Causal Rationales

A causal rationale must hold true under active intervention. If the explanation claims that feature X caused outcome Y, then manually setting X to a different value (while holding all else constant) must predictably change Y. This property distinguishes causal relationships from spurious correlations that disappear when the data-generating process is perturbed.

• Do-calculus formalizes this using the do(X=x) operator
• Contrast with observational P(Y|X) which captures mere association
• Example: A loan denial rationale citing 'low income' is only causal if increasing income would actually change the decision

True causal rationales are derived from an explicit Structural Causal Model (SCM) — a directed acyclic graph encoding domain knowledge about how variables influence one another. The rationale references the causal paths and mechanisms within this graph rather than surface-level feature importance.

• Nodes represent variables; edges represent direct causal relationships
• Structural equations define the functional form of each causal mechanism
• Enables answering counterfactual queries: 'What would have happened if...?'
• Example: A medical diagnosis rationale traces the path from symptom → underlying condition → observed lab result

A causal rationale explicitly accounts for and adjusts for confounding variables — hidden factors that influence both the purported cause and the effect. Without confounder adjustment, explanations may attribute outcomes to the wrong features entirely.

• Backdoor criterion: identifies which variables must be conditioned on to block spurious paths
• Front-door criterion: alternative adjustment when confounders are unobserved
• Example: A hiring rationale that credits a university degree for success must adjust for socioeconomic background, which influences both degree attainment and career outcomes
• Failure mode: Simpson's paradox, where trends reverse after proper stratification

Causal rationales support counterfactual reasoning — they can articulate not just why something happened, but what minimal changes would have produced a different outcome. This is the gold standard for actionable explanations and is impossible with purely associational models.

• Necessary cause: Without this factor, the outcome would not have occurred
• Sufficient cause: This factor alone would have produced the outcome
• Minimal sufficient cause: The smallest set of changes needed to flip the prediction
• Example: 'Your loan was denied because your debt-to-income ratio exceeded 43%. Reducing it to 36% would qualify you for approval.'

Rather than attributing importance to raw input features, causal rationales explain decisions at the level of generative mechanisms. They describe the process by which causes produce effects, not just which variables are correlated with the output.

• Distinguishes direct effects from indirect effects mediated through other variables
• Mediation analysis decomposes total causal effect into path-specific components
• Example: A predictive maintenance rationale explains that vibration anomaly → bearing wear → increased friction → imminent failure, rather than simply flagging 'vibration = high'
• Provides engineering teams with actionable diagnostic chains, not just alerts

Causal rationales remain valid under distribution shift — when the model encounters data from a different environment or population. Because they capture invariant causal mechanisms rather than brittle statistical patterns, these explanations do not break when the input distribution changes.

• Causal invariance principle: mechanisms generalize across domains; correlations do not
• Contrast with saliency maps that highlight different pixels under slight image rotations
• Example: A causal medical rationale linking a specific pathogen to symptoms remains valid across different hospital populations, while a correlational rationale might rely on population-specific artifacts
• Critical for high-stakes deployment where training and production distributions diverge