Letting a Logic Solver Choose the Structural Model: Discovering Paths Between Latent Constructs from Indicator Rank Tests

Structural equation modeling (SEM) is how much of psychology, education and public health studies chains of causes among quantities that cannot be measured directly, such as self-efficacy, stress or health literacy. Each construct is measured through several questionnaire items. SEM is confirmatory: the researcher writes down the arrows between constructs in advance, fits them to the item covariances, and reports coefficients. If the wrong arrows were drawn, the coefficients are biased. The fitted model also says nothing about which other arrow patterns would fit the data equally well.

Logic-based causal discovery offers the missing half. It turns statistical test results into clauses for a SAT or answer-set solver, a general-purpose search engine for logical constraints. The solver then reports which edges are forced, which are forbidden, and which the data cannot settle. So far it has worked only on directly observed variables. We propose running that machinery at the level of latent constructs. It would be fed tests that are valid despite measurement error: rank tests on cross-covariance blocks of the indicators, rather than tests on noisy construct scores. The output would be the set of all structural models consistent with the data and with the researcher's measurement design. Each surviving model is then fitted by ordinary SEM.

Core Claim

The measurement model is usually known by design, because item 7 was written to measure self-efficacy. When it is, a conditional independence among latent constructs, "construct $L_i$ is independent of $L_j$ given the set $L_S$," corresponds to a rank condition on observed covariances. The cross-covariance matrix between one indicator of $L_i$ together with one indicator per member of $S$, and one indicator of $L_j$ together with a disjoint indicator per member of $S$, should have rank at most $|S|$. The mechanism is that for pure indicators, this cross-covariance block factors through the covariance of the latents, so its rank is capped by the conditioning set.

Such rank tests are immune to the attenuation that corrupts tests on estimated factor scores. Conditioning on a noisy proxy of a mediator does not block the path through it, so partial correlations of factor scores report dependencies that are not there. Rank tests can therefore serve as soft constraints in existing logic encodings, with their weights derived from the test statistics. A second set of Boolean variables stands for possible departures from the declared measurement design, namely cross-loadings and correlated item errors. Each such variable switches off the rank constraints it would invalidate, and the solver pays a penalty for every one it uses.

The result would be the first logic-based discovery method for latent SEMs. It returns forced, forbidden and undetermined paths among constructs, accepts domain knowledge as hard clauses, and hands a small set of fully specified models to standard SEM estimation for coefficients.

Relation to Prior Work

The logic side supplies the method. Discovering Cyclic Causal Models with Latent Variables: A General SAT-Based Procedure encodes d-separation over cyclic graphs with confounders and computes the backbone, meaning the edges fixed across all consistent graphs. It assumes error-free tests, and it treats latents only as bidirected edges between observed variables, with no indicators. Ancestral Causal Inference makes such encodings robust to noisy tests by weighted answer-set optimization over ancestral relations, and A constraint-based approach to incorporate prior knowledge in causal models shows that background knowledge enters as extra clauses. We reuse all three: the encodings, the weighting scheme and the knowledge clauses. What changes is the source of the constraints and the variables they range over. Relaxed-faithfulness encodings, SAT-Based Causal Discovery under Weaker Assumptions and ASP-based Discovery of Semi-Markovian Causal Models under Weaker Assumptions, matter here because SEM data often show near-cancelling paths.

The latent-variable side both motivates and constrains the design. Latent Hierarchical Causal Structure Discovery with Rank Constraints already uses covariance rank deficiency to recover latent hierarchies. It is procedural, however: early rank-test errors propagate, and there is no global consistency check and no place for domain axioms. Score-Based Causal Discovery of Latent Variable Causal Models (SALAD) scores latent graphs by the degrees of freedom of the observed marginal. It supports exact search at roughly 10–15 variables and a continuous relaxation beyond that. It returns a best-scoring graph rather than the set of graphs the data cannot distinguish, and it cannot take logical side conditions.

Neither line uses the fact that SEM users typically know the measurement model. We treat that knowledge as the regime in which the logic encoding becomes tractable: the solver searches over tens of possible edges among constructs, not hundreds among indicators. On the application side, Aligning MLLM Benchmark With Human Preferences via Structural Equation Modeling imposed a Perception–Memory–Reasoning hierarchy on multimodal model benchmarks by confirmatory SEM. That makes it a natural case where the imposed ordering itself can be tested.

Research Program

We would build three components. The first is a test generator. For each pair of constructs and each candidate conditioning set up to a size cap, it draws several disjoint indicator selections and runs rank tests on the cross-covariance blocks. It pools evidence across selections and converts the result into a weight, either the log-odds of the p-value as in ACI or a bootstrap log Bayes factor.

The second is the encoding. The Hyttinen-style d-separation clauses, or ACI's ancestral clauses, are stated over the construct variables. Added to them are impurity variables for cross-loadings and correlated errors: each rank constraint is conditioned on the absence of impurities that would break it, and each impurity used carries a sparsity penalty. Researcher knowledge enters as hard clauses, for example "demographics are exogenous" or "wave-2 constructs cannot cause wave-1 constructs."

The third is a post-solver. It enumerates the optimal and near-optimal solutions and computes the backbone with a confidence margin, defined as the penalty increase needed to flip each edge. It then fits each surviving model by maximum-likelihood SEM, reporting coefficients and indirect effects with their spread across the equivalent models.

The analysis has two parts. We will characterize when latent conditional independence is identifiable from rank tests given at least two pure indicators per construct, and we will derive how the solver's search space scales with the number of constructs rather than the number of items. We will also study test dependence, the known weakness of additive weighting. Many rank tests reuse the same indicators, so we compare naive additive weights against weights discounted by bootstrap co-rejection rates among tests.

Baselines

The naive reference is standard practice. A researcher fits the hypothesized SEM, or a two-step pipeline estimates factor scores and then runs PC or ACI on them. The two-step variant doubles as the key ablation: our exact solver and encoding, fed partial-correlation tests on factor scores instead of indicator rank tests. Any gain over it isolates the value of measurement-error-robust tests. We expect it to hallucinate direct paths that bypass mediators, and it should do worse as item reliability drops.

The strongest existing competitor is SALAD. Score-Based Causal Discovery of Latent Variable Causal Models reports an F1 of 0.99 on linear one-factor models at T = 100 samples, versus 0.90 for the constraint-based FOFC method. That is the bar for plain one-factor structure recovery. We will run SALAD exact search, SALAD-CS where exact search is infeasible, and our method on identical simulated datasets. SALAD's search will be restricted to the true indicator clustering so both methods receive the same measurement knowledge, and all methods get the same wall-clock budget and the same significance-level or penalty tuning grid on held-out simulations.

Beating SALAD does not require higher F1 on pure one-factor data, where it is near ceiling. It means matching it there within noise while winning in three places: impure measurement, confounded or cyclic latent structure, and calibration of the "determined" edges.

The rank-constraint procedure of Latent Hierarchical Causal Structure Discovery with Rank Constraints is the procedural analogue of our tests. We run it with the same tests and significance levels; no reported result found on Lacuna for its accuracy in this file.

Why not just…?

Why not just enumerate every structural model and pick the best BIC? For four or five constructs this is feasible and is a real threat on accuracy, so we include it as a baseline in that range. Its limits are three. Model scores tie across equivalent models, so choosing the best one hides exactly the indeterminacy the backbone reports. Adding latent confounders among constructs or feedback loops multiplies the candidate set. And it cannot take logical knowledge beyond hand-filtering. The experiment settles it by reporting when enumeration stops finishing within budget as constructs grow from 4 to 12, and whether its top model lies inside our reported set.

Why not just compute factor scores and run existing logic-based discovery on them? Conditioning on an error-laden score of a mediator leaves residual dependence, so these tests systematically report extra direct paths. This is the attenuation problem SEM was invented to fix. The ablation above measures the damage directly as a function of item reliability. If the damage is negligible at reliabilities typical of published scales, our contribution shrinks to the impurity handling, and we would say so.

Why not just give SALAD the measurement clustering and stop? This is the most serious alternative for accuracy on clean data. We expect SALAD to be competitive or better there, because it pools evidence in one likelihood rather than across correlated tests. Our claim rests on outputs SALAD does not produce: calibrated forced and forbidden edges, hard domain constraints, and cycles or confounding among constructs. If SALAD's bootstrap edge frequencies turn out as well calibrated as our backbone margins, that claim weakens, so we measure exactly that.

Proposed Experiments

The primary testbed is simulation matched to the SEM literature. It uses 4–12 constructs with 3–5 indicators each, loadings chosen to give reliabilities between 0.6 and 0.9, and sample sizes from 200 to 2000, in the range of the N = 568 teacher study in The Relationship Between Preschool Inclusive Education Teachers' Organizational Support and Work Engagement. Structures include serial mediation chains of the kind in Perceived Academic Stress and Depression, latent confounders, one feedback loop, and 0–20% impure indicators.

The metrics are structural Hamming distance and edge F1 among constructs. We also measure calibration: the fraction of edges declared forced or forbidden that are correct, plotted against the confidence margin. Coefficient and indirect-effect error after SEM fitting come next, then solver runtime.

For real data, we reanalyze the Sachs protein data at the observed level, as a sanity check that the engine reproduces ACI. We then re-examine the multimodal benchmark task scores from Zou et al. and ask whether the Perception→Memory→Reasoning ordering is forced by the data or merely one of several fits. The number of evaluated models there is small, so we expect many undetermined paths and treat an honest "undetermined" as a legitimate finding.

Two outcomes would decide the proposal. The first is whether, at reliability 0.7 and N = 500, the rank-test solver has lower false-direct-path rates than the factor-score ablation. The second is whether its forced edges are correct substantially more often than SALAD's high-frequency bootstrap edges at matched coverage. If both fail, the idea should be dropped.