Hybrid Two-Stage Causal Screener

Updated 4 July 2026

Hybrid Two-Stage Causal Screener is a framework that decouples a broad, computationally efficient screening phase from a precise verification stage to identify candidate causal effects.
It leverages methods like marginal screening and targeted testing to control type I error and enhance power in high-dimensional regression and causal discovery settings.
Applications of this methodology span interaction detection, proximal causal learning, risk management in prediction markets, and confounder selection for doubly robust inference.

Searching arXiv for the cited papers and closely related work to ground the article. A hybrid two-stage causal screener denotes a class of procedures that separate causal discovery, causal estimation, or candidate-effect detection into an initial screening stage and a subsequent verification, estimation, pruning, or reranking stage. Across recent work, the architecture appears in high-dimensional interaction search, proximal causal learning with proxies, prediction-market lead–lag discovery, ultra-high-dimensional confounder screening, and causal graph recovery from observational data (Jonker et al., 2024, Mastouri et al., 2021, Kim et al., 4 Feb 2026, Tang et al., 2020). The unifying design is that the first stage is deliberately cheaper or broader, whereas the second stage is narrower and more inferentially targeted. The object being screened, however, varies by domain: in some settings it is an interaction term or a confounder set; in others it is a bridge function, a directed lead–lag relation, or a partially oriented causal graph.

1. General architecture and scope

The two-stage pattern is not tied to a single causal formalism. In some papers, the method is explicitly a causal estimator under potential-outcomes or proximal-identification assumptions; in others, it is a causal discovery or causal screening device; and in at least one important case it is an interaction-testing framework that is useful upstream of causal interpretation but is not itself formal causal inference in the potential-outcomes sense (Jonker et al., 2024).

Setting	Stage 1	Stage 2
High-dimensional interactions	Marginal Wald screening of variables	Interaction Wald tests with Bonferroni over tested pairs
Proximal causal learning	Learn $\mu_{W\mid A,X,Z}$ by kernel conditional mean embedding	Learn the bridge $h(A,X,W)$ or solve a kernel moment restriction
Prediction markets	Granger-causality ranking of directed pairs	LLM semantic reranking by plausible economic transmission mechanism
Doubly robust causal inference	Conditional ball covariance screening	Penalized refinement and doubly robust ACE estimation

This architectural recurrence suggests a common methodological principle: use one family of methods to reduce uncertainty before applying another family where it is most useful. In the literature considered here, the first stage may be marginal regression, kernel ridge regression, conditional dependence screening, FCI or PC, topological ordering, or Granger causality; the second stage may be interaction testing, bridge-function estimation, semantic filtering, Bayesian scoring, or local orientation and pruning (Bystrova et al., 2023, Dash et al., 2013, Hiremath et al., 2024, Chen et al., 2021).

2. High-dimensional interaction screening as a two-stage test

In very high-dimensional regression, especially when $p \gg n$ , exhaustive testing of all two-way interactions is computationally infeasible because the number of candidate pairs is of order $p^2$ . The two-stage testing procedure of "Two-Stage Testing in a high dimensional setting" formalizes a screen-and-test strategy for detecting pairs $(x_k,x_\ell)$ whose interaction term is associated with an outcome $Y$ (Jonker et al., 2024).

The full interaction model is

$h\big(\mathbb{E}(Y|(x_k,x_\ell),,)\big) = \beta_0 + \beta_1 x_{k} + \beta_2 x_\ell + \beta_3\; x_{k} x_\ell,$

where $\beta_3$ is the interaction parameter of interest. Stage 1 performs marginal screening variable by variable through

$h\big(\mathbb{E}(Y|{X},=(\gamma_0,\gamma_1),)\big) = \gamma_0 + \gamma_1 x_k,$

testing $H_0:\gamma_1=0$ with the Wald statistic

$h(A,X,W)$ 0

Variables with marginal $h(A,X,W)$ 1-values below a chosen threshold $h(A,X,W)$ 2 are retained. Stage 2 then tests only those pairs for which both variables survived screening, using

$h(A,X,W)$ 3

with interaction null $h(A,X,W)$ 4 and Wald statistic

$h(A,X,W)$ 5

Multiplicity correction in the second stage uses Bonferroni over the number of pairs actually tested there, $h(A,X,W)$ 6, so

$h(A,X,W)$ 7

The main theoretical contribution is an asymptotic-independence argument. For nested GLMs with canonical link, if $h(A,X,W)$ 8, then

$h(A,X,W)$ 9

are asymptotically independent, where $p \gg n$ 0 denotes the component of the stage-2 estimator orthogonal to the stage-1 parameter space. Under the full null, complete independence additionally requires mutually independent standardized covariates, independence from fixed covariates, and mean zero. Under these conditions, logistic, Poisson, and linear models yield asymptotic independence of stage-1 and stage-2 Wald statistics; analogous logic is developed for the Cox proportional hazards model using the partial likelihood.

The empirical results show type I error control and a power increase relative to exhaustive pairwise testing. Simulations were conducted for linear regression, Poisson regression, and Cox proportional hazards regression under both uncorrelated and correlated markers. Under the full null, empirical FWER for uncorrelated markers was close to $p \gg n$ 1 across the three models and several first-stage thresholds, whereas for correlated markers it was generally below $p \gg n$ 2 and conservative, with reported values around $p \gg n$ 3– $p \gg n$ 4. Power was higher than one-by-one pair testing when there were no main effects or when main effects were in the same direction; power could decline in the “opposite main effects” setting because marginal signals can cancel and the causal markers may fail screening.

Although this procedure is not formal causal inference, the paper explicitly notes that it is directly useful as a “hybrid causal screener” if “causal” is interpreted as “candidate effect of interest.” Fixed covariates such as sex, ethnicity, or smoking can be included in both stages, and the asymptotic-independence result provides a principled multiplicity argument after screening. The method therefore sits naturally upstream of causal interpretation rather than replacing it.

3. Two-stage causal estimation with proxies and structured exposures

A more explicitly causal use of the two-stage architecture appears in proximal causal learning with proxies for unobserved confounding. "Proximal Causal Learning with Kernels: Two-Stage Estimation and Moment Restriction" studies the target curve

$p \gg n$ 5

under the proxy assumptions

$p \gg n$ 6

together with completeness conditions ensuring that the proxies are sufficiently informative about the latent confounder $p \gg n$ 7 (Mastouri et al., 2021). Identification proceeds through the bridge function $p \gg n$ 8 defined by the Fredholm equation

$p \gg n$ 9

and the causal effect is recovered via

$p^2$ 0

Kernel Proxy Variable learning (KPV) is the paper’s two-stage estimator. In Stage 1, it learns the conditional mean embedding

$p^2$ 1

from samples $p^2$ 2 by ridge regression. In Stage 2, it learns the bridge function as an RKHS-valued regressor using a second sample $p^2$ 3 through

$p^2$ 4

The representer-theorem solution is a double sum over Stage 1 and Stage 2 samples,

$p^2$ 5

which underscores that the first stage does not merely estimate a nuisance quantity; it constructs the feature basis used in the second stage. The paper also proposes Proxy Maximum Moment Restriction (PMMR), based on the conditional moment restriction

$p^2$ 6

and shows that KPV and PMMR minimize aligned risk functionals under the bridge condition.

A different but structurally related two-stage causal estimator is developed for semi-continuous exposures in "Two-stage Estimation for Causal Inference Involving a Semi-continuous Exposure" (Wang et al., 26 Nov 2025). There the exposure has a point mass at zero and a continuous positive part, with $p^2$ 7 and $p^2$ 8 among the exposed. The paper introduces a two-part propensity structure,

$p^2$ 9

and a marginal structural model

$(x_k,x_\ell)$ 0

Stage I estimates the causal dose-response among the exposed, and Stage II estimates the exposure-status effect at reference dose $(x_k,x_\ell)$ 1 using the Stage-I estimate as an offset: $(x_k,x_\ell)$ 2 Stage II can be implemented by PS regression adjustment, IPW, or AIPW, with double robustness in Stage II for AIPW if either $(x_k,x_\ell)$ 3 or the imputation model is correct, provided Stage I is consistent. The paper explicitly interprets this architecture as “hybrid” because it combines a continuous-treatment stage with a binary-treatment stage, but it also notes that it is not a screener in the variable-selection sense.

These two lines of work show that “two-stage causal screener” can denote more than a filter for variables or edges. It may also mean a decomposition of the inferential problem itself: first construct a proxy-conditioned representation or estimate a dose-response component, then target the final causal functional.

4. Statistical-semantic screening in prediction markets

In prediction markets, the term “hybrid two-stage causal screener” is used directly for a pipeline that combines time-series evidence with mechanism-based semantic filtering. "LLM as a Risk Manager: LLM Semantic Filtering for Lead-Lag Trading in Prediction Markets" proposes a first stage based on Granger causality and a second stage based on LLM assessment of whether the proposed direction admits a plausible economic transmission mechanism (Kim et al., 4 Feb 2026).

The first stage starts from market-implied probability time series and tests directional predictability for every unordered event pair in both directions, sweeping lag lengths

$(x_k,x_\ell)$ 4

The direction with stronger statistical evidence is retained, candidate directed pairs are ranked by statistical strength, and the top $(x_k,x_\ell)$ 5 directed pairs form the candidate set. The default trading portfolio then selects the top $(x_k,x_\ell)$ 6 directed pairs. The paper characterizes this first stage as a high-recall statistical sieve: it is intended to find plausible directional dependence rather than to guarantee economic meaning.

The second stage does not discover new pairs. It re-ranks the statistically screened candidates using event titles, descriptions, the directed pair $(x_k,x_\ell)$ 7, and contextual information sufficient to judge whether the leader could plausibly transmit information to the follower. The prompt template is fixed across pairs, and the model is asked to assess whether a plausible economic transmission mechanism exists beyond correlation, assign a strength level, and predict the expected sign of co-movement, returning a structured JSON output. The mechanism types described in the paper include macro spillovers, cross-border trade links, policy transmission, financial linkages, and sectoral demand or supply effects.

Because causal ground truth is unobserved, evaluation is trading-based rather than graph-theoretic. On Kalshi Economics markets from October 2021 to November 2025, after filtering inactive or low-variation contracts, the study retained 554 event markets and evaluated the methods over 18 rolling windows, each with a 60-day training window and 30-day testing window. Under the default setting of $(x_k,x_\ell)$ 8 days, $(x_k,x_\ell)$ 9, and $Y$ 0, the hybrid method improved win rate from 51.4% to 54.5%, changed average win from \$Y$1636, improved average loss from -\$Y$2347, and increased total PnL from \$Y$312,500. The most prominent effect was a 46.5% reduction in average loss magnitude. The loss reduction persisted for same-event and different-event pairs and strengthened when leader repricings were large: for 5–10 point moves the win rate increased from 57.1% to 66.7%, and for 10+ point moves from 53.8% to 71.4%.

The paper’s interpretation is that the LLM behaves as a semantic risk manager layered on top of statistical discovery. It does not replace econometric screening; rather, it down-ranks fragile links that are statistically significant but economically dubious. The authors are explicit about the caveats: causal ground truth is unobserved, Granger causality is only a proxy for structural causation, the LLM score is not specified as a formal mathematical objective, and potential lookahead from pretraining knowledge remains a concern despite a post-cutoff evaluation after May 31, 2024.

5. Covariate screening for doubly robust causal inference

A hybrid two-stage screener also appears in ultra-high-dimensional confounder selection for observational causal inference. "Ultra-high Dimensional Variable Selection for Doubly Robust Causal Inference" studies settings in which

$Y$4

with target estimand

$Y$5

The proposed Causal Ball Screening (CBS) procedure is explicitly two-step: a model-free screening stage followed by refined selection and doubly robust estimation (Tang et al., 2020).

For each covariate $Y$6, Stage 1 computes the empirical conditional ball covariance

$Y$7

or equivalently

$Y$8

and retains the top $Y$9 variables with the largest values. The rationale is causal rather than predictive. The paper seeks to retain confounders and precision variables while screening out instrumental variables and null variables, because treatment-only predictors that are not confounders may inflate variance. This distinction is central to the method: causal variable selection is not the same as prediction-driven feature selection.

Stage 2 fits separate Lasso outcome regression models for treated and untreated groups and an adaptive-Lasso propensity score model on the screened set. The doubly robust estimator is then

$h\big(\mathbb{E}(Y|(x_k,x_\ell),,)\big) = \beta_0 + \beta_1 x_{k} + \beta_2 x_\ell + \beta_3\; x_{k} x_\ell,$0

A distinctive feature is that propensity score model selection is outcome model-free, which the paper argues is necessary for maintaining double robustness. Under minimal-signal and dimensional-growth conditions, the screening step has a sure screening property,

$h\big(\mathbb{E}(Y|(x_k,x_\ell),,)\big) = \beta_0 + \beta_1 x_{k} + \beta_2 x_\ell + \beta_3\; x_{k} x_\ell,$1

Additional results establish variable-selection consistency, double robustness of the final estimator if either the propensity or outcome model is correctly specified, and oracle asymptotic normality.

The empirical study compares CBS with Outcome Adaptive Lasso, Robust Inference, and Double/Debiased Machine Learning, and reports favorable performance across realistic simulation settings. In the ADNI application, after residualizing age, gender, and education, the method retained the top 30 SNPs, selected loci concentrated on chromosome 19 including APOE and APOC1, and estimated that high tau increases ADAS-11 by about 5.96 points with 95% confidence interval [4.15, 7.76]. Here the phrase “causal screener” is apt in the variable-selection sense: the first stage screens covariates for downstream doubly robust causal estimation.

Several causal-discovery papers instantiate the same two-stage logic in graph learning rather than effect estimation. "A Hybrid Anytime Algorithm for the Constructiion of Causal Models From Sparse Data" searches over essential graphs by repeatedly running the PC algorithm with varying significance level and variable-order parameters, then converts each essential graph to a DAG and scores it with the Bayesian metric $h\big(\mathbb{E}(Y|(x_k,x_\ell),,)\big) = \beta_0 + \beta_1 x_{k} + \beta_2 x_\ell + \beta_3\; x_{k} x_\ell,$ 2 (Dash et al., 2013). The method is “anytime,” and its central empirical conclusion is that essential-graph search is a more effective use of computation than direct greedy search in DAG space for sparse data.

In time-series discovery, "Causal Discovery from Time Series with Hybrids of Constraint-Based and Noise-Based Algorithms" defines two hybrid classes: NBCB, which uses a noise-based method to infer a causal order and then a constraint-based method to prune edges, and CBNB, which first infers a skeleton constraint-based and then orients unresolved instantaneous edges using a noise-based method within undirected cycle groups (Bystrova et al., 2023). The framework is correct under causal sufficiency, the causal Markov condition, adjacency faithfulness, an identifiable functional model, and consistency throughout time, and the paper emphasizes that the hybrids require only adjacency faithfulness rather than full faithfulness.

"Hybrid Top-Down Global Causal Discovery with Local Search for Linear and Nonlinear Additive Noise Models" also follows a two-stage pattern, but with a global-to-local organization (Hiremath et al., 2024). LHTS and NHTS first recover a hierarchical topological sort, and ED then performs local edge discovery using conditioning sets

$h\big(\mathbb{E}(Y|(x_k,x_\ell),,)\big) = \beta_0 + \beta_1 x_{k} + \beta_2 x_\ell + \beta_3\; x_{k} x_\ell,$ 3

where $h\big(\mathbb{E}(Y|(x_k,x_\ell),,)\big) = \beta_0 + \beta_1 x_{k} + \beta_2 x_\ell + \beta_3\; x_{k} x_\ell,$ 4 captures potential confounders and $h\big(\mathbb{E}(Y|(x_k,x_\ell),,)\big) = \beta_0 + \beta_1 x_{k} + \beta_2 x_\ell + \beta_3\; x_{k} x_\ell,$ 5 potential mediators. The contribution is a top-down causal discovery pipeline that exploits local causal substructures to reduce regression and conditioning-set size. In the presence of latent confounders, "FRITL: A Hybrid Method for Causal Discovery in the Presence of Latent Confounders" begins with FCI to obtain a conservative PAG, refines adjacent pairs by regression residual independence, applies the Triad condition to detect shared latent confounders, and only then uses overcomplete ICA on maximal unresolved cliques (Chen et al., 2021).

Taken together, these works show that the hybrid two-stage causal screener is not a single algorithm but a recurring methodological template. A recurring implication is that hybridization is most effective when the first stage removes enough of the combinatorial burden to make the second stage reliable, yet not so aggressively that it discards the signal the second stage must recover. The same body of work also clarifies a common misconception: screening is not itself causal identification. Interaction screening does not establish causality in a potential-outcomes sense; Granger-based lead–lag discovery is not structural causation; proxy-based proximal learning depends on completeness and proxy validity; graph-discovery hybrids depend on assumptions such as causal sufficiency, faithfulness or adjacency faithfulness, identifiable functional models, or non-Gaussianity; and semi-continuous exposure estimation depends on ignorability and positivity (Jonker et al., 2024, Kim et al., 4 Feb 2026, Mastouri et al., 2021, Wang et al., 26 Nov 2025). The main significance of the hybrid design is therefore methodological rather than ontological: it offers a way to combine cheap broad search with narrower, more principled second-stage inference under explicitly stated assumptions.