Skip to content

Expert Dependency Review

Facilitated review — instantiates Conditional Independence Boundary Mapping

A facilitated session where domain experts define the target and hand-draw the dependency structure — supplying edges, directions, and hidden variables the data alone can't reveal.

Expert Dependency Review is a structured elicitation in which domain experts do two things: pin down the target unit and the task it must serve, and draft or critique the dependency graph — which variables plausibly influence which, in which direction, and, critically, what is missing. Its distinguishing idea is that it is the human-knowledge front end of the archetype. It supplies exactly what observational data cannot: the causal direction that data leaves undetermined, and unmeasured variables that were never logged.[n1] It is the only mechanism here that can add a variable nobody recorded.

Example

A grid operator wants to know what drives transformer-failure risk at a substation, but failures are rare and the logged history is thin — too sparse to learn a structure from. So instead of mining data, they convene protection engineers. The engineers first define the target precisely: time-to-failure for a specific transformer class under load, framed to support a maintenance-scheduling decision. Then they sketch the dependency graph: load cycling drives winding temperature, which drives insulation aging, which drives failure; ambient temperature is a fork feeding both load and winding temperature; and — the decisive contribution — they posit moisture ingress as a hidden common cause that the SCADA system never measured. No data-driven method could have surfaced a variable that was never recorded. That drafted graph becomes the substrate the extraction and d-separation steps operate on.

How it works

The distinctive move is capturing tacit causal knowledge and running it deliberately before the data-driven steps, so those steps have a structured hypothesis to test rather than a blank slate. Experts assert edge directions, name latent variables, and — where they disagree — either reconcile or record the competing edges as explicitly uncertain. It is not a computation over data; it is an elicitation whose product is a graph and a target definition, each carrying the experts' confidence and their doubts.

Tuning parameters

The dials that adapt the session to a problem:

  • Panel breadth — a single expert versus a diverse panel; more experts surface more hidden variables and blind spots but require reconciliation.
  • Elicitation structure — free sketching versus constrained judgments (for example, forced pairwise direction calls); more structure reduces anchoring but costs time.
  • Disagreement handling — forcing consensus versus recording competing edges as uncertain; recording preserves honest uncertainty for the empirical steps to resolve.
  • Latent-variable license — how freely experts may posit unmeasured common causes; a generous license surfaces confounders but complicates every downstream step that must then account for them.

When it helps, and when it misleads

Its strength is supplying structure that data cannot: orientation within a Markov equivalence class, and entirely unrecorded variables. When data is thin, the system is novel, or the stakes demand a causal story rather than a correlational one, it is often the only way to get a graph at all.

Its failure mode is that experts are confidently wrong, share blind spots, and import motivated or fashionable causal stories; a plausible drawn edge is a hypothesis, not a fact. The classic misuse is treating the elicited graph as ground truth and skipping empirical validation, so a downstream extraction inherits the experts' errors as though they were proven. The discipline that keeps it honest is to mark every edge as a testable claim, feed the graph to the test suite, and reconcile it against any data-driven structure learning rather than trusting either source blindly.

How it implements the components

The review fills the model-supply components of the archetype — the human inputs the rest of the workflow consumes:

  • target_unit_definition — the session fixes what the target unit is and the decision it must serve, before any variable is drawn.
  • dependency_graph_or_model — the experts produce (or critique) the graph of who-depends-on-whom, including edge directions and posited hidden variables.

It does not compute or test anything on the graph it produces — reading the blanket off it is Bayesian Network Markov Blanket Extraction, and testing its claimed independences against data is the Conditional-Independence Test Suite.

Editorial Notes

Form Classification

Form family: Communication, Facilitation & Learning

Rationale: A facilitated elicitation session surfaces experts' tacit causal knowledge, reconciles or records disagreement, and builds shared target and dependency understanding.

Nearest alternative: Assessment, Review & Assurance — Experts judge edges, but the mechanism generates the hypothesis graph through participatory elicitation rather than auditing existing work.

Review outcome: Adjudicated after independent review; high confidence.

Origin Attribution

Primary origin: Statistics & Experimental Design

Origin pattern: Cross-disciplinary synthesis

Present-day reach: Specialized

Rationale: Expert elicitation of causal dependencies and hidden variables is established in causal graphical-model methodology.

Related originating lineages:

  • Data Science & Analytics — Bayesian-network and machine-learning practice materially operationalizes expert-drawn graph structure.
  • Operations Research — Decision-model construction materially uses facilitated experts to identify directed dependencies and hidden variables.

Review resolution: Both reviewers agree that statistics_experimental_design is primary. I retain data_science, operations_research only as formative origin lineages; cross_disciplinary_synthesis is appropriate because the final form materially combines the agreed primary with the retained formative lineages. Reach is specialized because the mechanism remains tied to a bounded professional technique, an applicability judgment kept separate from provenance. Encyclopedia synthesis is true because the exact generalized artifact is an encyclopedia-authored combination or refinement. No unresolved historical ambiguity remains after reconciling the secondary fields.

Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.

Review outcome: Reconciled after independent review; high confidence.

Notes

When the expert-drawn graph and a data-learned one disagree, the disagreement is signal, not noise — often the experts are holding a real unmeasured variable and the data is holding a real edge the experts rationalized away. Reconcile the two deliberately rather than defaulting to whichever source is nearer to hand.

[n1] Observational data determine a causal graph only up to its Markov equivalence class — structures sharing a skeleton and the same v-structures imply identical conditional independences, so some edge orientations, and any entirely unmeasured variable, cannot be recovered from data and must come from experiment or expert knowledge.