Latent Variable Causal Discovery under Selection Bias

AI summary

Overview

• Research area: Causal discovery / causal structure learning from observational data, specifically latent variable causal discovery in the presence of selection bias.
• Technical level: Advanced. The paper assumes familiarity with directed acyclic graphs (DAGs), d-separation, conditional independence, linear Gaussian structural equation models, covariance algebra, and the rank-based t-separation criterion of Sullivant et al. (2010).
• One-sentence scope: The paper introduces "generalized rank constraints," a graphical criterion describing the ranks of covariance submatrices in selection-biased linear Gaussian data, and uses them to identify latent structure — including the classical one-factor model — under selection bias.

What This Paper Is About

Causal discovery methods that handle latent variables normally rely on tools that go beyond conditional independence (CI) constraints, such as rank constraints, equality constraints, high-order moment constraints, matrix decomposition, copula models, and mixture oracles. The paper argues that none of these tools has been adapted to selection bias — the preferential inclusion of data points based on unknown mechanisms (Heckman, 1977) — so once selection bias is introduced, practitioners are forced back to basic CI constraints and effectively to the FCI algorithm, which cannot recover structure among latent variables. The goal of the paper is to close that gap by developing a tool that goes beyond CI constraints while handling both latent variables and selection bias at the same time.

Key Contributions

1. A generalized rank constraint with a graphical criterion. The paper shows that although selection makes the joint distribution far more complicated (e.g., truncation produces a truncated Gaussian whose covariances and higher moments are hard to express), the ranks of the biased covariance submatrices remain well-defined and informative. Theorem 1 states that for subsets A, B of the observed variables X, assuming genericity, rank(Σ^(S)_{A,B}) equals the minimum of |C| + |D| over pairs (C, D) drawn from X ∪ Y that t-separate (A ∪ Y, B ∪ Y) in the selection-augmented graph G^(S), minus k, the number of single selections. When there is no selection (S = ∅ and k = 0), this reduces to the original rank constraint criterion of Sullivant et al. (2010).

2. A formal model of selection compatible with rank constraints. The paper defines "linear selection mechanisms," a configuration S of k tuples (V_i, β_i, ε_i, 𝒴_i), where V_i ⊆ X is the subset of variables involved in the i-th selection, β_i is a vector of nonzero linear coefficients, ε_i is an independent noise term (possibly non-Gaussian or degenerate to a constant), and 𝒴_i ⊊ ℝ is the admissible value set. A sample is kept only if Y_i = β_iᵀV_i + ε_i lies in 𝒴_i for all selections. The authors show that hard truncation, the logistic selection model (Dubin and Rivers, 1989), the probit selection model (Heckman, 1977), and the stabilizing selection model (Lande and Arnold, 1983) all arise as special cases of this definition.

3. Identifiability results distinguishing latent variables from selection bias. The paper introduces "rank equivalence" for selection-augmented DAGs over the same observed variables, shows that latent variables and selection bias are sometimes distinguishable by CI constraints alone (Example 4), and shows that rank constraints can distinguish them even in CI-equivalent graphs (Example 5, using variants of the "spider" structure in Table 1).

4. Identification of the one-factor model under selection bias. The paper extends the one-factor model of Silva et al. (2003) to include selection on the latent variables (⋃_{j=1}^{k} V_j ⊂ L, with each latent L_i having at least two measurements, |X_i| ≥ 2), and claims that the one-factor model can be identified under selection bias, with Proposition 3 stating that CIs among the latent variables can be recovered from rank constraints on the observed data.

Main Findings

• Ranks survive selection. The inverse Tetrad example (Example 2.1) illustrates that four originally independent variables selected by truncating on a linear sum retain the same low ranks as the original Tetrad structure: rank(Σ'{X1,X2},{X3,X4}) = 1, rank(Σ'{X1,X3},{X2,X4}) = 1, and rank(Σ'_{X1,X4},{X2,X3}) = 1. The paper notes that the selected data no longer follows a linear Gaussian model, or even a linear structural equation model, yet these low-rank structures persist.

• Tetrad baseline. In the original (unselected) Tetrad structure, the same three submatrices have rank 1, explained by t-separation: (∅, {L}) or ({L}, ∅) t-separates ({X1, X2}, {X3, X4}), giving rank = 1 = 0 + 1. These low ranks do not follow from any conditional independence, as there are no CIs among {X1, X2, X3, X4}.

• Selection version of the same explanation. In the selection-augmented inverse Tetrad graph, ({Y}, {Y}) t-separates ({X1, X2, Y}, {X3, X4, Y}), which explains rank = 1 = 1 + 1 − 1.

• Sometimes latent variables and selection are distinguishable. Example 4 shows two graphs whose CI patterns differ: in the first, X1 ⟂ X4, X2 ⟂ X3, X1 not⟂ X4 | X3, and X2 not⟂ X3 | X4 imply that latent variables must exist between X3 and X4; in the second, the only two CIs among observed variables, X1 ⟂ X4 | X2, X3 and X2 ⟂ X3 | X1, X4, imply that selection bias involving all four variables must be present.

• Rank constraints add discriminating power beyond CI. In the spider example and its selection variants (Table 1), only A and B are observed and there are no CIs among them, yet low ranks exist and differ across graphs. In the original spider structure the rank between A and B is 2|C| rather than |C| — described as a property unique to the original structure, up to indeterminacies inside groups. In the second and third columns, while other ranks remain unchanged, either the rank between A and B increases or the rank between A1B1 and A2B2 decreases, which the paper says cannot be achieved by any graph without selection. The table's rank values include l + c, r + c, 2c, 2c + d, c, l + r + c, and l + r, under assumptions a, b ≫ l, r > c, d.

• Worked numerical illustration. Figure 2b visualizes the inverse Tetrad with σ_i² = 1, λ1, λ2, λ3, λ4 = 1, −2, 3, −4, and a, b = 3, 10, showing selected and unselected samples together with covariance values in the selected data.

• Not established: completeness. The paper states that fully characterizing the rank equivalence class — analogous to maximal ancestral graphs (MAGs) for CI constraints (Richardson and Spirtes, 2002) — remains an open challenge and is beyond its scope.

• Experiment details: The provided content states that simulations and real-world experiments confirm the effectiveness of the rank constraints, but the specific datasets, sample sizes, baselines, and numerical results are not reported in the available text (the content is truncated within the statement of Proposition 3).

Methodology in Plain English

The authors start from a known idea in the causal discovery literature: in linear Gaussian models, conditional independencies show up as zero partial correlations, which are equivalent to covariance submatrices having low rank. The t-separation criterion (Sullivant et al., 2010) explains exactly which low ranks are entailed by a graph, and it catches more than conditional independencies do. Their central question is whether those ranks still mean anything once the data have been filtered by selection.

Their strategy is to avoid modelling the full distribution after selection — because even simple truncation yields a truncated Gaussian whose covariance has no convenient closed form — and instead look for invariant patterns in the ranks of the biased covariance matrix. To do this, they treat selection as part of the causal process: they add a selection response node Y_i to the graph for each selection condition, with edges from every variable involved in that condition, following Bareinboim and Pearl (2012). They restrict attention to a class they call linear selection mechanisms, where the selection condition is a linear combination of variables plus independent noise, and admission to the dataset means this response lands in a specified set of values.

They then prove a graphical criterion (Theorem 1) relating the rank of any observed covariance submatrix in the biased data to a minimum t-separation quantity computed in the augmented graph, offset by the number of selection conditions. The intuition offered in the paper is a "dimensional bottleneck": when dependence in the data cannot be fully conditioned away, the number of latent or selection-induced channels through which the dependence flows leaves a trace in the ranks. Finally, they apply this criterion to a concrete, practically motivated model — the one-factor model with latent variables that are measured by multiple observed indicators, plus selection on the latent variables — to argue that latent structure can be recovered despite the bias.

Why This Matters

• Impact on research. Prior to this work, the only tool available for causal discovery when both latent variables and selection bias were present was the FCI algorithm, which operates only on observed variables and has no capability to identify relations among latent variables. This paper provides the first tool beyond CI constraints that is adapted to selection, positioning it as a stepping stone for new algorithms in the same way the original rank constraints enabled algorithms beyond FCI.

• Real-world applications:

  • Psychological and personality surveys: measured responses act as proxies for latent traits such as personality or political orientation, and people with certain traits are more likely to participate in the survey.
  • Recruitment via mail or phone: the contact method can systematically skew respondent groups by economic and education level.
  • Multi-criteria admissions and multi-criteria filtering systems: the model explicitly allows multiple selections to act simultaneously, matching situations where several independent conditions determine inclusion.
  • Stabilizing selection and truncation phenomena in biological or econometric settings: the paper notes the stabilizing selection model (Lande and Arnold, 1983), logistic selection (Dubin and Rivers, 1989), and probit selection (Heckman, 1977) as special cases of its linear selection mechanism.

• Industry relevance. Any dataset assembled by a filtering or eligibility step — credit scoring, hiring, admissions, clinical trial enrolment, opt-in product telemetry — violates the random-sampling assumption. The paper's framework gives such settings a way to reason formally about whether the observed dependence among variables reflects latent structure, the selection mechanism, or both.

Future Directions

• Characterize the rank equivalence class. The paper states that completeness in distinguishing latent variables from selection bias requires characterizing the rank equivalence class, analogous to maximal ancestral graphs for CI constraints, and explicitly leaves this as an open challenge beyond its scope.
• Handle more complex selection mechanisms. The authors raise the questions of what happens when selection involves randomness rather than truncation, and when multiple selection mechanisms act simultaneously — both are addressed within the linear selection framework, but the paper frames them as motivating questions for the general theory.
• Develop full algorithms. Just as the original rank constraints enabled algorithms beyond FCI, the generalized rank constraints are presented as paving the way for algorithms that recover both latent causal structure and selection structure; such algorithms are the natural next step.
• Broaden beyond linear Gaussian models. The entire framework assumes a linear Gaussian model with linear selection mechanisms, so extending rank-based reasoning to non-Gaussian or nonlinear settings is an implicit limitation. The paper refers readers to Section 6 for a discussion of potential limitations, though the details of that discussion are not included in the available content.

Target Audience

Researchers and graduate students in causal discovery, causal inference, and graphical models who are already comfortable with DAGs, d-separation, and linear Gaussian structural equation models. It is also relevant to applied statisticians and methodologists working with survey, psychometric, econometric, or any filtered/opt-in data where latent constructs are measured indirectly and selection bias is suspected. Practitioners looking for off-the-shelf software will not find it here; the paper is theoretical, contributing a graphical criterion and identifiability results rather than a ready-to-use pipeline.

Authors’ abstract

Addressing selection bias in latent variable causal discovery is important yet underexplored, largely due to a lack of suitable statistical tools: While various tools beyond basic conditional independencies have been developed to handle latent variables, none have been adapted for selection bias. We make an attempt by studying rank constraints, which, as a generalization to conditional independence constraints, exploits the ranks of covariance submatrices in linear Gaussian models. We show that although selection can significantly complicate the joint distribution, interestingly, the ranks in the biased covariance matrices still preserve meaningful information about both causal structures and selection mechanisms. We provide a graph-theoretic characterization of such rank constraints. Using this tool, we demonstrate that the one-factor model, a classical latent variable model, can be identified under selection bias. Simulations and real-world experiments confirm the effectiveness of using our rank constraints.