Using Causal Order Knowledge to Protect Structural Equation Models from Cancelling Paths

Structural equation models (SEMs) are the standard tool in psychology, education and health research for testing causal chains between latent constructs, meaning quantities like "stress" or "self-efficacy" that are never measured directly and are inferred from several questionnaire items each. In these chains a direct effect and an indirect effect through a mediator often have opposite signs. Researchers call this suppression. When the two effects nearly balance, the overall correlation between cause and outcome is close to zero. Any procedure that drops relationships with near-zero association will then delete a real causal edge. This includes informal correlation screening before specifying the model, and also automated causal discovery algorithms. The discovery literature calls the problem a failure of faithfulness, the assumption that every statistical independence in the data reflects a missing causal path. It shows the problem is common rather than exotic: Geometry of the faithfulness assumption in causal inference shows that once tests need associations to exceed a small margin, near-cancelling parameter settings can make up close to 90% of parameter space for a sparse 10-node graph.

This note proposes a simple change to how discovery uses the background knowledge SEM researchers almost always have. SEM researchers rarely know the full graph, but they usually know a partial order: stressors come before coping, coping comes before outcomes. Existing discovery tools such as PC or FCI with "temporal tiers" use this order only to orient edges after the edge set is fixed. They still delete an edge whenever any conditioning set, including the empty set, makes two constructs look independent, and cancellation is exactly what produces that pattern. We propose to use the order during edge deletion as well. An edge from an earlier construct to a later one is removed only if the two are independent given the later construct's full set of order-predecessors. That conditioning set is the one under which cancellation cannot hide a direct effect. Cancellation then becomes something we can locate. Between tiers it is handled without assuming faithfulness. Within tiers it is flagged as ambiguous rather than silently resolved. The method runs on construct covariances corrected for measurement error rather than on raw item sum scores.

Core Claim

Suppose a linear SEM has latent constructs $\eta_1,\dots,\eta_p$ grouped into ordered tiers. Let $\mathrm{pre}(j)$ be the constructs in strictly earlier tiers than $\eta_j$, and assume no unmeasured confounding between constructs. Then the coefficient of $\eta_i$ in the population regression of $\eta_j$ on $\mathrm{pre}(j)$ equals the direct structural effect $\beta_{ij}$. This holds because the regression residual is the structural error term, which is independent of all predecessors. Our cross-tier deletion rule is therefore

$$\text{remove } \eta_i \to \eta_j \iff \rho\big(\eta_i,\eta_j \mid \mathrm{pre}(j)\setminus\{\eta_i\}\big)=0,$$

where $\rho(\cdot\mid\cdot)$ is the partial correlation. The rule is correct whatever the indirect paths sum to. PC-style deletion instead removes the edge if the partial correlation vanishes for some subset. Cancellation never fools the first rule and routinely fools the second.

What the rule does not fix is just as important, and it defines the rest of the program. First, within a tier there is no order. A pattern where X and Y are independent but become dependent given M fits either a collider (X → M ← Y) or a cancelling triangle, and Gaussian data cannot tell them apart. Second, unmeasured confounding between constructs makes the predecessor regression pick up non-causal association. Third, if constructs are represented by item sum scores, conditioning on a noisy mediator does not block the indirect path. The partial correlations then show neither the cancellation nor the true direct effect cleanly. The measurement model is what makes the rule usable on real SEM data.

We do not claim the regression fact itself is new. It is the reason a saturated "everything earlier predicts everything later" path model is safe. The contribution is three things built on it. The first is a discovery procedure that uses tier knowledge in the deletion phase. The second is how that procedure combines with weakened-faithfulness reasoning within tiers and with latent confounding. The third is a measurement of how much of the cancellation problem this removes, and how much remains, across realistic SEM regimes.

Relation to Prior Work

Two papers motivate the problem. Geometry of the faithfulness assumption in causal inference shows that near-cancellation occupies positive and often large volume of parameter space at finite sample sizes. Weakening faithfulness: some heuristic causal discovery algorithms shows that standard PC and GES lose edges and orientations under near-cancellation. GES is greedy equivalence search, a score-based method that adds and removes edges to improve a fit score. The second paper repairs these failures with PC+GES and a conservative orientation rule (GES+c) under triangle-faithfulness, which assumes faithfulness holds only within three-node subgraphs.

That paper also supplies part of our method. We reuse its conservative orientation logic inside tiers, where order knowledge gives nothing. We differ from it in two ways. Its algorithms assume no latent confounders and no measurement error. They also never use background order, which is the most valuable information an SEM researcher brings.

The non-Gaussian line is the closest existing work at this intersection. Learning Linear Non-Gaussian Causal Models in the Presence of Latent Variables uses overcomplete independent component analysis to recover causal order when covariances cancel. Independent component analysis separates mixed signals by exploiting their non-Gaussian shape, and "overcomplete" means there are more hidden sources than observed variables. Causal Discovery in Linear Latent Variable Models Subject to Measurement Error handles measurement-error SEMs with a faithfulness condition split specifically to rule out cancellation, plus a sparsest-graph principle.

Both papers rely on higher-order moments that do not cancel. The second also requires its unobserved-root dual to have latents without parents, which excludes the latent mediator chains typical of applied SEMs such as Perceived Academic Stress and Depression: The Mediation Role of Mobile Phone Addiction and Sleep Quality. Our method works with Gaussian or near-Gaussian ordinal data and allows latent mediators. It pays for this by requiring order knowledge, an assumption non-Gaussian methods do not need.

Efficient Identification in Linear Structural Causal Models with Auxiliary Cutsets constrains the confounded case. Its ACID algorithm identifies direct effects given a known graph with correlated errors. Its guarantees are explicitly generic, meaning they hold almost everywhere in parameter space and can fail on exactly the cancellation sets we care about. Nor does it discover graphs. Instead of building on its guarantees, we use it as a downstream check. Algebraic problems in structural equation modeling explains, through the trek rule (each covariance is a sum of path products over connecting routes), why cancellation is a polynomial vanishing condition. It also explains why exact algebraic treatment does not scale.

Finally, the local latent-variable learners Local Causal Structure Learning in the Presence of Latent Variables and Local Learning for Covariate Selection in Nonparametric Causal Effect Estimation with Latent Variables are efficient but assume faithfulness. We treat them as confounding-aware competitors that should fail on cancelled edges.

Research Program

The procedure has four stages. Each corresponds to one of the failure modes above and can be ablated on its own.

The first stage fits a confirmatory factor model, the standard SEM measurement model in which each construct generates its own items. This produces an estimated construct covariance matrix with sampling uncertainty, so no structural assumptions enter yet. We compare against sum scores to quantify how much measurement error distorts the deletion tests.

The second stage applies the predecessor-set deletion rule to every cross-tier pair of constructs. The test is a Wald or partial-correlation test that accounts for the uncertainty of the measurement model. When predecessor sets are large relative to sample size, power drops. We will study this trade-off directly, and test a variant that conditions on the predecessors still adjacent after a first pass, only when those are a superset of every separating set PC would find.

The third stage handles within-tier structure. It runs a triangle-faithfulness search restricted to each tier, in the style of the PC+GES+c algorithm from the weakening-faithfulness paper. Collider-versus-cancellation triples are reported as ambiguous rather than oriented.

The fourth stage addresses latent confounding. The second stage errs toward keeping edges when constructs are confounded: conditioning on a predecessor that is a collider can open spurious paths, but cannot hide real ones. We therefore prune the candidate graph by fitting full SEMs that include bidirected (correlated-error) edges, scored by the Bayesian information criterion with a preference for sparser graphs. Each surviving direct effect is checked with ACID identification on the selected mixed graph.

Alongside the procedure we will produce a cancellation audit. For every edge the procedure keeps but a marginal or PC-style test would have dropped, it reports the estimated direct and indirect components and their ratio. This gives applied researchers the explicit suppression diagnosis they currently get only by accident.

The theoretical deliverable is a soundness statement with a clear scope. The cross-tier edges are consistent under no confounding without any faithfulness assumption. Within-tier output is sound under triangle-faithfulness. The confounded case gives a superset guarantee for the second stage. We will state where each guarantee stops rather than claim full identification.

Baselines

The naive reference point is the common applied practice: screen construct pairs by marginal correlation (or by sum-score correlation), then fit the SEM that includes only the surviving paths. This is the procedure suppression silently breaks, and beating it on recovery of cancelled direct edges is the minimum bar.

The main discovery competitor is PC with temporal-tier background knowledge, run through Py-Tetrad and RPy-Tetrad: A New Python Interface with R Support for Tetrad Causal Search, with FCI with tiers as its confounding-aware counterpart. It gets exactly the same order knowledge and the same construct covariance matrix as our method. This isolates our single change: using order during deletion rather than only during orientation.

The strongest existing cancellation-robust method is PC+GES+c from Weakening faithfulness: some heuristic causal discovery algorithms. On random DAGs of up to 40 variables with near-cancellation, that paper reports GES+c raising arrow precision from 0.35 for plain GES to 0.96, and PC+GES raising the true adjacency rate from 0.75 to 0.95. We will run it with the same tuning budget, give it the tier knowledge as forbidden edges, and reproduce its causally sufficient setting first to confirm our implementation reaches comparable numbers. Beating it means a higher true adjacency rate on cancelled edges and an arrow precision at least matching it. This must hold specifically in the regimes it was not designed for: measurement error and latent confounding.

For the non-Gaussian route, we run the measurement-error method of Causal Discovery in Linear Latent Variable Models Subject to Measurement Error (direct ordered grouping) and the overcomplete-ICA method of Learning Linear Non-Gaussian Causal Models in the Presence of Latent Variables on the same items without order knowledge. That reverse comparison asks what order knowledge buys relative to non-Gaussianity. The file reports no numerical results for these on comparable benchmarks: no reported result found on Lacuna for F1 of direct ordered grouping or overcomplete ICA under controlled cancellation. For confounded settings we also run the local learners MMB-by-MMB and LSAS.

The ablations isolate each ingredient: the deletion rule replaced by PC deletion, confirmatory factor estimates replaced by sum scores, the within-tier conservative stage replaced by plain orientation, and the fourth-stage pruning removed. The most important ablation is the saturated tier-forward regression with Wald pruning and no within-tier or confounding stages. That is the simplest version of our own idea, and it must be shown to be insufficient somewhere for the full procedure to be worth its complexity.

Why not just…?

Why not just fit the hypothesized SEM by maximum likelihood, since cancellation is measure zero? That answer assumes the researcher already has the right path diagram, and the open problem is how the diagram was chosen. The geometry paper shows that near-cancellation is not rare at realistic test margins, and the weakening-faithfulness results show discovery algorithms losing a quarter of adjacencies to it. The naive-screening baseline will measure directly how often a correct direct path never makes it into the model.

Why not just regress each construct on everything earlier in the order and prune by significance? This is a real threat, and it is the core of our second stage. Without confounding and with a total order it is all that is needed, and we expect our method to tie with it there. We will report that tie rather than hide it. The full procedure must earn its keep in three regimes. The first is large tiers with no internal order, where regression across tiers says nothing. The second is latent confounding between constructs, where regression keeps spurious edges that the fourth stage must remove. The third is sum-score data, where measurement error contaminates the regressions. The experiments sweep tier size and confounding density so that the crossover point, if any, is visible.

Why not just use non-Gaussian methods, which detect cancellation without order knowledge? Questionnaire data are usually ordinal and close to symmetric, which weakens the higher-moment signal these methods rely on. The measurement-error method also requires root latents. The overcomplete-ICA paper notes that its optimization degrades beyond about ten variables. We will vary the non-Gaussianity of the generated noise. If these methods match ours even at mild skewness, the right recommendation becomes "use order knowledge only to break their remaining ambiguities", and we will say so.

Why not just use FCI or the local latent-variable learners? They handle confounding but delete edges on any found independence. Cancellation produces exactly such independences, so they should lose cancelled edges at roughly the rate plain PC does. The matched-input comparison will test this.

Proposed Experiments

The primary testbed is synthetic linear SEMs with 6 to 30 latent constructs, three to five indicators each, and tiered structure. Parameters are drawn as in the geometry and weakening-faithfulness studies, with targeted triangles where the direct effect is set to cancel the indirect effect up to a controlled margin $\lambda$. We vary sample size from 200 to 5,000 to span applied SEM practice, along with indicator reliability, the number of constructs per tier, the density of construct-level correlated errors, and noise skewness. Metrics are true adjacency rate on cancelled edges, overall adjacency precision and recall, arrow precision, structural Hamming distance computed with Causal-learn: Causal Discovery in Python, the rate of within-tier triples correctly flagged as ambiguous, and absolute error of the identified direct effects.

To test whether conclusions transfer to realistic topologies, we will place linear-Gaussian SEMs on the WIN95PTS (76 nodes) and ANDES (223 nodes) skeletons used by the local-learning papers, with tiers taken from topological layers and a fraction of nodes hidden as confounders. We will also report the number of conditional independence tests. For context, on ANDES MMB-by-MMB uses about 5,000 tests against over 900,000 for global FCI.

On real data, Cattaneo2 provides a sanity check with a well-understood order. LSAS recovers a smoking effect of about −230g with 158 tests. Our procedure should recover a comparable effect, and it should flag any covariate whose relevance is masked by cancellation.

The decisive check is a single plot. It shows true adjacency rate on cancelled edges against the cancellation margin, for our method, PC with tiers, and PC+GES+c, separately for sum scores and factor-model inputs. If our curve stays flat as $\lambda \to 0$ while PC with tiers collapses, the central claim holds. The second check compares the full procedure with saturated tier-forward regression as tier size and confounding density grow. If the two never separate, the honest conclusion is that SEM researchers need only that regression plus a cancellation audit, and we will publish that simpler recommendation.