \gamma_k^* \neq 0\}$8

so linear or invertible-noise-free cases naturally produce low confidence rather than forced orientation (Jørgensen et al., 2020).

The literature also establishes impossibility results. By mapping linear ICP to a Gaussian multiple access channel, lower bounds on the support-recovery error probability show that environment diversity is not merely helpful but necessary. In the zero-rate setting with $S^* := \{k : \gamma_k^* \neq 0\}$9, any procedure must satisfy

$Y$00

and when environments are effectively identical the error probability is bounded away from zero even as $Y$01 (Goddard et al., 2022). This suggests that local ICP can reduce dimensionality, but it cannot overcome the fundamental absence of informative heterogeneity.

6. Applications, recent generalizations, and open issues

Local invariant causal prediction has been applied in settings where one wants direct causes of a single response rather than a full graph. In psychology, perturbation-graph designs provide repeated context changes for each variable; the paper on perturbation graphs argues that ICP generalizes the marginal perturbation-graph idea by conditioning on additional variables and thereby identifying direct causes instead of merely causal paths (Waldorp et al., 2021). In time series, seqICP has been used for monetary policy, where nonstationarity becomes an asset: local invariance across time blocks can reveal both lagged and instantaneous predictors of exchange-rate movements (Pfister et al., 2017). In generalized linear models, Pearson-risk invariance has been applied to fertility and high-income data, yielding local parent sets for count and binary targets under Poisson and logistic GAM assumptions (Polinelli et al., 2024).

Recent work broadens the predictive stakes of local invariant sets. In prediction–intervention games, the leader predicts a single target $Y$02 while a follower intervenes on covariates. The paper defines the stable blanket

$Y$03

proves that predictors based on the stable blanket are always better or as good as those based on the causal parents for two common classes of follower objectives, and gives sufficient conditions under which the stable-blanket predictor is worst-case optimal over the allowed interventions (Kühne et al., 16 May 2026). This suggests that local ICP is no longer only a discovery device; it is also a design principle for robust prediction under strategic distribution shift.

An even stronger development comes from multi-environment identifiability theory for nonlinear SCMs. Under acyclicity, invariance of mechanisms across environments, and sufficient variability, two auxiliary environments are sufficient to infer the full causal graph for arbitrary nonlinear mechanisms and to identify the SCM up to invertible elementwise reparameterizations of the sources (Montagna et al., 13 May 2026). A plausible implication is that classical local ICP-type guarantees for one target become corollaries of a stronger global invariance theorem: if the whole graph is identifiable from a constant number of environments, then the local parent set of any target is identifiable as well.

The open issues remain substantial. The environment set may be unknown or only approximately intervention-like; the functional class used for the target mechanism may be misspecified; local predictors that are invariant on training environments may fail under richer deployment shifts; and finite-sample invariance testing can be unstable in high dimensions. Across the literature, however, the same organizing idea recurs: identify a subset of variables for one target $Y$04 such that the mechanism for $Y$05 is stable under the relevant perturbations, and interpret that subset as causal, robust, or both.