Environment-Adaptive Covariate Selection (EACS)

• EACS is a framework that adapts covariate selection to varying data environments by identifying optimal predictor subsets based on environment-specific features.
• The methodology employs both discrete selectors and soft-gating networks to map environment summaries to tailored covariate sets, minimizing prediction error under covariate shift.
• Empirical and theoretical studies show that EACS improves OOD performance in simulations and real-world applications, including gene–environment interactions, by leveraging proxy and causal covariates.

Environment-Adaptive Covariate Selection (EACS) encompasses a class of methodologies for identifying covariate sets whose predictive value is environment-dependent—that is, the optimal subset of predictors for a target outcome varies conditional on the statistical or causal characteristics of the data environment. These methods stand in contrast to traditional covariate selection strategies that seek a single, static subset invariant across observed or unobserved environments. The EACS framework is motivated by the persistent failures of causal or invariant selection approaches under out-of-distribution (OOD) shifts, especially when only a subset of the true causes is observed and proxy or non-causal covariates may provide environment-specific utility (Zuo et al., 5 Jan 2026).

1. Formal Problem Setting and Motivation

EACS arises in the context of OOD prediction across a meta-distribution of environments $\mathcal E$, where each environment $e$ defines a data-generating process $P_e(X, Y)$ over covariates $X\in\mathbb R^p$ and outcomes $Y\in\mathbb R$. At test time, only unlabeled covariate samples from a new environment $e_{\mathrm{test}}\sim P_U$ are available, and the objective is to construct a predictor with minimal environment-specific mean squared error (MSE) under $P_{e_{\mathrm{test}}}$, accounting for covariate shift. EACS acknowledges that in many settings, especially when some causes are unobserved, non-causal covariates (often labeled as "spurious") may function reliably as proxies in certain environments, but can degrade performance when their proxy relationships are disrupted by shifts unique to the new environment (Zuo et al., 5 Jan 2026).

2. Core Methodologies and Algorithms

The EACS paradigm decomposes into two main algorithmic pathways: discrete environment-adaptive subset selection and continuous (soft-gating) variants.

Discrete Selector Framework:

• Environments are mapped to fixed-dimensional summaries $$u_e = f_{\mathrm{env}}(\{X_{i, e}\}_{i=1}^{n_e}) \in \mathbb R^d$$, using either engineered moments (means, variances, correlations) or learned invariant encoders such as DeepSets.
• A candidate library of covariate masks $Z\subseteq\{0, 1\}^p$ is constructed. Each $z\in Z$ defines a fixed subset predictor $f_z$ trained on pooled labeled data.
• For each training environment $e$, per-environment risk $R_e(z)$ is estimated for all $z$, labeling each $e$ with an optimal $z^*_e = \arg\min_{z\in Z}\widehat R_e(z)$ according to observed MSE.
• A multiclass classifier $g:\mathbb R^d\to Z$ is trained to map summaries $u_e$ to optimal masks $z^*_e$, producing a mapping from unlabeled target environments to selected covariate subsets.

Soft-Gating Approach:

• Replaces the discrete library $Z$ with a parametric gating network $f_{\mathrm{gate}}(u_e; \theta_{\mathrm{gate}})$ producing continuous gates $\tilde z_e = \sigma(\alpha_e / \tau)$, where $\sigma$ is the logistic sigmoid and $\tau$ is a temperature.
• The continuous mask $\tilde z_e$ adaptively reweights covariates for each environment, with the predictor $p_{\theta_p}(\tilde z_e \circ X_{i, e})$ trained using a joint MSE objective across all environments.
• Both the selector and predictor are optimized by gradient methods, enabling scalability beyond small $|Z|$.

At test time, only unlabeled covariates from the new environment are processed to obtain the summary $u_{e_{\mathrm{test}}}$, after which the environment-specific subset (discrete or continuous) is selected and used for prediction (Zuo et al., 5 Jan 2026).

3. Prior Knowledge and Theoretical Guarantees

EACS methods are designed to flexibly incorporate prior causal knowledge. Given a set $S$ of known causal covariates, the selection space can be restricted (for discrete selectors) to $Z_S = \{z\in Z: z_j = 1\ \forall j\in S\}$, or the soft-gating mask can be clamped so that $\tilde z_{e,j} = 1$ for $j\in S$. This regularization improves finite-sample performance, lowers effective hypothesis complexity, and aligns the learned predictors with known causal relationships (Zuo et al., 5 Jan 2026).

Theoretical guarantees for the discrete selection setting, under standard sufficiency and IID environment assumptions, include:

• Finite-Sample Oracle Inequality: For $n$ samples per environment and $|\mathcal E_{\mathrm{train}}|$ environments, the excess risk over the environment-wise oracle is bounded as

$$\mathbb E_{e\sim P_U}\bigl[R_e(g_{\mathrm{hat}}(u_e)) - \min_{z\in Z}R_e(z)\bigr] \leq C_1\sqrt{\frac{\log|Z|+\log(1/\delta)}{n}} + C_2 |\mathcal E_{\mathrm{train}}|^{-1/2}.$$

• Asymptotic Optimality: If $\log|Z|=o(n)$ and $n, |\mathcal E_{\mathrm{train}}|\to\infty$, the EACS predictor asymptotically matches the oracle environment-specific risk (Zuo et al., 5 Jan 2026).

4. Applications and Empirical Evidence

EACS has been empirically validated in several OOD prediction scenarios:

Simulation:

• In a canonical proxy-covariate generative model ($Y = C_1 + C_2 + \varepsilon_Y$, $X = C_1 - C_2 + \varepsilon_X$), EACS correctly determines the subset ($\{C_2\}$, $\{C_2,X\}$, or others) that is optimal for each environment, depending on how covariate shifts manifest (e.g., perturbations to $X$ destroy the proxy utility of $X$).
• Mean squared error curves for EACS approach the oracle as the number of environments and samples per environment increases.

Real Data:

• On daily bike-sharing data (731 environments, weather variables), EACS using summary-statistic–based selectors outperforms lasso, ICP, anchor regression, and fixed-subset oracles in mean per-environment MSE.
• In US census income prediction (51 state environments, high-dimensional tabular data), soft-gating EACS achieves the lowest per-state MSE compared to OLS, lasso, and anchor regression (Zuo et al., 5 Jan 2026).

A consistent empirical finding is that static causal or invariant selection may underperform ERM, while EACS—which adaptively leverages proxies when they remain reliable—yields uniformly lower OOD prediction error across diverse settings.

5. Relationship to Gene-Environment Interactions and Hierarchical Models

EACS principles are closely related to variable selection in high-dimensional gene–environment (G×E) interaction models. In this context, environment-adaptation manifests in models where the inclusion or exclusion of main and interaction effects depends explicitly on the observed environmental covariate:

• In hierarchical lasso frameworks (Zemlianskaia et al., 2021), selection of G×E interactions is regulated by penalties ensuring a “main-effect-before-interaction” hierarchy, tuning the set of active predictors in response to environmental shifts.
• Bayesian semi-parametric models for G×E selection (Ren et al., 2019) achieve environment-adaptation via hierarchical spike-and-slab priors associated with nonlinear basis expansions in $E$. The inclusion indicators dynamically select main and interaction effects according to the observed data patterns of $(E, Y)$, yielding context-specific sparsity.

6. Computational Strategies and Scalability

EACS frameworks adapt scalable optimization techniques for both selector training and inference in large-scale environments:

• Discrete selectors exploit multiclass classification or regression forests to map environment summaries to indices in $Z$.
• Soft-gating approaches leverage neural net–based gating functions, e.g., with DeepSets environment encoders and MLP gates, optimized by SGD.
• In variable selection for G×E modeling, block coordinate descent with dynamic screening (SAFE, Gap-SAFE), working sets, and active-set strategies allow hierarchical lasso methods to operate efficiently with $p\sim 10^5$–$10^6$ predictors (Zemlianskaia et al., 2021).

This computational infrastructure enables EACS procedures to accommodate high-dimensional predictor libraries, large numbers of environments, and complex summary mappings.

7. Limitations and Scope of Applicability

EACS achieves markedly improved prediction under OOD covariate shift by mapping environment-level covariate distribution signatures to targeted covariate sets, but retains several dependencies:

• Performance hinges on the summary mapping $u_e$ to accurately discriminate environments with preserved versus broken proxy relationships.
• Assumptions of IID sampling of environments and sufficient environment diversity are required for theoretical guarantees.
• When prior causal knowledge is incomplete or incorrect, restricting selection space may have unpredictable effects.

These aspects delimit the scope of EACS applicability. Nevertheless, empirical and theoretical analyses consistently demonstrate that the optimal covariate set for prediction is environment-specific and that EACS delivers near-oracle risk across diverse real-world and synthetic settings (Zuo et al., 5 Jan 2026, Zemlianskaia et al., 2021, Ren et al., 2019).